Docstoc

International Journal of Computer Science - PDF - PDF

Document Sample
International Journal of Computer Science - PDF - PDF Powered By Docstoc
					IJCSIS Vol. 6, No. 3, December 2009 ISSN 1947-5500

International Journal of Computer Science & Information Security

© IJCSIS PUBLICATION 2009

IJCSIS Editorial
Message from Managing Editor
The IJCSIS Volume 6 No. 3 December 2009 issue contains a set of papers (with acceptance rate of ~ 33%), selected after rigorous journal-style review process. In this publication, you will find high-quality contributions from researchers in the diverse area of computer science, networking, emerging technologies and information security. With a continuing open-access policy, we are pleased to invite our dear readers to appreciate this collection of remarkable computer science research works.

Special thanks to our reviewers and technical sponsors for their valuable service.

Available at http://sites.google.com/site/ijcsis/ IJCSIS Vol. 6, No. 3, December 2009 Edition ISSN 1947-5500 © IJCSIS 2010, USA.

Indexed by (among others):

TABLE OF CONTENTS
1. Genealogical Information Search by Using Parent Bidirectional Breadth Algorithm and Rule Based Relationship (pp. 001-006) Sumitra Nuanmeesri, Department of Information Technology, King Mongkut’s University of Technology North Bangkok, Bangkok, Thailand Chanasak Baitiang, Department of Applied Science, King Mongkut’s University of Technology North Bangkok, Bangkok, Thailand Payung Meesad, Department of Information Technology King Mongkut’s University of Technology North Bangkok, Bangkok, Thailand 2. Web-Based Expert System for Civil Service Regulations: RCSES (pp. 007-016) Mofreh Hogo, Dept. of electrical engineering, Technology, Higher Institution of Technology Benha, Benha University, Egypt. Khaled Fouad, Central Lab. for Agricultural Expert Systems (CLAES) Fouad Mousa, Business management Dept, Faculty of commerce, Assuit university 3. A Wide-range Survey on Recall-Based Graphical User Authentications Algorithms Based on ISO and Attack Patterns (pp. 017-025) Arash Habibi Lashkari, Computer Science and Information Technology, University of Malaya (UM) Kuala Lumpur, Malaysia Dr. Rosli Saleh, Computer Science and Information Technology, University of Malaya (UM) Kuala Lumpur, Malaysia Samaneh Farmand, Computer Science and Information Technology (IT), University Malaya (UM) Kuala Lumpur, Malaysia Dr. Omar Bin Zakaria, Computer Science and Data Communication (MCS), University of Malaya (UM) Kuala Lumpur, Malaysia 4. A New Method to Extract Dorsal Hand Vein Pattern using Quadratic Inference Function (pp. 026030) Maleika Heenaye- Mamode Khan, Department of Computer Science and Engineering, University of Mauritius, Mauritius Naushad Ali Mamode Khan, Department of Mathematics, University of Mauritius, Mauritius 5. Architecture of Network Management Tools for Heterogeneous System (pp. 031-040) Rosilah Hassan, Rozilawati Razali, Shima Mohseni,Ola Mohamad and Zahian Ismail Department of Computer Science, Faculty of Information Science and Technology, Universiti Kebangsaan Malaysia, Bangi, Selangor, Malaysia 6. A Topological derivative based image segmentation for sign language recognition system using isotropic filter (pp. 041-045) M. Krishnaveni, Department of Computer Science, Avinashilingam University for Women, Coimbatore, India. Dr. V. Radha, Department of Computer Science, Avinashilingam University for Women, Coimbatore, India 7. A Framework for Validation of Object-Oriented Design Metrics (pp. 046-052) Devpriya Soni, Department of Computer Applications, Maulana Azad National Institute of Technology (A Deemed University), Bhopal 462007 India Ritu Shrivastava,& M. Kumar, SIRT, Bhopal (India) 8. A New Image Steganography Based On First Component Alteration Technique (pp. 053-056) Amanpreet Kaur, Renu Dhir, and Geeta Sikka, Department of Computer Science and Engineering, National Institute of Technology, Jalandhar, India

9. Evaluating Effectiveness of Tamper-Proofing on Dynamic Graph Software Watermarks (pp. 057063) Malik Sikandar Hayat Khiyal, Aihab Khan, Sehrish Amjad, Department of Computer Science, Fatima Jinnah Women University, The Mall, Rawalpindi, Pakistan M. Shahid Khalil, Department of Mechanical Engineering, University of Engineering, Texila. 10. A Novel Trigon-based Dual Authentication Protocol for Enhancing Security in Grid Environment (pp. 064-072) V. Ruckmani, Senior lecturer, Department of Computer Applications, Sri Ramakrishna Engineering College, India Dr G Sudha Sadasivam, Professor, Department of Computer Science and Engineering, PSG College of Technology, Coimbatore, India 11. Design and Analysis of a Spurious Switching Suppression Technique Equipped Low Power Multiplier with Hybrid Encoding Scheme (pp. 073-078) S.Saravanan, Department of ECE, K.S.R.College of Technology, Tiruchengode-637215, India. M.Madheswaran, Department of ECE, Muthayammal Engineering College, Rasipuram-647408, India 12. Using Sloane Rulers for Optimal Recovery Schemes in Distributed Computing (pp. 079-083) R. Delhi Babu, Department of Computer Science & Engineering, SSN College of Engineering, Chennai, India P. Sakthivel, Department of Electronics& Communication Engineering, Anna University, Chennai, India 13. ICD 10 Based Medical Expert System Using Fuzzy Temporal Logic (pp. 084-089) P.Chinniah, Research Scholar, Department of ECE, CEG, Anna University, Chennai, INDIA. Dr.S.Muttan, Professor, Centre for Medical Electronics, CEG, Anna University, Chennai, India 14. DNA-MATRIX: a tool for constructing transcription factor binding sites Weight matrix (pp. 090092) Chandra Prakash Singh, Department of Computer Sciences, R.S.M.T., U.P. College, Varanasi (India). Feroz Khan, MSB Division, Central Institute of Medicinal & Aromatic Plants (CSIR), Lucknow (India) Sanjay Kumar Singh, Department of Computer Sciences, R.S.M.T., U.P. College, Varanasi (India). Durg Singh Chauhan, Institute of Technology, B.H.U., Varanasi (India). 15. Multiprocessor Scheduling For Tasks With Priority Using GA (pp. 093-100) Dr.G.Padmavathi, Professor and Head, Dept.of Computer Science, Avinashilingam University for Women, Coimbatore – 43, India. Mrs.S.R.Vijayalakshmi, Lecturer, School of Information Technology and Science, Dr.G.R.D College of Science, Coimbatore -14, India. 16. Measurement of Nuchal Translucency Thickness for Detection of Chromosomal Abnormalities using First Trimester Ultrasound Fetal Images (pp. 101 -106) S. Nirmala, Center for Advanced Research, Muthayammal Engineering College,Rasipuram V. Palanisamy, Info Institute of Engineering, Kovilpalayam, Coimbatore – 641 107. 17. An Improved Image Mining Technique For Brain Tumour Classification Using Efficient Classifier (pp. 107-116) P. Rajendran, Department of Computer science and Engineering, K. S. Rangasamy College of Technology, Tiruchengode-637215, Tamilnadu, India. M. Madheswaran, Center for Advanced Research, Department of Electronics and Communication Engineering, Muthayammal Engineering College, Rasipuram – 637 408, Tamilnadu, India. 18. Mining Spatial Gene Expression Data Using Negative Association Rules (pp. 117-120) M. Anandhavalli, & M. K. Ghose, Department of Computer Science Engineering SMIT Majitar, India K. Gauthaman, Department of Drug Technology Higher Institute of Medical Technology, Derna, Libya

19. Hierarchical Route Optimization by Using Tree information option in a Mobile Networks (pp. 121-123) K. K. Gautam & Menu Chaudhary, Department of Computer Science & Technology, Roorkee Engineering & Management Technology Institute, Shamli-247 774 (INDIA) 20. Seeing Beyond the Surface: Understanding and Tracking Fraudulent Cyber Activities (pp. 124135) Longe O. B. & Mbarika V., Int. Centre for IT & Development, Southern University, Baton Rouge, LA 70813 Kourouma M, Dept. of Computer Science, Southern University , Baton Rouge, LA 70813 Wada F. & Isabalija R, Nelson Mandela School of Public Policy, Southern University, Baton Rouge, LA 70813 21. On the Efficiency of Fast RSA Variants in Modern Mobile Phones (pp. 136-140) Klaus Hansen, Troels Larsen, Kim Olsen, Department of Computer Science, University of Copenhagen, Denmark 22. An Efficient Inter Carrier Interference Cancellation Schemes for OFDM Systems (pp. 141-148) B. Sathish Kumar, K. R. Shankar Kumar, R. Radhakrishnan Department of Electronics and Communication Engineering, Sri Ramakrishna Engineering College, Coimbatore, India. 23. High-Precision Half-Wave Rectifier Circuit In Dual Phase Output Mode (pp. 149-152) Theerayut Jamjaem, Department of Electrical Engineering, Faculty of Engineering, Kasem Bundit University, Bangkok, Thailand 10250 Bancha Burapattanasiri, Department of Electronic and Telecommunication Engineering, Engineering Collaborative Research Center, Faculty of Engineering, Kasem Bundit University, Bangkok, Thailand 10250 24. Internal Location Based System For Mobile Devices Using Passive RFID And Wireless Technology (pp. 153-159) A. D. Potgantwar, Lecturer in Computer Engg, SITRC Nashik India Vijay M.Wadhai, Prof & Dean Research, MITSOT, MAE, Pune India 25. High-Precision Multi-Wave Rectifier Circuit Operating in Low Voltage + 1.5 Volt Current Mode (pp. 160-164) Bancha Burapattanasiri, Department of Electronic and Telecommunication Engineering, Engineering Collaborative Research Center, Faculty of Engineering, Kasem Bundit University, Bangkok, Thailand 10250 26. Classifying Application Phases in Asymmetric Chip Multiprocessors (pp. 165-170) A. Z. Jooya, Computer Science dept., Iran University of Science and Technology, Tehran, Iran M. Analoui, Computer Science dept., Iran University of Science and Technology, Tehran, Iran 27. Syllable Analysis to Build a Dictation System in Telugu language (pp. 171-176) N. Kalyani, Assoc. Prof, CSE Dept, G.N.I.T.S, Hyderabad, India. Dr K. V. N. Sunitha, .Professor & HOD, CSE Dept., G.N.I.T.S, Hyderabad, India 28. Sinusoidal Frequency Doublers Circuit With Low Voltage + 1.5 Volt CMOS Inverter (pp. 177180) Bancha Burapattanasiri, Department of Electronic and Telecommunication Engineering, Engineering Collaborative Research Center, Faculty of Engineering, Kasem Bundit University, Bangkok, Thailand 10250 29. Speech Recognition by Machine: A Review (pp. 181-205)

M. A. Anusuya, Department of Computer Science and Engineering, Sri Jaya chamarajendra College of Engineering, Mysore, India S. K. Katti, Department of Computer Science and Engineering, Sri Jayachamarajendra College of Engineering, Mysore, India 30. An Extension for Combination of Duty Constraints in Role-Based Access Control (pp. 206-215) Ali Hosseini, ICT Group, E-Learning Center, Iran University of Science and Technology, Tehran, Iran Mohammad Abdollahi Azgomi, School of Computer Engineering, Iran University of Science and Technology, Tehran, Iran 31. An Improved Approach to High Level Privacy Preserving Itemset Mining (pp. 216-223) Rajesh Kumar Boora, Ruchi Shukla, A. K. Misra Computer Science and Engineering Department, Motilal Nehru National Institute of Technology, Allahabad, India – 211004 32. Call Admission Control performance model for Beyond 3G Wireless Networks (pp. 224-229) Ramesh Babu H.S., Department of Information Science and Engineering, Acharya Institute of Technology Gowrishankar, Department of Computer Science and Engineering, B.M.S. College of Engineering, Satyanarayana P.S, Department of Electronics and Communication Engineering, B.M.S. College of
Engineering, Bangalore, India

33. Efficient Candidacy Reduction For Frequent Pattern Mining (pp. 230-237) Mohammad Nadimi-Shahraki, Faculty of Computer Engineering, Islamic Azad University, Najafabad branch, Iran, & Ph.D. Candidate of Computer Science, University of Putra Malaysia Norwati Mustapha, Faculty of Computer Science and Information Technology,University of Putra Malaysia (UPM), Selangor, Malaysia. Md Nasir B Sulaiman, Faculty of Computer Science and Information Technology, University of Putra Malaysia (UPM), Selangor, Malaysia. Ali B Mamat, Faculty of Computer Science and Information Technology,University of Putra Malaysia (UPM), Selangor, Malaysia. 34. Application of a Fuzzy Programming Technique to Production Planning in the Textile Industry (pp. 238-243) I. Elamvazuthi , T. Ganesan, P. Vasant, Universiti Technologi PETRONAS, Tronoh, Malaysia J. F. Webb, Swinburne University of Technology Sarawak Campus, Kuching, Sarawak, Malaysia 35. The Application of Mamdani Fuzzy Model for Auto Zoom Function of a Digital Camera (pp. 244249) I. Elamvazuthi, P. Vasant, Universiti Technologi PETRONAS, Tronoh, Malaysia J. F. Webb, Swinburne University of Technology Sarawak Campus, Kuching, Sarawak, Malaysia 36. Comparative Evaluation and Analysis of IAX and RSW (pp. 250-252) Manjur S Kolhar, Mosleh M. Abu-Alhaj, Omar Abouabdalla, Tat Chee Wan, and Ahmad M. Manasrah National Advanced IPv6 Centre of Excellence, Universiti Sains Malaysia, Penang, Malaysia

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Genealogical Information Search by Using Parent Bidirectional Breadth Algorithm and Rule Based Relationship
Sumitra Nuanmeesri
Department of Information Technology King Mongkut’s University of Technology North Bangkok, Bangkok, Thailand

Chanasak Baitiang
Department of Applied Science King Mongkut’s University of Technology North Bangkok, Bangkok, Thailand

Payung Meesad
Department of Information Technology King Mongkut’s University of Technology North Bangkok, Bangkok, Thailand

Abstract—Genealogical information is the best histories resources for culture study and cultural heritage. The genealogical research generally presents family information and depict tree diagram. This paper presents Parent Bidirectional Breadth Algorithm (PBBA) to find consanguine relationship between two persons. In addition, the paper utilizes rules based system in order to identify consanguine relationship. The study reveals that PBBA is fast to solve the genealogical information search problem and the Rule Based Relationship provides more benefits in blood relationship identification. Keywords-Genealogical Information; Search; Rule Based; algorithm; Bidirectional Search; Relationship

capturing genealogical information [2]. The WINGIS is a Window-based system which allows users to enter the genealogical information in a database and search for people according to name, date of birth or ID number (unique to WINGIS). It currently contains information on more than 550,000 South Africans ranging from 1615 onwards. It is obvious that, searching and browsing information in such a large database are the big problems, especially when many people have the same or similar surnames. Moreover, the visualization of large information spaces is also a common problem. The information needs to be displayed in a meaningful way in order to facilitate analysis and identification process. Family tree, however, are unique and do not correspond exactly to hierarchical data structures. Zoomable user interfaces, called ZoomTree [3,4], could facilitate the dynamic exploration and browsing of family trees within WINGIS [7]. It has been, therefore, used to address theses problems in similar situations, especially with hierarchical data [1, 8]. Yen and Chen explained the design theory and implementation ideas of GIS for Chinese genealogy. Method of using the XML technology to create the metadata of genealogy, maintaining relations among the individuals, and develop management and visualization utilities to represent pedigree information will be introduced [5, 6]. Expert System technology is applied in genealogy search. C Language Integrated Production System (CLIPS) is a productive development and delivery expert system which provides an environment for the construction of rule in the CLIPS library. PHLIPS is integrated software that provides PHP with a basic interface to a CLIPS library in order to help individuals to tract their member history and find relatives [9]. Ontology is a new alternative method to represent the consanguine relationship of family tree. Members in a family are represented by nodes and their consanguine relationships are represented by edges [10]. Ontology based knowledge is proposed prototype system for manage model for Chinese Genealogical Record of Very Important Persons in Nationalist Party and Communist Party of China [11].

I. INTRODUCTION Due to a rapid increase in the world population, therefore genealogical information typically contain large amount of families’ information and relationship between members. Generally genealogical information of a house includes first name, last name, ID, date of birth, sex, father’s ID, and mother’s ID. It is not difficult to know weather or not two people in a house have the consanguine relationship. Example, when a daughter gets married and moves to live with her husband, she changes her last name. Her data information in her parent’s house will be removed. Her information is in the new house (her husband house).When she has a daughter; her daughter might do the same thing, get married and move away. This moving of daughters occurs again and again. The problem arises when we want to know consanguine relationships of two people who have not live in the same house. There are two reasons that make this a hard problem. The first reason is that personal information is contained in only one house, the other reason is that more houses and more information is generated each year in the country. When investigating the consanguine relationships of two people, the problem knows what the short of relationship they have. It starts from the first person nobody knows the short of direction to link between the two people. This paper proposes to develop algorithms to find the consanguine relationship of two people. WINGIS is a genealogical information system (GIS) developed at the University of Port Elizabeth (UPE) for

1

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Most review paper presented a family tree by using different presented technique, however, the papers started from only one person. In this paper, we show Parent Bidirectional Breadth Algorithm (PBBA) to find consanguine relationship between two persons by starting from individuals and try to link them together and Rule Based Relationship of identify name of consanguine relationship. II. PROBLEM FORMULATION

we describe detail of the methodology applied to find the consanguine relationship and rule to identify the relationship between two people.

Everybody know their relative, such as uncle, aunt, niece, cousin and grandparent, however, they might not know all members of their families. Moreover, information in house does not include all family relationship information. Each house includes information of all people who live in the house. Each person’s information consists of their first name, last name, gender, date of birth, father ID and Mother ID. Thus, other people in their family (for example uncle aunt grandparent) are not included in the information of their house. The problem rises when we want to investigate consanguine relation between two persons. There are three questions that are whether or not two persons are in the same family, what kind of the consanguine relationship do they have and how fast can one find the consanguine relationship. Those questions can be answered by using database that is present hereafter. In this section, we represent problem related to the consanguine relationship. Given is a directed network G = (N, A). Node i represents people. Directed arcs (i, j) represent consanguine relationship from parent (i) to son or daughter (j). The name of relationship depends on gender of two people. If node i and node j are nodes at the starter and the end of a directed arc. Node i is a parent and node j is an offspring. If gender of node i is male and the gender of node j is female. The arc will represent relationship from a father to his daughters shown in Figure 1. When a person has father, mother, and three children, the network is depicted in Figure 2. Because there are million populations in a country, their network is very complicated, as it can be seen in Figure 3. i j

Figure 3 The complex relationship of population in a country.

III.

METHODOLOGY

In this section, we present PBBA to find consanguine relationship between two people and the Rule Based Relationship that is used to identify the relationship name. Parent Bidirectional Breadth Algorithm (PBBA) is developed by combine the Bidirectional approach and Breadth First Search and we define parent direction that is the only important direction for search. So, the algorithm is faster than the original Bidirectional approach and Breadth First Search for finding consanguine relationships. Let S be searching set that the program use to find the consanguine relationship, P is a person node, fp is the father of person P and mp is mother of person P.xp Represents kind of sex for person P, Ap is age of person P, B is binary variable. It equals one when a computer finds the consanguine relationship, it is zero if it is otherwise. The dept of searching is Lmax. Lc is current level for searching. In general, information of people includes father, mother, sex, and date of birth, and age of people can be calculate by using date of birth. Figure 4 shows procedure of Parent Bidirectional Breadth Algorithm. After we discover that two people have the same blood relationship, we should identify the name of relationship too. That is the name of relationship depends on the level of generation, kind of sex, and age of person. Because of this, we can adapt rule according to countries. In this paper, we present a rule based system using an English rule. First we classify people by two classes. The one class is descendant generation. That means offspring in family. The other one class is ancestor generation. That means the person who has previous generation of the lower class, for example, parent, grandparent, great-grandparent, etc.

Figure 1 The relationship from node i to node j.

Figure 2 The direction from one node to the other nodes.

After we find an existence of the consanguine relationship between two persons, we should identify what kind of the relationship is there. We record the direction that is used to link the consanguine relation between them. In the next section,

The level is distance that count from one node to the intersection node (linking node). We compare between two people that want to know there relationship. The person node that has the level of generation more than the other person node is the descendant generation. Otherwise, the other person node is Ancestor generation. After that, the descendant generation

2

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

uses name followed Figure 5 and the Ancestor generation uses name followed Figure 6.

Figure 5. That is the fist person might be grand father or grand mother if the connection between them is direct. On the other hand the first person might be a grand uncle or grand aunt if the connection is indirect. Form this example the computer compares data from the parent of the parent, so the second person has the same ID number with first person. The connection is direct; in addition, the computer checks what kind of data is stored about the first person. When the sex is a female, the first person is a grandmother. The second person uses the same technique with Figure 6. Kind of sex of the second person is male. This case the second person is a grandson.

Figure 5 The name of relationship for descendant generation.

Figure 6 The name of relationship for ancestor generation.

IV.

EXPERIMENTAL RESULT

Figure 4 The procedure of Parent Bidirectional Breadth Algorithm.

In brief, the family relationship identification starts from finding the level of generation. Descendant generation uses rule from Figure 5 and ancestor generation uses Figure 6. From Figure 6, Line represents directly connection of two people. We use the line if ID of the parent is descendant generation which equals to the ID of ancestor generation. Form the family member who does not have the direct parent relationship, we use dash line to present. For example, if first and second person are different level of generation. First person is upper level and the second person is lower. The different level of two people is two levels. First person use Figure 6 and the second person use

This section performance of PBBA is show by two ways. The first way is complexity. That is used to explain the difficult of the problem. If this algorithm is high performance, it should find the solution in a short computational time. We show performance of algorithm by proof mathematic. The second way is correctness of finding name of consanguine relationship. We test correctness by generating family data .Uniform distribution is use to generate data. The effectiveness of the PBBA is present by comparing with the effectiveness of BFS that is well known. The effectiveness of BFS is presented. The algorithm starts from the first person node that wanted to link relationship to the second person node. From this first node, there are many nodes for choosing to move. They include of father node, mother node, and child nodes. That shows as follow Figure 7.

3

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

relationship between the first and the second person nodes 2Lmax. Total numbers of searching node are ∑ (2 + V ) nodes. If one node spends one unit time, ∑ (2 + V ) nodes spend ∑ (2 + V ) time too.
L
2 Lmax L =1 L

2 Lmax L =1

2 Lmax L =1

L

Figure 7 Relation between current node other nodes.

PBBA performs in this problem with a shorter time. When the direction to search is parent node direction, the first person node has 2 ways for moving .Those are father node and mother node. In the next step, the father node and the mother node also have 2 nodes for moving too. For Lmax steps, there are 2Lmax nodes for moving. For Bidirectional approach, the first person node and the second person node can search concurrently at the same time with the same behavior. The numbers of moving nodes from the first and second nodes are 2X2L nodes at Lmax steps. By the same reason, PBBA use 2 ∑ 2 time to find the solution.
Lmax l =1 L

Figure 8 The alternative node for the second time search.

If the numbers of moving nodes of two algorithms are compared, each node of breath first search algorithm has (2+V) moving nodes and each node of PBBA has 2 moving nodes. In the fact, the bread first search cannot reduce nodes by searching with parent node direction because it could not find the solution. On the other hand, Bidirectional approach can reduce because it also finds the solution. Absolutely, PBBA is faster than the BFS in consanguine relationship between two persons. When we compare the computational time of those two algorithms, we meet that ∑ (2 + V ) times are more 2 ∑ 2 than times. So, PBBA is higher performance than BFS for the consanguine relationship problem.
2 Lmax L =1 L Lmax l =1 L

Table 1 shows the comparison data of the BFS, Bidirectional approach and PBBA. The first column shows level of searching. The second column contains numbers of nodes that are used if the consanguine relationship, which can be found between L levels. The third and the fourth columns contain the number of nodes that are used when the solution can be found between L levels. Figure 10 shows the trend of used nodes for the three algorithms. Figure10 reveals that BFS and Bidirectional approach uses many nodes to find the solution and PBBA used a low amount of nodes to find the solution. The fact that a low amount of nodes spent a short time to calculate the results shows that the PBBA is a best algorithm to solve the consanguine relationship problem. We present correctness of the PBBA by testing algorithm with family data. First, we generate one family data. After that, we use the algorithm to find consanguine relationship from two people from the data. If the results are correct, we add the new family data to the old data. So, the new database includes two families. Then we run the algorithm to test correctness again. If the correctness satisfies, this process runs again until there are five families in the data.

Figure 9 The directions and levels for search.

If V is number of child node, there are 2+V alternative nodes (ways) for moving from current node to the other nodes. For the second time of searching, each node of 2+V has 2+V way to move. There are (2+V)L alternative nodes for moving. The alternative nodes are presented in Figure 8. Trust, there are (2+V)L alternative nodes when the searching time is L. From the first person node to the second person node computer passes the node that is called Intersection node in Figure 9. In worst case, although the first person node and the second person node are at the same level in the family, the difference level between the level of the first person node and the intersection node is Lmax and the difference level between the level of the second person node and the intersection node is Lmax. That is BFS must use level for linking

4

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
TABLE I. THE DIRECTIONS AND LEVELS FOR SEARCH. Bidirectional approach 10 60 310 1560 7810 39060 195310 976560 4882810 24414060

Level 1 2 3 4 5 6 7 8 9 10

BFS 30 780 19530 488280 12207030 305175780 7629394530 4.76837E+12 4.76837E+12 1.19209E+14

PBBA 4 12 28 60 124 252 508 1020 2044 4092

family. The second error is if the computer identifies that they have the some consanguine relationship, but they do not have the same consanguine. In this case, we create one column to check family data. The people who are in the same family have the same value in the checking column. For example, the members of the first family have value one and the members of the second have value two. This column is not load to the memory because it use for checking the correctness by human. When computer identify that two people are not the same family, we can check the correctness by using this column.
TABLE II. THE EXAMPLE OF GENERATED PEOPLE DATA IN DATABASE.

number of node for finding Blood relationship
200000 180000 160000 140000 120000 100000 80000 60000 40000 20000 0 0 2 4 6 Level 8 10 12

Number of nodes

BFS Bi-directional approch PBBA

Figure 10 Comparison the effectives of three algorithms by the increasing of node number.

We start from one man, the computer generates his wife. Absolutely the man and his wife should have the same last name. In addition, the computer generates the next generation. There are V numbers of children. Sex of children is randomly generated. Each offspring has fifty percent to be a man and fifty percent to be women. After that, all children must be married and they must have new offspring. If the offspring is male, he has the last name of his father. When offspring is women, her last name that is changed depends on her husband. Her husband‘s name and his last name are generated randomly. The family is created until limited number of generation. Final we random two people from this data and check consanguine relationship. If all solution is correct. We add new family by the same technique to database and random to check solution again. When all solutions are correct, we add a new family to database again. We check the correctness of the algorithm until there are five families including in database. There are two errors for correctness test. The first one is computer identify that two people have the same family but they have not. In this case, we check by computer routing to link relation between the people. If the computer can link routing completely, two people are the same family. If computer cannot link the routing, two people are not the same

Table 2 shows at first the program starts with 2 descendants per family, only 1 family is created then 5 pairs of each family member will be picked up randomly. Next program will search for their genealogical information. Second step program will create 2 families, each family has 2 descendants then 5 pairs of each family member will be picked up randomly and each pair will be looked up. Next step the all same old processes have to be done according to Table 2. Then test its accuracy 60 times. All tests have shown that even though there family name has changed the search of genealogical information with rule-base relationship is still valid. The last one is the module to find blood relationship between two people from the large population in a database. The study reveals the result is correct 100 percent. V. CONCLUSION AND FUTURE WORK

Parent Bidirectional Breadth Algorithm (PBBA) provides higher effective performances to solve the genealogical information search problem. This method is a beneficial technique to find the connection between two people due to PBBA being able to link relationship between two nodes and parent node only. The parent direction technique helps to minimize unnecessary directions, thus, the computational time of this algorithm is short. In addition, we present Rule Based Relationship for identifying name or type of relationship; we used English rule as an example because it is a good guideline. Readers can adapt it for their own countries. For the future research, we will develop a new algorithm to explore the consanguine relationship of the group of people. The new algorithm can cluster several people into a group regarding their family.

5

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

REFERENCES
[1] [2] M. D. Plessis, “WINGIS”, Department of Computer Science and Information Systems, University of Porth Elizabeth, 2001. S. Pook, et al., “Context and Interaction in Zoomable User Interface”, Proceedings of 5th IEEE International Working Conference on Advance Visual Interface, USA, 2000, pp. 227-231. C. Oosthuizen, “The Design of A ZoomTree Interface for WINGIS”, Department of Computer Science & Information System, University of Port Elizabeth, 2003. J. Wesson, et al., “A ZoomTree Interface for Searching Genealogical Information”, Proceedings of 3rd ACM International Conference on Computer graphics virtual reality, Africa, 2004, pp.131-133. J. H. Yeh, and C. Chen, “Knowledge Management in a Chinese Genealogy Information System”, Proceedings of International Conference on Digital Library, 2003, pp.427-431. J. Lamping, and R. Rao, “Visualizing Large Trees Using the Hyperbolic Information Browser”, Proceedings of ACM Conference on Companion on Human factors in computing systems, USA, 1996, pp.388-389. J. H. Yeh, and C. Chen, “The Design and Implementation Chinese Genealogy Information System”, The College of Engineering, NT.U., 2004, pp.43-48. M. Perror, and D. Barber, “Tagging of Name Records for Genealogical Data Browning”, Proceedings of 6th IEEE/ACM joint International Conference on Digital Library, USA, 2006, pp.316-325. H. Yang, et al., “A PHLIPS-Based Expert System for Genealogy Search”, Proceedings of Conference IEEE Southeast, VA, 2007, pp.165-170.

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10] J. Ying, and H. Dong, “Ontology Based Knowledge Modeling of Chinese Genealogy Record”, Proceedings of 8th IEEE International Workshop on Semantic Computing and Systems, USA, 2008, pp.33-34. [11] H. Dong, S. Yu and Y. Jiang, “Knowledge Representation of Chinese Genealogical Record of VIPs in KMT and CPC”, Proceedings of IEEE International Conference on Hybrid Intelligent and Systems, China, 2009, pp.116-120.

AUTHORS PROFILE Sumitra Nuanmeesir received the B.S. degree (1nd Class Honor) in General Management, M.S. degree in Information Technology. She is currently Ph.D. Student in Information Technology. Her current research interests Information System, Algorithm and Application, Expert System, Database System.

Chanasak Baitiang received the B.Ed. (2nd Class Honor), M.S. degree in Mathematics., and Ph.D. degree in Higher Education. His current research interests Discrete Mathematics, Graph Theory, Applied Linear Algebra, Model of Computation.

Phayung Meesad received the B.S., M.S., and Ph.D. degree in Electrical Engineering. His current research interests Fuzzy Systems and Neural Networks, Evolutionary Computation and Discrete Control Systems.

6

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Web-Based Expert System for Civil Service Regulations:RCSES
Mofreh A. Hogo*
Electrical engineering Technology Dept., Higher Institution of Technology Benha, Benha University, Egypt.
*

Khaled Fouad
Central Lab. for Agricultural Expert Systems (CLAES)

Fouad Mousa
Business management Dept, Faculty of commerce, Assuit university

the corresponding author, work in [2-5] are one of most commercially successful branches of AI [6]. Although there have been reports of ES failures [7] and [8], surveys [9] and [10] show that many companies have remained enthusiastic proponents of the technology and continue to develop important and successful applications. Internet technology can change the way that an ES is developed and distributed. For the first time, knowledge on any subject can directly be delivered to users through a web based ES. Grove [11] provided some examples of web-based expert systems in industry, medicine, science and government and claimed that “there are now a large number of expert systems available on the Internet.” He argued that there are several factors that make the Internet, by contrast to standalone platforms, an ideal base for KBS (knowledge based system) delivery. These factors include: The Internet is readily accessible, web-browsers provide a common multimedia interface, several Internet-compatible tools for KBS development are available, Internet-based applications are inherently portable, and emerging protocols support cooperation among KBS. He also identified several problems in the development of web-based KBS: Keeping up with rapid technological change to servers, interface components, inference engines, and various protocols; and reducing the potential delivery bottleneck caused by communication loads and a limited infrastructure Adams [12] pointed out that “there are numerous examples of expert systems on the web, but many of these systems are small, non-critical systems.” The most successful example is probably the web-based legal expert system reported by Bodine [13], who remarked that “Law firms are collecting hundreds of thousands of dollars in subscription fees from clients who use their question-andanswer advisory services based on the web.”. Contrary to Grove and Adams, Huntington [14] stated that “there are not many ES on the web” due to the fact that the Internet was not created with applications such as expert systems in mind. As a result, the manner in which the web and web browsers interface made it difficult to perform the actions required by ES. Athappilly [15] reported a “dynamic web-based knowledge system for prototype development for extended enterprise.” He suggested that the use of emerging Internet

Abstract— Internet and expert systems have offered new ways of sharing and distributing knowledge, but there is a lack of researches in the area of web-based expert systems. This paper introduces a development of a web-based expert system for the regulations of civil service in the Kingdom of Saudi Arabia named as RCSES. It is the first time to develop such system (application of civil service regulations) as well the development of it using web-based approach. The proposed system considers 17 regulations of the civil service system. The different phases of developing the RCSES system are presented, as knowledge acquiring and selection, ontology and knowledge representations using XML format. XML-Rule-based knowledge sources and the inference mechanisms were implemented using ASP.net technique. An interactive tool for entering the ontology and knowledge base, and the inferencing was built. It gives the ability to use, modify, update, and extend the existing knowledge base in an easy way. The knowledge was validated by experts in the domain of civil service regulations, and the proposed RCSES was tested, verified, and validated by different technical users and the developers’ staff. The RCSES system is compared with other related web based expert systems, that comparison proved the goodness, usability, and high performance of RCSES. Keywords- Knowledge base;Ontology; RCSES; and Civil regulation.

I.

INTRODUCTION

An expert system is a computer program that simulate the problem-solving behavior of a human, it is composed of a knowledge base (information, heuristics, etc.), inference engine (analyzes the knowledge base), and the end user interface (accepting inputs, generating outputs). The path that leads to the development of expert systems is different from that of conventional programming techniques. The concepts for expert system development come from the subject domain of artificial intelligence (AI), and require a departure from conventional computing practices and programming techniques. Expert systems (ES) emerged as a branch of artificial intelligence (AI), from the effort of researchers to develop computer programs that could reason as humans [1]. Many organizations have leveraged this technology to increase productivity and profits through better business decisions the

7

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

technology made the development of multifunctional AI systems relatively easy and less expensive, but that many users in the business arena were unaware of these technologies and their potential benefits. Consequently, the business community was not aware of the value or educated to deploy the potential available from these technologies in making their business more efficient and competitive. The goals of the work presented in this paper are: Firstly to develop a web-based expert system equipped with an integrated knowledge for the regulations of the civil service system in Saudi Arabia, secondly building a knowledge base regarding the rules and regulations of the civil services based XML technique. Thirdly to develop an easy and attractive interface for the user to deal with the system for the introduction of inputs (entering ontology and knowledge base, queries or conditions) and view the results easily. The overall structure of the proposed RCSES system consists of two main units the knowledge base builder unit and the RCSES interface unit. The knowledge base builder unit includes the ontology builder tool, morphological analyzer, and model builder tool. The expert system interface unit includes: data for the user selection, inference engine, the system results, and the reasoning engine. The final form of the proposed RCSES is presented too, with different web pages, the URL for the proposed work is: http://rcses.com. The rest of the paper is organized as: Section 2 introduces the knowledge acquisition and selection. Section 3 presents the knowledge representation of the civil service regulations including the Ontology and the Knowledge Rules representation. Section 4 introduces the RCSES system development stage. Section 5 introduces the results (system usability and testing). Section 6 is reserved for the comparison with other related web based expert systems. Finally section 7 is reserved for the conclusion and future extensions.

expert system. The main reason for this bottleneck is communication difficulties between the knowledge engineer and the domain expert. Knowledge acquisition in the proposed RCSES was not a big problem; where the regulations of the civil service were selected from the web site of the ministry of the civil service then we use it as it is after the validation from the experts in the civil service domain. The source of the regulations is: http://www.mcs.gov.sa. The system is developed based on 17 regulations of the civil service system in Saudi Arabia. More details can be obtained from the web site of the proposed web-based expert system: http://rcses.com. III.
THE KNOWLEDGE REPRESENTATION

After the step of the domain identification and knowledge acquiring from a participating expert of civil service regulation documents, a model for representing the knowledge must be developed. Numerous techniques for handling information in the knowledge-base are available; however, most expert systems utilize rule-based approaches. The knowledge engineer, working with the expert, must try to define the possible best structure. Other commonly used approaches include decision trees, blackboard systems and object oriented programming. Knowledge representation has been defined as "A set of syntactic and semantic conventions that make it possible to describe things. The syntax of a representation specifies a set of rules for combining symbols to form
Ontology

Parent- Regulation name

Child- Regulation Element

..........

Child- Regulation Element

Concept

Concept

Concept

Concept

Property

Property

………

Property

Property

II. KNOWLEDGE ACQUISITION AND SELECTION A precise domain is required by an expert system, the domain must be compact and well organized. The quality of knowledge highly influences the quality of expert system. The knowledge base is the core component of any expert system; it contains knowledge acquired from the domain expert. Building the knowledge base with the help of domain expert is the responsibility of knowledge engineer. The first task of any expert system development is the knowledge acquisition; which is one of the most important phases in the expert system development life cycle. The process of knowledge acquisition is difficult especially in case if the knowledge engineer is unfamiliar with the domain. The goal of knowledge acquisition step is to obtain facts and rules from the domain expert so that the system can draw expert level conclusions. Knowledge acquisition is crucial for the success of an expert system and regarded as a bottleneck in the development of an

Ontology Builder Tool Value

Value

Value

...................

Value

Model
Model of Regulation

Knowledge Base

Rule

..........

Rule Models Builder Tool

Condition

Result

Condition

Result

Concept

Concept Property Value

.............

Concept

Concept Property Value

......................... ...................

Figure 1. Knowledge structure.

8

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

expressions in the representation language. The semantics of a representation specify how expressions so constructed should be interpreted (i.e. how meaning can be derived from a form). In the proposed system RCSES, the knowledge representation methodology uses XML format. Where, two elements of knowledge, ontology and model rules are represented using XML format. The overall knowledge structure is shown in Fig.1. A. Ontology Representation An ontology is defined as ‘‘formal, explicit specification of a shared conceptualization’’. It represents the concepts and the relationships in a domain promoting interrelation with other models and automatic processing. Ontologies are considered to be a proper mechanism to encode information in modern knowledge-intensive applications, so they have become one of the most used knowledge representation formalism. They allow the enrichment of data with semantics, enabling automatic verification of data consistency and making easier knowledge base maintenance as well as the reuse of components. Ontology is a conceptualization of a domain into a human understandable, machine-readable format consisting of entries. The ontology entry has three layers: concepts, properties of concept and values of the properties. The proposed ontology structure is illustrated by a real example for the ontology structure in the proposed RCSES as shown in Fig. 2; which describes the structure of separation regulation in civil service system using XML format. The interpretation of the example as following:
1.Node of “OntParent” represents regulation data and has attribute “ParentName” its value means the regulation name as: “ ‫اﻟﺘﻌﯿﯿﻦ ﻓﻲ‬ ‫.”اﻟﻮظﺎﺋﻒ اﻟﻌﺎﻣﺔ‬ 2.Child nodes “OntChild” represent the decision result for each part in regulation data and has attribute “ChildName” its value means the decision result of the regulation as: "‫."اﻟﺘﻮظﯿﻒ ﻓﻲ وظﺎﺋﻒ اﻟﻤﺮﺗﺒﺔ اﻟﺴﺎدﺳﺔ ﺣﺘﻲ اﻟﻌﺎﺷﺮة - ﻣﺆﻗﺖ‬ 3.Node “OntConcept” represents the concepts in the regulation, and has attribute “ConceptName” its value is the concept names in the selected part in the regulation as:"‫"اﻹﻋﻼن‬ 4.Node “OntVal” represents the values in the selected part in the regulation, it has an attribute “ValueName”, which specifies the values of the OntVal, as "‫."ﯾﻮﺟﺪ إﻋﻼن"," ﻻﯾﻮﺟﺪ إﻋﻼن‬

<KSA_Civil_Ontology> <OntParent ParentName="‫>"اﻟﺘﻌﯿﯿﻦ ﻓﻲ اﻟﻮظﺎﺋﻒ اﻟﻌﺎﻣﺔ‬ <OntChild ChildName="‫>"اﻟﺘﻮظﯿﻒ ﻓﻲ وظﺎﺋﻒ اﻟﻤﺮﺗﺒﺔ اﻟﺴﺎدﺳﺔ ﺣﺘﻲ اﻟﻌﺎﺷﺮة - ﻣﺆﻗ ﺖ‬ <OntConcept ConceptName="‫>"اﻹﻋﻼن‬ <OntVal ValueName="‫>/"ﯾﻮﺟﺪ إﻋﻼن‬ <OntVal ValueName="‫>/"ﻻ ﯾﻮﺟﺪ إﻋﻼن‬ </OntConcept> </OntChild> <OntChild ChildName="‫>"اﻟﺘﻮظﯿﻒ ﻓﻲ وظﺎﺋﻒ اﻟﻤﺮﺗﺒﺔ اﻟﺮاﺑﻌﺔ ﻋﺸﺮة ﻓﻤﺎ ﻓﻮق‬ <OntConcept ConceptName="‫>"ﻗﺮار ﻣﻦ ﻣﺠﻠﺲ اﻟﻮزراء‬ <OntVal ValueName="‫>/"ﯾﻮﺟﺪ ﻗﺮار‬ <OntVal ValueName="‫>/"ﻻ ﯾﻮﺟﺪ ﻗﺮار‬ </OntConcept> </OntChild> </OntParent> </KSA_Civil_Ontology> Figure 2. Sample of ontology structure in XML format.

<KSA_Civil_Regulation> <Model ModelName="‫>"إﻧﮭـﺎء اﻟﺨـﺪﻣﺔ‬ <Rule Name="R1" RegItem="‫ "إﻧﮭـﺎء اﻟﺨـﺪﻣﺔ ﺑﺎﻹﺳﺘﻘﺎﻟﺔ‬NoTrueFinding="0"> <Finding Cpt="‫ "اﻹﺳﺘﻘﺎﻟﺔ‬Prop="Value" Val="‫"ﺗﻘﺪﯾﻢ اﻻﺳﺘﻘﺎﻟﺔ وﻗﺒﻮﻟﮭﺎ‬ Equal="Yes" ExistInWM="No"/> </Rule> <Rule Name="R2" RegItem="‫"إﻧﮭـﺎء اﻟﺨـﺪﻣﺔ ﺑﻄﻠﺐ اﻹﺣﺎﻟﺔ ﻋﻠﻰ اﻟﺘﻘﺎﻋﺪ‬ NoTrueFinding="0"> <Finding Cpt="‫ "طﻠﺐ اﻹﺣﺎﻟﺔ ﻋﻠﻰ اﻟﺘﻘﺎﻋﺪ ﻗﺒﻞ ﺑﻠﻮغ اﻟﺴﻦ اﻟﻨﻈﺎﻣﯿﺔ‬Prop="Value" Val="‫ "ﺗﻘﺪﯾﻢ اﻟﻄﻠﺐ ﻗﺒﻞ ﺑﻠﻮغ اﻟﺴﻦ اﻟﻨﻈﺎﻣﯿﺔ وﻗﺒﻮﻟﮫ‬Equal="Yes" ExistInWM="No"/> </Rule> </model> </ KSA_Civil_Regulation> Figure 3. Sample of rules structure in XML format

The knowledge base builder tool

Ontology builder tool

Morphological analyzer Domain expert Model builder tool

Knowledge Base

The proposed system developed a tool to build the domain ontology, which is easy to use and dynamic one for the acquired knowledge. B. Rules of Knowledge Representation The knowledge can be formulated as shown in the following simple statements: IF the ‘traffic light’ is green THEN the action is go, as for example: IF the ‘traffic light’ is red THEN the action is stop. These statements represented in the IFTHEN form are called production rules or just rules. The term ‘rule’ in artificial intelligence, which is the most commonly type of knowledge representation, can be defined as IF-THEN

Expert system interface
Data of user selection

Inference engine

User

System result Reasoning engine

Figure 4. Structure of the proposed system RCSES.

9

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Figure 5. (a): Inference engine block diagram. Start

Yes

Check: do the values of model, rule, and finding exist?

No

Create instance from class “XmlDocument” Use method “Load” to open the KB file Display Message, the values not exist

Locate all nodes whose name of model Use “CreateElement” method to create element of rule, and Finding Use “AppendChild” method for each element rule to create a child of the model Check if the model name is existed? Yes Locate all nodes with the same model name Use “SetAttribute” method for each element rule to create a attribute of the rule Check if the rule names for model exist? Use method “Save” to save changes in XML file Yes No Display KB in XML file Check if the rule finding is existed?

End

Yes

Figure 5.

(b): The algorithm for the inference engine.

10

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

structure that relates given information or facts in the IF part to some action in the THEN part. A rule provides some description of how to solve a problem. Rules are relatively easy to create and understand. Any rule consists of two parts: the IF part, called the antecedent (premise or condition) and the THEN part called the consequent (conclusion or action). The basic syntax of a rule is: IF < antecedent>THEN < consequent> . The rules in XML format have a different structure with the previous meaning but in different format. Sample of rule built in the proposed RCSES is shown in Fig.3; it can be interpreted as following: 1. <KSA_Civil_Regulation>; represents the regulations’ in the domain of the civil regulations. total

2.The node of “Model” represents a regulation data and has attribute “ModelName” its value takes the regulation name as “‫.”إﻧﮭـﺎء اﻟﺨـﺪﻣﺔ‬ 3.The child nodes “Rule” represent the decision rule for each part in regulation; takes R1, R2,…and so on. 4.The node attribute “RegItem” represents the consequent in the selected rule in the regulation. 5.The node “Finding” represents antecedent in the rule. 6.The condition equal to attribute “Cpt” value as concept, attribute “Prop” value as property, and attribute “Val” value as value. The condition for example equal to separation “‫ﺗﻘﺪﯾﻢ اﻻﺳﺘﻘﺎﻟﺔ =اﻻﺳﺘﻘﺎﻟﺔ‬ ‫.”وﻗﺒﻮﻟﮭﺎ‬

Figure 6. Home page of the proposed RCSES.

The proposed system RCSES, developed a tool component to build the domain model rules for the acquired knowledge in an efficient, easy, and dynamic way for changing and/or modifying the rules as required. IV.
RCSES DEVELOPMENT

The development of the proposed RCSES system includes the development of different sub-systems. Fig.4 shows the block diagram of the entire proposed system RCSES. The following section describes the different sub-systems in the proposed RCSES system as well as its functionality in the proposed system RCSES: 1. The knowledge builder unit: Developed to enable the expert or knowledge engineer to enter and save domain ontology and domain rules in an efficient and easy way. 2. The morphological analyzer unit: Developed to check if the entered words or sentences in the domain model rules are found in the ontology stored in the knowledge base or not; to avoid the unwanted reasoning results or the results concluded due to the user inputs errors and mistakes. 3. Web page interface of the proposed RCSES system: For using purposes includes entering ontology, knowledge, and the use of the expert system by entering the questions and get the output results with reasoning too. 4. The entire control and operation of the system is done by the inference engine; that is developed using Visual

Figure 7. Web page for entering the ontology & knowledge base.

Basic dot net (VB.NET); which handles the knowledge in format of XML to get the result from the XML file (Knowledge base) that stores the knowledge rules. Fig. 5 shows the block diagram of inference engine. The main roles of the inference engine are summarized as: It applies the expert domain knowledge to what is known about the present situation to determine new information about the domain. The inference engine is the mechanism that connects the user inputs in the form of answers to the questions to the rules of knowledge base and further continues the session to come to conclusions. This process leads to the solution of the problem. The inference engine also identifies the rules of the knowledge base used to get decision from the system and also forms the decision tree. A. Reasoning Mechanism The reasoning mechanism consists of three main components namely: working memory manager (WM manager), XML matcher, and result browser. Fig.5 (a)

11

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

depicts the main components of the reasoning mechanism, while Fig.5 (b) illustrates the inference engine flowchart. • WM Manager Component: The WM manager interacts with the user-friendly interface to get the concepts and its properties as well as the values of those properties by using communication model. The user-friendly interface permits the user to edit his complaints easily. This complaint is considered as a user finding. When the user selects concept, property, and value to be entered in the working memory, the WM manager creates an XSL query statement, which represents these findings. • XML Matcher Component: In this methodology the rule is succeeded when all its child nodes are existed in working memory. This is achieved when the attribute ‘ExistInWM’of every child node is set to “Yes”. So, the matcher gets those succeeded rules by comparing the value of the attribute ‘NoTrueFindings’ of every parent node and the number of the child nodes in this rule and select the matched one. The succeeded rules are store in the result store for later use by display result component. • Result Browser Component: Result Browser component gets the value of the attribute disorder for every rule in the result store, and pass value to the user interface by using communication model to display it to the user. V. RESULTS OF THE PROPOSED SYSTEM RCSES The proposed RCSES was evaluated with different users, including “RCSES” developers’ staff, and technical people. The system is validated by experts in the field of Civil service regulations. Tests of the system were carried out by the developers to make sure the system would work correctly as well as the RCSES is web based expert system, another validation and evaluation for the RCSES will be carried out through the using through the web and the feedbacks from the users will be considered for any comments and modifications. This output results from the developed RCSES will be interpreted via three parts as: The home page (http\\www. rcses.com) of the system, the web page for ontology building and knowledge building of the civil service regulations, and the web pages of the using the RCSES as shown in Figures 6,7,8, 9, and 10. Description of these units and sub-units of web pages are as following: • The home web page is illustrated in Fig.6, which represents the user and expert interface to the RCSES system. • The web page for the ontology and knowledge base builders is shown in Fig. 7; it assists the ability to the experts in entering the knowledge base ontology from the web page and the modification or extension of the knowledge base ontology. • The web page for the ontology builder is shown in Fig. 8(a), while Fig. 8(b) shows the web page for the knowledge base builder; it assists the ability to enter the knowledge base from the web page and the modification or

extension of the knowledge base. • The web page for using the RCSES system that is shown in Fig. 9; it assists the ability to use the RCSES system, by entering the question (select the attributes, then selection the attributes values, continue until select all or the majority of the attributes then find the expected results and decisions. Finally you can see the reasoning; which can be viewed in form of HTML). • Samples from the use of the RCSES and the corresponding output results, and for more details about the use of the RCSES system visit the rcses.com, we presented a Case of Employing in the governmental jobs ‫اﻟﺘﻌﯿﯿﻦ ﻓﻲ اﻟﻮظﺎﺋﻒ‬ " ‫ ," اﻟﻌﺎﻣﺔ‬the steps for selecting the attributes and the conclusion inferred are illustrated through Figures 10(a..j); this case describes the all steps needed from the starting selecting the attributes, followed by the selection of the attributes vales, followed by the conclusion and the expected results, and the sure results. The system presented in this paper RCSES was not easy to be developed, but after the development it can be used by anyone with Internet access and a web browser easily. The web based ES made the evaluation and maintenance of “RCSES” easier than a conventional ES. There is no need to install the system in advance. It is easy to collect feedback. Visitors can be easily traced and analyzed, by collecting information, it was possible to profile the users and determine the value of the system. Compared with traditional ES development tools, the web design software simplifies the user interface design. XML-based user interfaces allow the incorporation of rich media elements. Hyperlinks provide an extra facility in enhancing ES explanation and help functions; users can access the relevant web site easily. The WWW helps in acquiring the knowledge needed in constructing the knowledge base. Any knowledge updating and maintenance can be handled centrally. Useful links are incorporated to help the user understand and interpret the ES recommendations. The e-mails, feedback forms and other Internet communication functions allow users to question and comment on the system. VI.
COMPARISON WITH OTHER RELATED WBES

The comparisons with other similar web based expert systems are carried from different points of views as following: 1. KR: the Knowledge Representation, 2. KBR: Knowledge Base Repository, 3. KM: Knowledge Modification, 4. KB: Knowledge Builder, 5. IR: Inference Engine, 6. EP: Execution Platform, 7. PE: Programming Environment, 8. RT: Response Time for results seen. From our searches of the literature and the presented comparison, there is no reporting on the topic, there is no any

12

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Figure 8. (a): Page for entering the ontology.

Figure 10. (a): Real example using RCSES step 1.

Figure 8. (b): Page for entering the Knowledge base.

Figure 10. (b): Real example using RCSES step 2.

Figure 9. Using the RCSES system for questions and inferring.

Figure 10. (c): Real example using RCSES step 3.

13

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Figure 10. (d): Real example using RCSES step 4.

Figure 10. (g): Real example using RCSES step 7.

Figure 10. (e): Real example using RCSES step 5.

Figure 10. (h): Real example using RCSES step 8.

Figure 10. (f): Real example using RCSES step 6.

Figure 10. (j): Real example using RCSES step 9.

14

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
TABLE I. COMPARISON ANALYSIS WITH 6 RELATED WEB B ASED EXPERT SYSTEMS. KB IE EP PE RT to show result *****The proposed presented work: RCSES***** By backward Rules were built Very easy to a developed chaining .NET in XML format XML modify or change toolWeb-Based Fast ,depends on technology as required Knowledge XML parser Builder) 1. Web-based weather expert system (WES) for Space Shuttle launch [16] Rules are built in Backward Boolean Difficult, by built in Java Java servlet with Java chaining Web-Based Need to download data as inputs object programmer program technology program. technique 2. Web-based expert system for advising on herbicide use in Great Britain [17] Prolog-based Prolog KB Difficult, By Perl, and Takes more time (reads KB files Prolog KB Prolog inference Web-Based files programmer HTML & construct HTML engine 3. Dr. Wheat: A Web-based Expert System for Diagnosis of Diseases and Pests in Pakistani Wheat [18] Forward and e2gLite Little difficult, e2gLite™ Rather fast, because interface is Form of IF-THEN backward knowledge using e2gLite expert Web-Based JAVA built by java that processes the rules chaining base shell system shell built knowledge technique 4. A Novel Web based Expert System Architecture for On-line and Off-line Fault Diagnosis and Control (FDC) of Power System Equipment [19] Upload the Rather slow because it depends is developed Form of Text files for Knowledge base Developed .NET on saving data in data base and using Web-Based production rules, if-else rules. files in text Shell technology upload the knowledge base after VB.NET format each updating 5. Web-based expert system for food dryer selection [20] backward Takes more times ,create the Depends on chaining, by Rule-based By Developed Java servlet HTML forms for presenting three ReSolver KB ReSolver Web-Based knowledge ReSolver technology intermediate and final results modules inference and help screens mechanism 6. Knowledge Representation and Reasoning Using XML [21] backward Very hard as As text file chaining Visual basic 6.0 using the XML XML building from represents Client-side reducing the response time ,Depends on with XML start XML format XML parser KR KBR KM

web-based ES on the civilian service regulations on the web. There also appears to be a lack of a general methodology for developing web-based expert systems. The proposed system provides better performance as the KR was done by XML format as well as the XML ontology format, the homogeneity between the KR and the platform on the WEB so the performance is very high compared with other systems that needs a mediator to match between the knowledge representation methods and the web platform representation or to load the knowledge from any file or by executing a programs as presented in the comparison. The dynamic property for changing and modifying the Knowledge as needed in a fast and very easy way. VII. CONCLUSION The work presented in this paper tries to overcome the general lack of research in the area of web-based expert systems (WBES). The paper addressed the issues associated with the analysis, design, development, and use of web-based expert system for the regulations of the civil service system in the K.S.A “RCSES”. It is the first time to develop such system; which is web based, with the new methodology, the ontology,

the knowledge representation, and the tools as well. The presented work introduces a comparison with other related WBES. The work considered 17 regulations for the civil service and its ability to modify or updating and the extending of the existing regulations. The “RCSES” was verified, and validated by different technical users and the developers as well to be usable in the real world governmental departments. The research provides benefits to the employees and the ability to solve the contradiction in the confused problems, as well as the providing for the suggested solutions. The developed RCSES is fully implemented to run on the web using ASP.net techniques as the main programming language, and a new server-side technology, XML-Rule-based knowledge sources and the inference mechanisms were implemented using ASP.net. The work presented in this paper can be extended to add another version of the system in English language, and to include the uncertainty using the fuzzy knowledge representation and inference too. Acknowledgment The authors wish to thank Taif University for the financial support of this research and the anonymous reviewers for their valuable comments.

15

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
Mofreh A. Hogo is a lecturer at Benha University, Egypt. He is a lecturer of Computer Science and Engineering. Dr. Hogo holds a PhD in Informatics Technology from Czech Technical University in Prague, Computer Science and Engineering Dept. 2004. He is the author of over 40 papers that have been published in refereed international Journals (Information Sciences, Elsiver, UBICC, IJICIS, IJCSIS, IJPRAI, ESWA, IJEL, Web Intelligence and Agent Systems, Intelligent Systems, international journal of NNW, IJAIT Journal of Artificial Intelligence Tools, IJCI) and Book chapters (Neural Networks Applications in Information Technology and Web Engineering Book, Encyclopedia of Data Warehousing and Mining, and Lecture Notes in Artificial Intelligence Series), and international conferences (Systemics, Cybernetics and Informatics Information Systems Management, IEEE/WIC, IEEE/WIC/ACM, ICEIS). His areas of interest include Digital Image Processing, Multimedia Networks, Intrusion detection, Data Mining, Data Clustering and classification, pattern Recognition, character recognition, fuzzy clustering, artificial Neural Networks, Expert systems, Software Engineering. Khaled Fouad is an assistant lecturer at Central Lab. for Agricultural Expert Systems (CLAES), as he working in his Ph.D program. He interested in Programming, Web design, Internet and its applications Development, and Expert Systems applications Development.

REFERENCES [1] [2] [3] [4] [5] K. Darlington, The Essence of Expert Systems, Prentice-Hall, Essex, England (2000). E. Coakes and K. Merchant, Expert systems: a survey of their use in UK business, Information and Management 30 (1996) (5), pp. 223–230. J.D. Durkin, Expert systems: a view of the field, IEEE Expert 11 (1996) (2), pp. 56–63. T.G. Gill, Early expert systems: where are they now?, MIS Quarterly 19 (1995) (1), pp. 51–81. E. Oz, J. Fedorowicz and T. Stapleton, Improving quality, speed and confidence in decision making: measuring expert system benefits, Information and Management 24 (1993) (2), pp. 71–82. K. Metaxiotis and J. Psarras, Expert Systems in business: applications and future directions for the operations researcher, Industrial Management and Data system 103 (2003) (5), pp. 358–361. R.M. O’keefe and D. Rebne, Understanding the applicability of expert systems, International Journal of Applied Expert Systems 1 (1993) (1), pp. 3–24. B.K. Wong and J.A. Monaco, Expert system applications in business: a review and analysis of the literature, Infxormation and Management 9 `(1995) (3), pp. 141–152. Y. Yoon, T. Guimaraes and Q. O’neal, Exploring the factors associated with expert systems success, MIS Quarterly 19 (1995) (1), pp. 83–106. A.S. Kunnathur, M. Ahmed and R.J.S. Charles, Expert systems adoption: an analytical study of managerial issues and concerns, Information and Management 30 (1996) (1), pp. 15–25. W.D. Potter, X. Deng, J. Li, M. Xu, Y. Wei, I. Lappas, M.J. Twery and D.J. Bennett, A web-based expert system for gypsy moth risk assessment, Computers and Electronics in Agriculture 27 (2000) (1–3), pp. 95–105. J.A. Adams, The feasibility of distributed web based expert systems, Proceedings of the 2001 IEEE International Conference on Systems, Man, and Cybernetics Tucson, AZ, October (2001). L. Bodine, Finding new profits: delivering legal services via web-based expert systems. The LawMarketing Portal, at http://lawmarketing.com/publications/legalmarketingtech/pub34.cfm. D. Huntington, Web-based expert systems are on the way: Java-based web delivery, PC AI Intelligent Solutions for Desktop Computers 14 (2000) (6), pp. 34–36. K. Athappilly, A dynamic web-based knowledge system for prototype development for extended enterprise, Proceedings of the 3rd International Conference on the Practical Applications of Knowledge Management PAKEM 2000, Manchester, April (2000). Rajkumar, T. Bardina, J.E (2003). “Web-based weather expert system (WES) for Space Shuttle launch”, Volume: 5, On page(s): 5040- 5045 vol.5, ISBN: 0-7803-7952-7, Systems, Man and Cybernetics, 2003. IEEE International Conference. T.Alan., W. Ian (2004). “A web-based expert system for advising on herbicide use in Great Britain”, Computers and Electronics in Agriculture 42 (2004) 43–49, Published by Elsevier Science B.V. S. Fahad, R. Saad,I. Kashif, M. Fahad, F. Ahmad, I. Inam, and A. Tauqeer, (2008). “Dr. Wheat: A Web-based Expert System for Diagnosis of Diseases and Pests in Pakistani Wheat”, Proceedings of the World Congress on Engineering 2008 Vol I ,WCE 2008, July 2 - 4, 2008, London, U.K. M. Babita, M. B. Srinivas,and Amit Jain (2008).” A Novel Web based Expert System Architecture for On-line and Off-line Fault Diagnosis and Control (FDC) of Power System Equipment”, 978-1-4244-17629/08.2008 IEEE. S. Haitham , J. Christopher (2003), “Web-based expert system for food dryer selection”, Elsevier Science Ltd. All rights reserved. doi:10.1016/S0098-1354(03)00020-6. B. Khaled, R. Mahmoud, El. Salwa, and T. Khaled. (2004), Knowledge Representation and Reasoning Using XML, 2nd international Artificial Intelligent, Cairo

[6]

[7]

[8]

[9] [10]

[11]

Fouad Mousa is associate professor at Business management Dept,Faculty of commerce, Assuit university.

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

16

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

A wide-range survey on Recall-Based Graphical User Authentications algorithms based on ISO and Attack Patterns
ARASH HABIBI LASHKARI

Computer Science and Information Technology, University of Malaya (UM) Kuala Lumpur, Malaysia .

SAMANEH FARMAND Computer Science and Information Technology (IT), University Malaya (UM) Kuala Lumpur, Malaysia .

DR. ROSLI SALEH

Computer Science and Information Technology, University of Malaya (UM) Kuala Lumpur, Malaysia .

Dr. OMAR BIN ZAKARIA Computer Science and Data Communication (MCS), University of Malaya (UM) Kuala Lumpur, Malaysia
.

Abstract- Nowadays, user authentication is one of the important topics in information security. Text-based strong password schemes could provide with certain degree of security. However, the fact that strong passwords being difficult to memorize often leads their owners to write them down on papers or even save them in a computer file. Graphical user authentication (GUA) has been proposed as a possible alternative solution to text-based authentication, motivated particularly by the fact that humans can remember images better than text. In recent years, many networks, computer systems and Internet-based environments try used GUA technique for their user’s authentication. All of GUA algorithms have two different aspects which are usability and security. Unfortunately, none of graphical algorithms were being able to cover both of these aspects at the same time. This paper presents a wide-range survey on the pure and cued recall-based algorithms in GUA, based on ISO standards for usability and attack patterns standards for security. After explain usability ISO standards and attack patterns international standards, we try to collect the major attributes of usability and security in GUA. Finally, try to make comparison tables among all recall-based algorithms based on usability attributes and attack patterns those we found. Keywords - Recall-Based Graphical User Authentication, Graphical Password, Usability and security, ISO 9241-11, ISO 9126, ISO 13407, Attack Patterns, Brute force, Dictionary attacks, Guessing, Spyware, Shoulder surfing, Social engineering (description).

I. INTRODUCTION In recent years, computer and network security has been formulated as a technical problem. A key area in security research is authentication which is the determination of whether a user should be allowed access to a given system or resource. In this respect, the password is a common and widely authentication method still used up to now. The use of passwords goes back to ancient times when soldiers guarding a location by exchange a password and then only allow a person who knew the password. In modern times, passwords are used to control access to protect computer operating systems, mobile phones, auto teller machine (ATM) machines, and others. A typical computer user may require passwords for many purposes such log in to computer accounts, retrieving e-mail from servers, accessing to files, databases, networks, web sites, and even reading the morning newspaper online. In graphical password, the problem arises because passwords are expected to have two fundamentals requirements: i. Password should be easy to remember. ii. Password should be secured. In a graphical password system, a user needs to choose memorable image. The process of choosing memorable images depends on the nature of the process of image and the specific sequence of click locations. In order to support memorize ability, images should have meaningful content because meaning for arbitrary things is poor.

.
17 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

II. RECALL-BASED ALGORITHMS A literature on most of articles regarding graphical password techniques from 1994 till 2009 shows that graphical password can be categorized into pure recall, cute recall and recognition groups. This section tries to explain Pure and cute Recall-Based algorithms by focusing on their lacks and weaknesses. 1. Passdoodle (Pure recall) Passdoodle is a graphical password comprised of handwritten designs or text, usually drawn with a stylus onto a touch sensitive screen. In their 1999 paper, Jermyn et al. prove that doodles are harder to crack due to a theoretically much larger number of possible doodle passwords than text passwords [18] . Figure 1 will be shown a sample of Passdoodle password.

easier than DAS Password. The other weakness is that, the users tend to choose frail graphical passwords that are vulnerable to the graphical dictionary attack [11]. 3. Grid Selection (Pure recall) In 2004, Thorpe and van Oorschot further studied the impact of password length and stroke-count as a complexity property of the DAS scheme. Their study showed that stroke-count has the largest impact on the DAS password space -- The size of DAS password space decreases significantly with fewer strokes for a fixed password length. The length of a DAS password also has a significant impact but the impact is not as strong as the stroke-count. To improve the security, Thorpe and van Oorschot proposed a “Grid Selection” technique. The selection grid is an initially large, fine grained grid from which the user selects a drawing grid, a rectangular region to zoom in on, in which they may enter their password (figure 3). This would significantly increase the DAS password space [20].

Figure 1: An Example of a Passdoodle

Weaknesses: with reference to the [18], they found that people could remember complete doodle images as accurately as alphanumeric passwords, but they were less likely to recall the order in which they drew a doodle than the resulting image. In the other research [24], users were fascinated by the doodles drawn by other users, and frequently entered other users’ login details merely to see a different set of doodles from their own. 2. Draw A Secret (DAS) (Pure recall) In 1999, this method present by allowing the user to drawing a simple picture on a 2D grid as in Figure 2. The interface is consisting of a rectangular grid of size G * G. Each cell in this grid is denoted by discrete rectangular coordinates (x,y). As it can be seen in the figure, the coordinate sequence generated by drawing is [7]: (2,2), (3,2), (3,3), (2,3), (2,2), (2,1), (5, 5). Figure 2 shows a sample of DAS password.

Figure 3: A sample of Grid Selection method

Weaknesses: This method just significantly increases the DAS password space but the lacks of DAS doesn’t solve yet [20]. 4. Qualitative DAS (QDAS) (Pure recall) In 2007, QDAS method designed as an enhancement of DAS method created by encoding each stroke. The raw encoding consists of its starting cell and the sequence of qualitative direction change in the stroke relative to the grid. A direction change is considered when the pen cross a cell boundary in a direction different from direction the cross the previous cell boundary. The research shows that, the image which has more area of interest (Hot Spot) could be more useful as a background image [2]. Figure 4 shows a sample of QDAS password. Figure 4 shows a sample of QDAS password.

Figure 2: Draw a Secret (DAS) method on a 4*4 Grid

Weaknesses: Goldberg in 2002 had a survey which showed that most of the uses forgot their stroke order. On the other hand, he showed that the user can remember text password
18

Figure 4: A sample of Qualitative DAS Algorithm

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Weaknesses: This model uses dynamic grid transformation to hide the process of creating password so this method could be safer that original DAS to shoulder surfing attack. Although this model have more entropy than previous DAS but it has less memorable than the original one [2]. 5. Syukri Algorithm (Pure recall) Syukri algorithm proposes a system where authentication is conducted by having user drawing their signature using mouse. See Figure 5 [21]. This technique includes two stages, namely, registration and verification. During the registration stage, user will first be asked to draw their signature with mouse, and then the system will extract the signature area and either enlarges or scale-down signatures, rotates if needed, (also known as normalizing). The information will later be saved into the database. The verification stage first takes the user input, and does the normalization again, and then extracts the parameters of the signature. The system conducts verification using geometric average means and a dynamic update of database. According to the study [21], the rate of successful verification was satisfying. The biggest advantage of this approach is that there is no need to memorize one’s signature and signatures are hard to fake. Figure 5 shows a sample of Syukri password.

Figure 6: A sample of Blonder method

Weaknesses: A problem with this scheme was that the number of predefined click regions was relatively small so the password had to be quite long to be secure. Also, the use of pre-defined click objects or regions required simple, artificial images, for example cartoon-like images, instead of complex, real-world scenes [22]. 7. PassPoint (cute recall) In 2005, PassPoint created in order to cover the limitation of Blonder Algorithm which was limitation of image. The picture could be any natural picture or painting but at the same time should be rich enough in order to have many possible click points. On the other hand the image is not secret and has no role other than helping the user to remember the click point. Another source of flexibility is that there is no need for artificial predefined click regions with well-marked boundaries like blonder algorithm. The user is choosing several points on picture in a particular order [16]. Figure 7 shows a sample of PassPoint password.

Figure 5: A sample of Syukri algorithm

Weaknesses: However, not everybody is familiar with using mouse as a writing device; the signature can therefore be hard to drawn. One possible solution to this problem would be to use a pen-like input device, but such devices are not widely used, and adding new hardware to the current system can be expensive. In this study, researchers believed such technique is more useful to small devices [21]. 6. Blonder (cute recall) In 1996, this method designed by Greg E. Blonder which a pre-determined image presented to the user on a visual display and user should be point to one or more predetermined positions on the image (tap regions) in a predetermined order as a way of point out his or her authorization to access the resource. Originator maintained that the method is secure according to a millions of different regions. Figure 6 shows a sample of Blonder password. Figure 6 shows a sample of Blonder password.

Figure 7: A sample of Passpoint method

Weaknesses: Users in PassPoint system were able to easily and quickly create a valid password, but they had more difficulty learning their passwords than alphanumeric users, taking more trials and more time to complete the practice, On the other hand the login time, in this method is longer than alphanumeric method [16]. 8. Background DAS (BDAS) (cute recall) In 2007, this method proposed by adding background image to the original DAS for improvement, so that both background image and the drawing grid can be used to providing cued recall [11]. The user starts by using three different ways: i. The user have secret in mind to begin, and then draw using the point from a background image.

19

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

ii. The user’s choice of secret is affected by various characteristic of the image. iii. A mix of two above methods. Figure 8 shows a sample of BDAS algorithm.

kitchen, bathroom, bedroom or others (See Figure 10). To enter a password, user can click and/or drag on a series of items within that image [20].

Figure 8: A sample of BDAS algorithm Figure 10: A sample of PASSMAP method

Weaknesses: With reference to a research on BDAS, memory decaying over a week is one of the major problems in this algorithm. Users had no problem in recreating it in the five-minute test, but a week later they could not do better than producing the secret password as previous. Also shoulder-surfing and interference between multiple passwords are concerns for BDAS [11]. 9. PASSMAP (cute recall) One of the main problems with passwords is that very good passwords are hard to remember and the one which are easy to remember are too short of simple to be secured. From the studies of human memory, we know that it is relatively easy to remember landmarks on a well-known journey [19]. Figure 9 will be shows a sample of PassMap password.

Weaknesses: There are some disadvantages such as the size of password space is small. There are limited places that one can take vegetables, fruits or food from and put into, therefore causing the passwords to be somewhat guessable or predictable [20]. 11. VisKey SFR (cute recall) VisKey is a recall-based authentication scheme that currently has been commercialized by SFR Company in Germany. This software was designed specifically for mobile devices such as PDAs. To form a password, users need to tap their spots in sequence (Figure 11) [20].

Figure 9: A sample of PASSMAP method

Figure 11: A sample of VisKey SFR method

Weaknesses: Additionally the PassMap technology is not very susceptible to "shoulder surfing" as can be clearly seen from Figure 8. Noticing a single new edge in a large graph or even an absence of some edge in the map is not a trivial task, for someone just passing by. But it is respect to Brute Force attacks while at the same time considering how good those mechanisms are in terms how memorable they are [19]. 10. Passlogix v-Go (cute recall) Passlogix Inc. is a commercial security company located in New York City USA. Their scheme called Passlogix v-Go uses a technique known as “Repeating a sequence of actions” which means creating a password by a chronological situation. In this scheme, user can select their background images based on the environment, for example in the
20

Weaknesses: The problem with this technique is the input tolerance. Since it is difficult to point to the exact spots on the picture, Viskey permits all input within a certain tolerance area around it. The size of this area can be predefined by users. Nonetheless, some precautions related to the input precision needs to be set carefully, as it will directly influence the security and the usability of the password. For a practical setting of parameters, a four spot VisKey can offer theoretically almost 1 billion possibilities to define a password. However, is not large enough to avoid the off-line attacks by a high-speed computer. At least seven defined spots are needed in order to overcome the brute force attacks [20].

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

12. Pass-Go Scheme In 2006, this scheme is an improvement of DAS algorithm which kept the advantages of the DAS plus adding some extra security features to it. Pass-Go is a grid-based scheme which requires a user to select intersections, instead of cells so the new system refers to a matrix of intersections, rather than cells as in DAS. As an intersection is actually a point which doesn’t have an area, it would be impossible for a user to touch it without an error tolerance mechanism. Therefore sensitive areas defined to address this problem.

Figure 12: Pass-Go Scheme, 2006

Figure 12: The 17 Parts of ISO 9241

Changing the format of typing from cell to intersection bring the user more free choices. The other difference between these two algorithms is that the size of grid in enhanced method changes to 9*9. III. ISO STANDARDS ON USABILITY The International Organization for Standardization (ISO) developed a variety of models to measure usability, but none of these models cover all usability aspects. This section tries to scrutiny three models in ISO and finally makes a complete table of usability attributes base on these ISO standards. ISO 9241: ISO 9241 is a series of international standards of ergonomics requirements for office work with visual display terminals (Figure 12). The definitions of Part 11 of ISO 9241 are built from a different usability viewpoint. Its key components are: effectiveness that describes the interaction from a process point of view, efficiency that is the attention for results and resources implied and satisfaction that refers to a user point of view. ISO 9241 provides requirements and recommendations concerning hardware, software and environment attributes that contribute to usability, and subjacent ergonomic principles. Parts 3 to 9 deal with hardware design requirements and guidelines that can have implications on software. Parts 10 to 17 deal with software attributes [25].

According to this standard, the measurement of system usability consists of three usability attributes: 1. Effectiveness: How well do the users achieve the goals they set out to achieve using the system? 2. Efficiency: The resources consumed in order to achieve their goals. 3. Satisfaction: How the users feel about their use of the system? ISO 9126: ISO 9126 address software quality from the product point of view. It is probably the most extensive software quality model, even if it is not exhaustive. Initially published in 1991, the approach of its quality model is to present quality as a whole set of characteristics. It divides software quality into six general categories: functionalities, reliability, usability, effectiveness, maintainability and portability (figure 13) [25]. Part four of ISO 9126 defined the usability as "A set of attributes that bear on the effort needed for use and on the individual assessment of such use, by a stated or implied set of users". It proposed then a product oriented usability approach. Usability was seen like an independent factor of software quality. It treated software attributes, mainly its interface that makes it easy to use. As you see in figure 13, the major attributes are: Understandability, Learnability, Operability, and Attractiveness [25].

21

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Table 1: Our finalized Usability attributes from ISO Standards

Figure 13: The 6 Parts of ISO 9126

ISO 13407: This International Standard provides guidance on humancentred design activities throughout the life cycle of computer-based interactive systems. It is aimed at those managing design processes and provides guidance on sources of information and standards relevant to the humancentered approach. For the purposes of this International Standard, there are eight terms and definitions inclusive Interactive system, Prototype, Usability, Effectiveness, Efficiency, Context of use, and user that third term as usability define as below: Extent to which a product can be used by specified users to achieve specified goals with effectiveness, efficiency and satisfaction in a specified context of use [ISO 9241-11]. Finally, The Usability Model that defined by ISO 13407 is comprised of five stages, which are implicitly joined in a loop. Figure 14 decided this model graphically [26].

IV. USABILITY IN RECALL-BASED TECHNIQUES With reference to table 1, now we can make a comparison table among all recall-based algorithms in two categories as pure and cued recall-based algorithm that you can find in tables below (Table 2, 3):

Table 2: The Usability features in Pure Recall-Based Techniques

Figure 14: The 5 Parts of ISO 13407

Finally, after this survey on ISO standards (9241, 9126, 13407) in usability we find more attributes for each feature that you can see in the table below (Table 1).

22

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Table 3: The Usability features in Cued Recall-Based Techniques

Password guessing attacks can be broadly categorised into online password guessing attacks and offline dictionary attacks. In an online password guessing attack, an attacker tries a guessed password by manipulating the inputs of one or more oracles. In an offline dictionary attack, an attacker exhaustively searches for the password by manipulating the inputs of one or more oracles. As many users try to select their password based on their personal information like the name of their pets, passport number, family name and so on, the attacker try to guess. SPYWARE Spyware is a type of malware which installed on computers with the aim of collecting sensitive information of users, using a key logger or key listener. This information gathered without user’s knowledge and report back to an outside source. During graphical password authentication the attacker attempt to gain sensitive information like user names or selected passwords images by intercepting information exchanged.

V. LIST OF ATTACKS ON GUA With reference to the Common Attack Pattern Enumeration and Classification (CAPEC) Standard Abstraction Attack Pattern List (Release 1.3) and other resources of attacks, finally we found six attacks method that is efficient in graphical user authentication (GUA) algorithms [28]. Now, this section tries to explain these attacks methods and then next section tries to make a comparison table among all recall-based algorithms that explained in section II. BRUTE FORCE It is more difficult to carry out a brute force attack against graphical passwords than text-based passwords. The attack programs need to automatically generate accurate mouse motion to imitate human input, which is particularly difficult for recall based graphical passwords. Overall, we believe a graphical password is less vulnerable to brute force attacks than a text-based password. The main defense against brute force search is to have a sufficiently large password space. The speed which an attacker discovers a secret is directly related to the resources that the attacker has. This attack method is resource expensive as the attackers’ chance for finding user’s password is high only if the resources be as complete as possible. DICTIONARY ATTACKS Since recognition based graphical passwords involve mouse input instead of keyboard input, it will be impractical to carry out dictionary attacks against this type of graphical passwords. For some recall based graphical passwords, it is possible to use a dictionary attack but an automated dictionary attack will be much more complex than a text based dictionary attack. More research is needed in this area. Overall, we believe graphical passwords are less vulnerable to dictionary attacks than text-based passwords. GUESSING
23

SHOULDER SURFING Shoulder surfing refers to using direct observation techniques, such as looking over someone's shoulder, to get information. Shoulder surfing is effective in crowded places because it's really easy to stand near someone and watch them entering a PIN number at an ATM machine. This attack is also possible at a distance using vision-enhancing devices like miniature closed-circuit television cameras can be concealed in ceilings, walls or fixtures to observe data entry. To prevent shoulder surfing, it is advised to shield paperwork or the keypad from view by using one's body or cupping one's hand.

SOCIAL ENGINEERING (DESCRIPTION) In this kind of attack an attacker uses human interaction to obtain or compromise information about an organization or computer systems. An attacker possibly claiming to be a new employee, or researcher and even offering credentials to support that identity. However, by asking questions, he or she may be able to piece together enough information to infiltrate an organization's network. If an attacker is not able to gather enough information from one source, he or she may contact another source within the same organization and rely on the information from the first source to add to his or her credibility. VI.
IMPRESSIBILITY ON ATTACKS IN RECALL-BASED ALGORITHMS

Previous part tries to explain and define six major attack patterns in graphical user authentication technology. Now, this part makes a comparison table among all recall-based algorithms based on attack patterns and impressibility of algorithms on attack patterns.

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Table 4: The attacks peruse in Recall-Based algorithms
[5] [6]

[7]

[8]

[9]

[10]

[11]

[12]

VII. CONCLUSIONS In this study, twelve algorithms from Recall-Based explained in two Pure and Cued categories and peruse most of their weaknesses and vulnerabilities. In the first part of research, we found three ISO standard included ISO 9241, ISO 9126 and ISO 13407 and find three main usability categories and then for each category define some sub attributes that you can find in table1. Then, tables 2 and table 3 showed the comparison among all pure and cued recall-based algorithms on our useability founded attributes and sub attributes. In the second part, we found the effective attack patterns on graphical user authentication (GUA) and explained them. Finally, in the last part we made a comparison table among impressibility of all recall-based algorithms based on standard attack patterns. ACKNOWLEDGMENT We would like to express our appreciation to our parents and all the teachers and lecturers who help us to understand the importance of knowledge and show us the best way to gain it. REFERENCES
[1] Ahmet Emir Dirik, Nasir Memon and Jean-Camille Birget, “Modeling user choice in the PassPoints graphical password scheme”, Symposium on Usable Privacy and Security 2007. Pittsburgh, Pennsylvania, USA. ACM. 20-28; July 2007. Di Lin, Paul Dunphy, Patrick Olivier and Jeff Yan, “Graphical Passwords & Qualitative Spatial Relations”, Proceedings of the 3rd symposium on Usable privacy and security. Pittsburgh, Pennsylvania. ACM. 161-162 ; July 2007. Eiji Hayashi , Nicolas Christin, “Use Your Illusion: Secure Authentication Usable Anywhere”, Proceedings of the 4th symposium on Usable privacy and security (SOUPS). Pittsburgh, PA USA, ACM. 35-45, July 2008. Furkan T., A. Ant Ozok, and Stephen H. Holden, ”A Comparison of Perceived and Real Shoulder-surfing Risks between Alphanumeric and Graphical Passwords”, Symposium on Usable Privacy and

[13]

[14]

[15]

[16]

[17]

[18]

[19] [20]

[21]

[2]

[22]

[3]

[23]

[4]

[24]

Security (SOUPS). Pittsburgh, Pennsylvania, USA. ACM. 56-66; July 2006. Greg E. Blonder , Graphical Password U.S. Patent No. 5559961, 1996. Haichang Gao, Xuewu Guo, Xiaoping Chen, Liming Wang, and Xiyang Liu, “YAGP: Yet another Graphical Password Strategy”, 2008 Annual Computer Security Applications Conference. IEEE. 121-129; 2008. Jermyn Ian, A. Mayer, F. Monrose, M. K. Reiter and A. D. Rubin,“The design and analysis of graphical passwords”, Proceedings of the Eighth USENIX Security Symposium. August 23-26 1999. USENIX Association 1–14, 1999. Julie Thorpe, P.C. van Oorschot,“Towards Secure Design Choices for Implementing Graphical Passwords”, Proceedings of the 20th Annual Computer Security Applications Conference. Ottawa, Ont., Canada, IEEE. 50 – 60; Dec 2004. L. Y. POR, X. T. LIM; “Multi-Grid Background Pass-Go”, WSEAS TRANSACTIONS on INFORMATION SCIENCE & APPLICATIONS, ISSN: 1790-0832, Issue 7, Volume 5, July 2008. Paul Dunphy , James Nicholson , Patrick Olivier, “Securing Passfaces for Description”, Symposium on Usable Privacy and Security (SOUPS), Pittsburgh, PA USA; 2008. Paul Dunphy, Jeff Yan, “Do Background Images Improve “Draw a Secret” Graphical Passwords?”, Proceedings of the 14th ACM conference on Computer and communications security. Alexandria, Virginia, USA. ACM. 36-47; 2007. Rachna Dhamija, “Hash visualization in user authentication”, Proceedings of CHI 2000. The Hague, the Netherlands. ACM 279– 280; 2000. Roman Weiss, Alexander De Luca, “PassShapes – Utilizing Stroke Based Authentication to Increase Password Memorability”, Proceedings of the 5th Nordic conference on Human-computer interaction: building bridges. Lund, Sweden. ACM. 2008. 383-392; October 2008. Sacha Brostoff & M. Angela Sasse, “Are Passfaces1 More Usable Than Passwords? (A Field Trial Investigation)”, Department of Computer Science, University College London,London, WC1E 6BT; 2008. Saranga Komanduri, Dugald R. Hutchings, “Order and Entropy in Picture Passwords”, Proceedings of graphics interface 2008. Windsor, Ontario, Canada. Canadian Information Processing Society. 115-122; May 2008. Susan Wiedenbecka, Jim Watersa, Jean-Camille Birgetb and Alex Brodskiyc, Nasir Memon. PassPoints, “Design and longitudinal evaluation of a graphical password system”, Academic Press, Inc. 102-127, July 2005 Xiaoyuan Suo, Ying Zhu and G. Scott. Owen, “Graphical Passwords: A Survey”, Proceedings of the 21st Annual Computer Security Applications. IEEE. 463-472; 2005. Christopher Varenhorst,” Passdoodles; a Lightweight Authentication Method “ , Massachusetts Institute of Technology, Research Science Institute, July 27,2004. Roman V. Yampolskiy, “User Authentication via Behavior Based Passwords”; IEEE Explore, 2007. Muhammad Daniel Hafiz, Abdul Hanan Abdullah, Norafida Ithnin, Hazinah K. Mammi; “Towards Identifying Usability and Security Features of Graphical Password in Knowledge Based Authentication Technique”; IEEE Explore, 2008. Ali Mohamed Eljetlawi, “Study and Develop a New Graphical Password System”, University Technology Malaysia, Master Dissertation, 2008. Susan Wiedenbeck, Jean-Camille Birget, Alex Brodskiy;” Authentication Using Graphical Passwords:Effects of Tolerance and Image Choice”, Symposium On Usable Privacy and Security (SOUPS), Pittsburgh, PA, USA, 2005. Paul Dunphy, Jeff Yan; “Do Background Images Improve “Draw a Secret” Graphical Passwords?”; CCS’07, Alexandria, Virginia, USA, 2007. Karen Renaud, “On user involvement in production of images used in visual authentication” ; Elsevier, Journal of Visual Languages and Computing,2008.

24

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

[25]

[26] [27]

Alain Abran, Witold Suryn, Adel Khelifi, Juergen Rilling, Ahmed Seffah; “Consolidating the ISO Usability Models”, Concordia University, Montreal, Canada International Standard ISO 13407, “Human-centred design processes forinteractive systems”, First edition, 1999-06-01 Common Attack Pattern Enumeration and Classification (CAPEC) Standard Abstraction Attack Pattern List (Release 1.3); http://capec.mitre.org/data/lists/patabs_standard.html, Access on October 2009.

25

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

A New Method to Extract Dorsal Hand Vein Pattern using Quadratic Inference Function
Maleika Heenaye- Mamode Khan
Department of Computer Science and Engineering, University of Mauritius, Mauritius.

Naushad Ali Mamode Khan
Department of Mathematics, University of Mauritius, Mauritius .

Abstract—Among all biometric, dorsal hand vein pattern is attracting the attention of researchers, of late. Extensive research is being carried out on various techniques in the hope of finding an efficient one which can be applied on dorsal hand vein pattern to improve its accuracy and matching time. One of the crucial step in biometric is the extraction of features. In this paper, we propose a method based on quadratic inference function to the dorsal hand vein features to extract its features. The biometric system developed was tested on a database of 100 images. The false acceptance rate (FAR), false rejection rate (FRR) and the matching time are being computed. Keywords-dorsal hand vein pattern; quadratic inference function;generalised method of moments

Extensive researches are carried out on vein patterns and researchers are striving hard to find methods and techniques to develop dorsal hand vein security system. Any biometric system consists of four main steps namely the preprocessing, feature extraction, processing and matching phase. Feature extraction is a crucial step in biometric system and its capability directly influence the performance of the system. In this work, the method proposed aims at reducing the dimension of the training set by building an adaptive estimating equation or a quadratic inference function [8],[9] that combines the covariance matrix and the vectors in the training set. The organisation of the paper is as follows: in section II we describe the pre-processing phases applied on the hand dorsal vein pattern, feature extraction using quadratic inference function is presented in section III, we explain the vein pattern matching in section IV and finally the experimental results are presented in chapter V. II. PREPROCESSING PHASES APPLIED ON THE DORSAL HAND
VEIN BIOMETRIC

I.

INTRODUCTION

There is an increasing interest for biometric in the research community since traditional verification methods such as passwords, personal identification numbers (PINS), magnetic swipe cards, keys and smart cards offer very limited security and are unreliable[1],[2]. Biometric which involves the analysis of human biological, physical and behavioral characteristics is being developed to ensure more reliable security. The most popular biometric features that are used are fingerprints, hand geometry, iris scans, faces, as well as handwritten signatures. Recently dorsal hand vein pattern biometric is attracting the attention of researchers and is gaining momentum. Anatomically, aside from surgical intervention, the shape of vascular patterns in the back of the hand is distinct from each other [3], [4], [5]. Veins are found below the skin and cannot be seen with naked eyes. Its uniqueness, stability and immunity to forgery are attracting researchers. These feature makes it a more reliable biometric for personal identification [1]. Furthermore, the state of skin, temperature and humidity has little effect on the vein image, unlike fingerprint and facial feature acquirement [6]. The hand vein biometrics principle is non- invasive in nature where dorsal hand vein pattern are used to verify the identity of individuals [7]. Vein pattern is also stable, that is, the shape of the vein remains unchanged even when human being grows.

First of all, it is necessary to obtain the vein pattern in the image captured. The preprocessing phases which consist of image acquisition, hand segmentation, vein pattern segmentation, noise filtering and thinning of the vein pattern are applied on the hand dorsal vein pattern. A. Image Acquisition Vein pattern is found beneath the skin and is invisible to the naked eye. Up to now, there is currently no publicly available hand vein pattern database available to the research community [2]. Each researcher has to capture their own images by devising their own setup. Vein images can only be captured by using either near infrared or far infrared light. However, according to research better quality images can be obtained using near infrared light [10]. A thermal camera or alternative setup like using a charge coupled device (CCD) with alternative devices can be used to capture the dorsal hand vein pattern. In this work a database of 100 hand dorsal vein pattern was obtained by a group of researcher, Prof.Ahmed Badawi from

26

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

University of Tennessee Knoxville [7]. The images were taken with a commercially available conventional charge couple device (CCD) monochrome camera. The hand was presented as a clenched fist with the thumb and all the other fingers hidden. In the setup the intensity of the IR source is attenuated by the use of diffusing paper, which also helps for obtaining an equally distributed illumination on the hand area. A frame grabber is used to capture the image for computer processing. Images were captured using a 320W x 240H pixels video digitizer with a gray- scale resolution of 8- bits per pixel [7], [11], [12]. All the images obtained are of width 320 pixels and of height 240 pixels. It is to be noted that all image templates are of the same size. The figure below shows one of the samples of the dorsal hand vein pattern obtained.

Estimation of Background

Subtraction of Background

Contrast Adjusment

Figure 1. Sample of a hand dorsal vein pattern

B. Hand and Vein Pattern Segmentation There is a need to remove all the unnecessary information obtained during data capture. When a hand image is obtained, the hand background is first segmented from the image. For hand segmentation, morphological operations are applied on the hand image to estimate the background of the hand region. The two morphological operations are dilation and erosion, where dilation is an operation that “grows” objects and erosion “thins” objects in a binary image. Erosion followed by dilation was used and this creates an important morphological transformation called opening. The opening of an image X by structuring element B is denoted by XoB and is defined as follows [13]:

Vein Pattern thresholding

Vein Pattern Thinning

X o B = ( X ⊕ B) ⊕ B

(1)
Figure 2. Biometric Procedure

The background was then subtracted from the original image. This allows us to obtain the region of interest. The contrast that varies all over the vein image has been adjusted. After this operation the hand is being segmented. In order to obtain the vein pattern, the image is then thresholded. Thresholding is the most common segmentation method which is computationally fast and inexpensive. Figure 2 shows the resulting images after applying all the above steps.

C. Enhancement and Thinning of the Vein Pattern The clearness of the vein pattern varies from image to image. Thus, we had to enhance the quality of the image to obtain the vein structures. To achieve this, different filters was applied on these vein patterns segmented. Match filter, Wiener filter and smoothing filter as proposed by S.Zhao et al, [14] was used to suppress noises that exist in the vein pattern. This allowed us to obtain clearer vein pattern for feature extraction.

27

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

As the size of veins grow as human beings grow, only the shape of the vein pattern is used as the sole feature to recognize each individual. A good representation of the pattern’s shape is via extracting its skeleton. A thinning algorithm was devised to obtain a thinned version of the vein pattern. Pruning, which eliminates the shadow in the images was applied on the image. III.FEATURE
EXTRACTION OF THE VEIN PATTERN USING QUADRATIC INFERENCE FUNCTION (QIF)

where Let

ψ jk

1 I ∑ xij I i=I 1 I and ψ k = ∑ yik I i=I = (ψ j ,ψ k ) .

ψj =

(5) (6)

After obtaining the vein pattern, the coordinates were extracted from the pattern. Each coordinate represent the pixel values of the image We assume I images for the training set X, i.e,

φij = xij − ψ j and φik = y ij − ψ k .Thus, we can write φijk = X ijk − ψ jk ,we then construct another training set φ = [φ1 , φ 2 ,K, φi ,K, φ I ] in the same way as the training

X = [ X 1 , X 2 ,K X i ,K X I ]
⎡ X i11 ⎢X X i = ⎢ i 21 ⎢ M ⎢ ⎣ X iM 1
where

(2)

X i12 X i 22 M K

K K X ijk K

X i1N ⎤ X i2N ⎥ ⎥ M ⎥ ⎥ X iMN ⎦

set X using the new co-ordinate systems defined in equations ( 5) and ( 6). To reduce the dimension and avoid the original training matrix X , we construct a new matrix of lower dimension that is based on the following quadratic inference formula [8],[17]:

(3)

Q = g T C −1 g

(7)

where g is a matrix of dimension M × 2 N and its elements are g jk = ψ (4)
jk

and C =

1 I T ∑ φiφi . I i =1

X ijk = ( xij , yik )
where

i is the index for the i th image, j is the th corresponding index for the x -coordinate of the i image, k is the corresponding index for the y − coordinate of the

Note the dimension of the training set has been from M × 2 NI to 2 N × 2 N following equation (7). To generate the eigenvein, we use Qvi = μ i vi (8) For each eigenvector, a family of eigenvein has to be generated. However, many eigenveins are being generated. In order to determine how many eigenveins are required, the following formulae are being used. We have accounted for 90 % and 95% of the variation in the training set.

i = 1,L, I , j = 1,L , M and i th image where k = 1, K, N . Thus, the training matrix X is of dimension M × 2 NI . Note that for each image X i in the training set X , the number of x co-ordinates is chosen using the condition M = min( M 1 , M 2 , K , M i , K , M I ) ,where M i = the number of x co-ordinates in image i and the number of y co-ordinates is chosen using the condition N = min( N 1 , N 2 , K , N i , K , N I ) ,where N i = the number of y co-ordinates in image i . These two
conditions are very important in stabilizing the dimensionality of the matrix X i . However, this way of representing the training set has two major drawbacks. Firstly, there may be variation in the elements of the training set and secondly, the dimensions of the matrix is very large especially when I is large. Thus, it becomes difficult to work with the original training matrix X . To solve these problems, we standardize the coordinates by averaging, i.e,

∑μ ∑μ
j =1 2N' i =1 2N j =1 i =1 2N

2N'

i

> 0.9

(9)

j

∑μ ∑μ

i

> 0.95

(10)

j

We have already obtained 2N ' eigenveins. For each element in the training set, the weight is calculated. This weight will demonstrate the contribution of each eigenvein to respective training element. If the weight is bigger, then the eigenvein has shown the real vein. If the value is less, there is no big contribution with the real vein for that particular

28

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

eigenvalue. The following operation shows how each element in the training set is projected onto the vein space:
ωk =
1 NM

100

0.0400

0.0300

∑∑
i =1

N

M

j =1

(Q υ k )T ( X

T ij

−φT) j

(11)
TABLE II. FAR AND FRR USING QUADRATIC INFERENCE FUNCTION

where,

1 < k < 2 N ' ,1 ≤ i ≤ 2 N , j = 1...M
Each element in the training set has a weight to determine their contribution to the vein space.
IV.VEIN PATTERN MATCHING

To recognize an image means to check whether the image exist in the database. When a person wants to get access to the system, the picture of the vein, known as the test image is captured. The coordinates of the test image are obtained and represented as the training set. The weight of the new image is calculated and projected on the vein space [15],[16]. The vein space contains all the vein images. Thus, we have to check whether the input image exist in that space. The Euclidean distance between the projected image and those stored is being calculated. First of all, our system checks whether the test image is a vein by testing it with an arbitrary value. Then the Euclidean distance is computed to check whether the test image exist in the database. If it is vein image, then it is accepted. The results were recorded and analyzed. V.EXPERIMENTAL RESULTS The hand dorsal vein biometric was tested using pixel by pixel method and the quadratic inference method discussed in this paper. It is to be noted that pixel by pixel method test each individual pixel by counting the number of overlapped pixel in the test image and that of the template found in the database. In order to test the efficiency and accuracy of the method proposed, false acceptance rate (FAR) and false rejection rate(FRR) are computed. False Acceptance Rate refers to the total number of unauthorized persons getting access to the system over the total number of people attempting to use the system. False Rejection Rate refers to the total number of authorized persons not getting access to the system over the total number of people attempting to get access to the system. The table below shows the FAR and FRR for 20,40,60,80 and 100 images tested.
TABLE I. FAR AND FRR USING PIXEL BY PIXEL METHOD

Number of images 20 40 60 80 100

FAR(%) 0.0500 0.0250 0.0340 0.0250 0.0200

FRR(%) 0.0600 0.0500 0.0340 0.0125 0.0300

According to the results obtained, the FAR and FRR is less when using quadratic inference function compared to pixel by pixel method. In order to test the efficiency of our proposed method, we have computed the matching time of the method illustrated in the table below:
TABLE III.
COMPARISON OF MATCHING TIME

Number images

of

Matching time using pixel by pixel(in second)

20 40 60 80 100

275 580 843 1130 1400

Matching time using Quadratic Inference function (in second) 130 300 450 575 700

From the results obtained, it is noticed that the matching time of the proposed method is less compared to the pixel by pixel method. According to the experimental results, quadratic inference function method is on average twice faster compared to pixel by pixel method.
VI.CONCLUSION

Number of images 20 40 60 80

FAR(%) 0.1000 0.0250 0.0340 0.0375

FRR(%) 0.1500 0.0750 0.0670 0.0250

The new method proposed that is the quadratic inference function was successfully applied on hand dorsal vein pattern providing satisfactory results. The FRR and FAR were computed and are found to be less when using our proposed method. It also reduces the dimension of the matrices which consequently has an impact on matching time. The matching time is improved in our proposed method and this is desired in all biometric security system. ACKNOWLEDGMENT We express our deepest thanks to Prof. Ahmed M. Badawi, from University of Tennessee, Knoxville, for providing us with a dataset of 200 images of hand dorsal vein pattern.

29

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

REFERENCES
[1] L.Wang and G.Leedham, “Near- and- Far- Infrared Imaging for Vein Pattern Biometrics” ,Proceedings of the IEEE International Conference on Video and Signal Based Surveillance, 2006 L. Wang, G.Leedham and D.Cho. “Minutiae feature analysis for infrared hand vein pattern biometrics”. The Journal of the pattern recognition society, Volume 41 , Issue 3 (March 2008) Pages 920-929 Year of Publication: 2008 ISSN:0031-3203 C. Lin and K. Fan. “ Biometric Verification Using Thermal Images of Palm- Dorsa Vein Patterns” , IEEE Transactions on Circuits and Systems for Video Technology, VOL.14,NO.2, 2004.

[2]

[3]

[4]

J.Cross and C.Smith, “Thermographic Imaging of Subcutaneous Vascular Network of the Back of the Hand for Biometric Identification”, IEEE 29th Annual 1995 International Carnahan Conference, (1995) 20- 35
C. Deepika and A.Kandaswamy, “ An Algorithm for Improved Accuracy in Unimodal biometric Systems through Fusion of Multiple Feature Sets” , ICGST- GVIP Journal, ISSN 1687-398X, Volume (9), Issue (III), June 2009.

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

L.Chen, H.Zheng, L.Li , P.Xie and S.Lui, “ Near-infrared Dorsal Hand Vein Image Segmentation by Local Thresholding Using Grayscale Morphology”. The 1st International Conference on Bioinformatics and Biomedical Engineering,2007. Page(s):868 – 871 A.Badawi, “Hand Vein Biometric Verification Prototype: A Testing Performance and Patterns Similarity”. In Proceedings of the 2006 International Conference on Image Processing, Computer Vision, and Pattern Recognition (IPCV'06: June 2629, 2006, Las Vegas, USA A. Qu, B.Lindsay,B.Li. “Improving generalized estimating equation using quadratic inference function”, Journal of Biometrica, Vol. 87, Pg. 823-836, 2000. L.Hansen, “Large sample properties of generalized method of moments estimnators”, Journal of Econometrica, Vol.50, Pg. 1029-1054, 1982. S. Zhao, Y.Wang and Y.Wang.“Extracting Hand Vein Patterns from Low-Quality Images: A New Biometric technique Using Low-Cost Devices”, IEEE, 4th International Conference on Image and Graphics, 2007 M.Shahin, A.M. Badawi and M.Kamel, „Biometric Authentication Using Fast Correlation of Near Infrared Hand Vein Patterns”, Volume 2 Number 3 (2007) M.K.Shahin, A.M. Badawi and M.E.Rasmy, “A Multimodal Hand Vein, Hand Geometry, and Fingerprint prototype design for high security biometrics”, In the Proceedings of CIBEC’08 M. Sonka, V.Hlavac, R.Boyle. “Image Processing: Analysis and Machine Vision”, Thomson- Engineering; 2nd Edition (September 30, 1998). S.Zhao, Y.Wang, Y.Wang.E “Extracting Hand Vein Patterns from Low-Quality Images: A New Biometric technique Using Low-Cost Devices”, IEEE, 4th International Conference on Image and Graphics, 2007 M.Heenaye-Mamode Khan, R.K.Subramanian and N. Mamode Khan,2009. Low dimensional representation of dorsal hand vein features using Principle Component Analysis, In the Proceedings of World Academy of Science, Engineering and Technology, Pg 1091- 1097, Dubai, United Arabs Emirates, Volume 37, January 28-30, 2009, ISSN 2070- 3740 M. Heenaye- Mamode Khan, R.K.Subramanian R.K., and N. Mamode Khan,2009. Representation of Hand Dorsal Vein Features Using a low Dimensional Representation Integrating Cholesky Decomposition, In the IEEE 2nd International Congress on Image and Signal Processing, Vol.3, 17-19 October 2009, Tianjin, China. A. Qu and B.Lindsay,2003, Building adaptive estimating equations when inverse of covariance estimation is difficult, Journal of Royal Statistical Society, Vol.65, pg. 127-142, 2003. .

30

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Architecture of Network Management Tools for Heterogeneous System
Rosilah Hassan, Rozilawati Razali, Shima Mohseni, Ola Mohamad and Zahian Ismail Department of Computer Science, Faculty of Information Science and Technology Universiti Kebangsaan Malaysia, Bangi, Selangor, Malaysia
. Human: where human manager defines the policy and organization approaches. Methodology: defines the architectural framework and the functions to be performed. Instrumentation: the actual operational aspects that establish the procedures, methods and algorithms for data collection, processing and reporting, and analysis of problems, their repair, prediction or forecasting of service levels and probable improvements to enhance performance. S&NM aims to provide network managers a complete view of the whole network through a visualized Network Management Tool (NMT). The International Organization for Standards (ISO) [1] has categorized five main management functions that can be managed by such tools: Fault, Configuration, Accounting, Performance, and Security (FCAPS), as simplified in Table 1. To illustrate how these functions are interrelated, an example of simple S&NM applications is shown in Fig. 1. It can be seen that a user interface is used to manage the functions, which are originated from various software, hardware, firmware, and end-users. Most of the existing S&NM systems are developed in an individual fashion, where each system is designed to operate within its own defined area or scope.s This creates a number of incompatibilities and lack of integration does not allow a common view of the system and networks to be managed. Also, this causes lack of data flows between these incompatible systems, resulting in inconsistencies of data, event correlation and maintenance of the different data bases. It may also cause many systems with low level of inter and intra communications among them.

Abstract— Managing heterogeneous network systems is a difficult task because each of these networks has its own curious management system. These networks usually are constructed on independent management protocols which are not compatible with each other. This results in the coexistence of many management systems with different managing functions and services across enterprises. Incompatibility of different management systems makes management of whole system a very complex and often complicated job. Ideally, it is necessary to implement centralized metalevel management across distributed heterogeneous systems and their underlying supporting network systems where the information flow and guidance is provided via a single console or single operating panels which integrates all the management functions in spite of their individual protocols and structures. This paper attempts to provide a novel network management tool architecture which supports heterogeneous managements across many different architectural platforms. Furthermore, an architectural approach to integrate heterogeneous network is proposed. This architecture takes into account both wireless fixed and mobile nodes. Keywords-component; Network Tools Architecture; Services Management; Heterogeneous System;

I. INTRODUCTION System and Network Management (S&NM) is concerned with observing, monitoring, testing, configuring, and troubleshooting the different network components, services, systems and users. The management process wraps all the network system elements starting from the end-users, through the applications and supporting data, the end system’s network connectivity edge and deep into the network infrastructures themselves. Typically, S&NM comprises the following aspects:

31

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
TABLE I. Category of Management Fault Management (FM) ISO SPECIFIED FCAPS FUNCTIONS Task Recognizing problem situation, isolating the problem to the source, Providing notification to the appropriate person, and tracking problems through resolution Controlling the health of network by keeping regular scheduled configuration back ups and having careful controlled implementation and also changing the procedures Measuring the usage statistics and allocation of costs associated with billing for time and services Concerned with gathering network and system statistics including utilization, errors, response time are valuable tools in network trends Controlling access rights, usually network infrastructure and access to network hardware components

Following are the most used protocols in network management systems: Simple Network Management Protocol. Telecommunication Management Network. Fault, Configuration, Accounting, Performance, and Security (FCAPS) A. Simple Network Management Protocol Simple Network Management Protocol (SNMP) is an Internet Engineering Task Force (IETF) de facto standard known as Requests For Commands (RFC) [2]. SNMP is a framework for managing Internet devices or network elements using the TCP/IP protocol suite [3]. The SNMP management model contains a management station. The actual management process takes place in this management station. Other managed devices and available network peripherals communicate via a network management agent embedded inside the device with management station. The SNMP model of a managed network consists of four components: Managed nodes (Agent), Network Management stations (NMS), Managed information (MIB), and a management protocol (SNMP), as shown in Fig. 2. The manager and managed agent style is used in SNMP where the manager monitors a group of agents. The dialogue relationship between manager and agent is shown in Fig. 3. A manager can check the agent behavior through a set of parameters; also it can force agent to behave in a certain way by resetting value of those parameters. Agents can send all or correlated alarms to the manager of any faulty situation. Different network managers can exchange information about each other’s networks using SNMP. SNMP works in cooperation with a Structure of Management Information (SMI) and Management Information Base (MIB), as well as the underlying supporting S&NM protocol suite. SMI is responsible for defining the general rules for naming objects (hardware and system, non-physical such as programs, and administrative information), defining objects types, and show how to encode objects and values [4]. MIB is a conceptual database that can be distributed on different sites or assembled in a single location. It creates the objects needed and names them according to SMI rules. Then, it assigns them to their proper object types. For each entity such as a device, a set of defined objects will be created and kept in the database. Fig. 4 shows the ISO MIB that consists of the data which reflect the FCAPS that are requested by different network management architectures.

Configuration Management (CM)

Accounting Management (AM)

Performance Management (PM)

Security management (SM)

Network management Application User Interface

Fault management

Configuration management

Accounting management

performance management

Security management

Hardware

Software

Firmware

Users

Network componants

Figure 1. Network management application components

32

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
NMS
Managing entity Agent data

B. Telecommunication Management Network The Telecommunication Management Network (TMN) was defined by the International Telecommunication Union – Telecommunication Standardization Sector (ITU-T) [5]. It is a framework to achieve communication between heterogeneous telecommunication networks. TMN defines a standard interface for network elements that handle the communication process. In this way, network elements can be managed by a single network management system regardless of their different manufacturers. The framework identifies four logical layers of network management [5]: Network Management Layer (NML): Enables telecom operators to perform integrated fault management and service provisioning in multivendor multi-platform environments. Service Management Layer (SML): is where Telecom operators have sought to differentiate themselves by purchasing numerous applications for managing service usage, activation and assurance. Element Management Layer (EML): Contains individual network elements handling functions. Enables capabilities related to network monitoring, inventory management, service assurance, and network provisioning. Business Management Layer (BML): Performs functions related to business aspects, analyzes trends and quality issues provide a basis for billing and other financial reports. Telecommunications technology architectures are typically expressed in a more simplified palette using the TMN model in Fig. 5.
BML

Managed node
Agent data

SNMP protocol

Managed node
Agent data

Agent data

Managed node

Managed node

Figure 2. SNMP management model components

Request for information

Reply to manager request

Reseat parameters values

Unusual situation alarm

NM information exchange

Figure 3. Manager Agent Request Response Exchange Protocol

Operations Manager

System Planer

Service Engineer

Helpdesk Manager

MIB

Other Networks managers

Manager

Agent

SML NML EML

Configuration subsystem management

Fault subsystem management

Security subsystem management

Accounting subsystem management

Perfomance subsystem management

NEL

Figure 5. TMN Network Management Architecture Figure 4. ISO Management Information Base (MIB)

33

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

C. Fault, Configuration, Accounting, Performance, and Security (FCAPS) FCAPS is a framework defined by the ISO and the name is a contraction of the five management categories that were mentioned earlier in Table 1: Fault, Configuration, Accounting, Performance, and Security. This categorization is the functional one and does not consider the business-related role of management systems within the telecommunication network. The concept of FCAPS was initially developed by ITU-T to assist in managing the telecommunication networks. But, it was really the ISO who applied the concept to data networks. However, it turned out that those five protocols would be very similar, thus ISO decided to make them under one protocol, namely Common Management Information Protocol (CMIP) [6]. CMIP allows the communication between the manager and managed entities through a typical communication protocol. It can request the setting of parameter values, events and activities. II. NETWORK HETROGENITY A heterogeneous network systems environment can be viewed as connecting computers and network devices such as switches, repeaters and routers with different protocols and different operating systems which varies in type, size, and topology. Fig. 6 illustrates three networks belonging to different autonomous organizations but providing a composite service to end-users. The illustration in this figure can be described as being immature. Each network has a different environment and those environments can communicate with each other by a gateway. But in a real communication environment, the same device is surrounded by different technologies overlapping on the air such as WiFi technology cellular radio, Bluetooth, Zigbeee, and other various technologies with different capabilities and different coverage. The 3G devices, for example, support this variety and overlapping where user can have many types of connectivity using the same device in the same place at the same time. On the other hand, this is a mature heterogeneity. The term heterogeneous in the near past means having different devices run by different operating systems to work under different technologies. This heterogeneity term is becoming less and less by embedding the variant capabilities in one device under one operating system, leading to more homogeneity.
Internet

LAN Cellular

WLAN

Mobile Ad Hoc Network

Figure 6. Example of Heterogeneous Network

It is the feature of modern communication world, where different service providers offering wide variety of services using different technologies, and targeting different user’s interests. Forth generation (4G) technologies which bring high-speed broadband devices and services to the wireless world are good instances of heterogeneous networks where the user environment will be an integrated environment based on terminal heterogeneity and network heterogeneity. Heterogeneity could be derived from different issues as following: Terminal Heterogeneity: refers to the variety of user’s terminals with different characteristics such as operating system, processing power, memory, storage size, battery life, technology supported. Network Heterogeneity: refers to the variety of network types and technologies used for example WiFi, Ethernet, Zigbee, Bluetooth, Cellular networks, GPS, etc. Protocol Heterogeneity: refers the variety of technologies and network, where each type of network uses different communication protocols. Data type Heterogeneity: different networks, protocols, terminals can use different types of data. III. INTEGRATED NETWORK MANAGEMENT TOOL In a heterogeneous environment, it becomes difficult for network managers to control all these networks or subnetworks for different reasons, such as incompatibilities or vendor specific S&NM architectures. There are also other reasons such as: Mastering problem: many management tools need to be mastered at least one for each type of device or network.

34

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Needs of expertise: an expert manager is needed to handle the heterogeneous system and network. Expert network managers are hard to find. Also, They are in short supply. Different protocols: subnetworks within the same network can use different suites of protocols. Data inconsitancy: different subnetwoks have different representation of data types and data flows. Integrated S&NM needs to ensure that it creates a meta-level view and control of these different S&NMs. This need for a single centralized global view of a set of distributed components become esential in order to make managesment effective and efficient. Integrated System and Network management(IS&NM) was proposed to meet this need. The purpose IS&NM is to provide a single set of tools for managing network resources within a “networked set of components” regardless of the type of the subnets. IS&NM reduces errors in mastering multiple different subnetworks; minimizes the human skill levels needed to manage the network; and enhances network management flexibility simulatenously. IV. RELATED WORK There are a number of network management tools that claimed to integrate network management functions. Those tools aimed to manage network components from the same vendor or the same service provider. Therefore, we are mainly interested in bulding a comprehensive network management tool through a single interface. We aim to provide a central control in a distributed management fashion. In this section, we will highlight several integrated network management architectures. Aslo, Table 2 summarizes three network management architectures and tools described is this section. Alcatel [7] offers a complete suite of network management systems for
TABLE II. Integrated Network Management & Tools Alcatel-Lucent Network

sub networks in Mobile Service Providers (MSP) environment. They can be integrated with Alcatel's mobile Operating Support System (OSS) solutions to provide full end-to-end service management. To help MSPs to meet these new challenges, Alcatel NMS and OSS have been designed as open solutions that facilitate the smooth integration of multi-technology and multi-vendor elements. Alcatel integrated fault management based on MicroMuse Netcool® consolidates real-time alarm information across network elements and application servers, providing real-time alerts relating to problems that affect service, thereby reducing the time to repair. Alcatel performance management based on Metrica™/NPR includes the collection of performance data and its consolidation across the entire network and application servers to provide end-to-end Quality of Service (QoS) indicators. Alcatel problem management based on Action Request System® from Remedy® integrates with previous applications, supporting network and service problem resolution by handling the workflow of problem resolution. It reduces network and service downtime and thus improves QoS and customer satisfaction. It can provide repair time and reaction time metrics as a basis for evaluating Service Level Agreements (SLA). The Evolium™ OMC-CN provides efficient cross-network element configuration capabilities and extensive performance monitoring functions tuned to the core network technologies. The Evolium™ OMC-R combines the integration of Radio Network Optimizer (RNO) with Alcatel's worldwide expertise in 2G mobile networks and its UMTS experience in Japan, to realize the best radio network management system on the market. In the CORDS project [8], several management tools have been developed independently. The tools are Network Modeling Tool (NetMod), Network Simulation Testbed (NEST), Network Management Analysis and Testing Environment (NETMATE), Hy+ and Shoshin Distributed Debugger.

NETWORK MANAGEMENT ARCHITECTURE FOR ALCATEL, CORDS, AND NETDISCO In house Service Provider Integrated Service Provider

Mobile network

Alcatel The Evolium™ OMC-CN Alcatel Evolium™ OMC-R Columbia University: NEST and NETMATE -

MicroMuse Netcool® Metrica™/NPR Action Request System® from Remedy® University of Toronto: Hy+, University of Waterloo: Shoshin Distributed Debugger Devices : Airespace, Allied Telesyn, Aruba, Asante, Bay, Cisco, Dell, Enterasys Networks, Extreme Networks, Foundry Networks, HP, Juniper, Net-SNMP, NetScreen, Nortel, Proxim, Sun, Synoptics and Zyxel.Software: Perl, Mason, Net-SNMP, PostgreSQL, Apache 1, GraphViz and MIBs

CORDS Netdisco

Large-scale local network Heterogeneous

35

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

NetMod is a software package that predicts the performance of new network technologies in a largescale local network environment. NEST is a graphical environment for distributed network systems rapid prototyping and simulation developed at Columbia University. The NETMATE was developed at Columbia University to provide a unified and comprehensive software environment for network management to oversee and orchestrate the operations of diverse devices and protocol. Hy+ is a visual database system being developed at the University of Toronto. The system is capable of manipulating data by visually expressed queries on large complex systems such as a computer network. Shoshin Distributed Debugger is being developed at the University of Waterloo to support debugging of distributed and parallel applications Network Discovery and Management (Netdisco) [9] is an Open Source web-based network management tool hosted by Sourceforge. Netdisco is a network management application targeted at large corporate and university networks. It integrated devices and software into one integrated system. The integrated devices are Airespace, Allied Telesyn, Aruba, Asante, Bay, Cisco, Dell, Enterasys Networks, Extreme Networks, Foundry Networks, HP, Juniper, Net-SNMP, NetScreen, Nortel, Proxim, Sun, Synoptics , and Zyxel [9]. These devices provide hardware such as switch, router and hub. The integrated softwares are Perl, Mason, Net-SNMP, PostgreSQL, Apache 1, GraphViz and MIBs [9]. V. PROPOSED WORK In this paper, we propose a network management tool architecture that supports heterogeneous network management system in many different architectural platforms. The proposed system should meet the following requirements. Easy to use: to minimize the level of expert needed to manage the network. Easy to access: web-based architecture is the best way to achieve this issue. The network management tool should have a standardized set of FCAPS function objects. This is needed to eliminate the data inconsistency that can be resulted from different standards in different networks. Able to interpret between definitions in different managed devices through a common nomenclature. A. Architectural Components The proposed architecture is a collection of network management tools. The main components of this

environment are the console, user interface, Agents Interface, and a database facility called MIB. User Interface (UI): This provides the ability to visually navigate and control complex network scenarios. The UI presents numerous objects and relationships into a visual representation for a human network manager. An important feature of user interface is that a user is able to see multiple views of network simultaneously. This feature of UI is shown in Fig. 7. Management Information Base (MIB): This part of architecture is containing information about objects (hardware and system, non-physical such as programs, and administrative information).MIB is organized by grouping of related objects and defines relationship between objects. Also, it is not a physical database; it is a virtual database that is complied into management model Agents Interface: the agents Interface of this architecture are entities providing information with a standardized interface for MIB. The gateway interacts with these agents interface asks the agents for information about the different network devices. These agents have a very simple structure and usually only communicate in response to the requests of variables contains in the MIB. Gateway: A gateway stands for communicating between networks that use different protocols, or which have the same protocols but do not otherwise communicate. Gateways integrate various data sources and create the appearance that all data resides in a single, logical relational database.

Figure 7. Example of User Interface

B. Process of proposed Integrated Network Management Tool The process of proposed network management tool in this paper is as following. The local network

36

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

management system in each network gathers the information needed about the devices, terminals, software and users in that network using its own protocols. The gathered data has to be sent to the Agents Interface in the Console part. The Agents Interface communicates with the agents and then stores those data in the MIB database. The Management Functions combines the FCAPS functions (Fault, Configuration, Accounting, Performance, and Security), which can be accessed by users through the User Interface. The user interface (UI) is a web-based interface that visualizes and shows the attached networks management information. In this architecture, agents on different networks interact with a gateway to communicate with any other agents. Networks such as LAN, GSM, ATM, and ISDN are local networks and each one has its own network management systems. Through gateway agent, each local network has an agent and they speak to gateway which leads to centralized control with distributed management. A conceptual view of the proposed architecture is shown in Fig. 8.

Message Translator, Message Rebuilder and Message Dispatching Relay Agents. The information received by the gateway from different networks would be processed as in Fig. 9. In proposed architecture, the in-coming data from agents of different types of networks such as GSM, ATM, LAN, and ISDN are translated in the gateway agent. The information in the gateway agent is translated from actual to generic or vice versa. The functions of those elements are as follows: Common Information Model (CIM): This is an object oriented information model for specifying management information in a way that is independent of applications, platforms, protocols and implementations. CIM defines a way to exchange the data from any source and network. The data represented using CIM can be understood and analyzed by any network management tools or applications that understand CIM. The heterogeneous networks can make correlations between information coming from different locations in the network Extensible Markup Language: The Desktop Management Task Force (DMTF) recently advanced the CIM encoding to standard Extensible Markup Language (XML). In order to interoperate with each other for applications, it is necessary to represent actual management data in a standard way. Extensible Markup Language (XML) is a markup language for representing information in a standard format in order to make heterogeneous platforms and applications interoperate with each other. CIM operation over HTTP: This concept stands for mapping of CIM operations onto the Hyper Text Transfer Protocol (HTTP). All management functions can be accessed by users through the User Interface, which is a web based interface that visualize and show the attached networks management information. Gateway: A gateway which also described in the last section stands for communicating between two networks that use different protocols, or which have the same protocols but do not otherwise communicate. Gateways integrate various data sources and create the appearance that all data resides in a single, logical relational database. It is a combination of hardware and software that translates between two different protocols and acts as the connection point to the Internet.

Console User Interface Management Functions Agents Interface MIB Local NMS ISDN Gateway
LAN

Local NMS

Local NMS ATM Local NMS GSM

Figure 8. The Conceptual View of the Proposed Network Management Tool Architecture

C. Proposed Network Management Tool Requirements There are four elements involved in the proposed management environment, as shown in Table 3. The block diagram of a gateway application is shown in Fig. 9. It consists of five components: Message Accepting Rely Agents, Message Extractor,

37

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
Message Extractor Heterogeneous Same Message Extractor Message Translator Message Rebuilder Message Dispatching Relay Agent

Message Accepting Relay Agent

Homogeneous

Figure 9. Block Diagram of the Gateway Application

TABLE III. Requirements

REQUIREMENTS OF MANAGEMENT ENVIRONMENT Application Common Information Model (CIM) Gateway Extensible Markup Language Hyper Text Transfer Protocol (HTTP)

Data description lTransport Encoding Representing information Operations to manipulate the data

zigbee, and ethernet networks like local area networks LAN are integerated in this experimental work. The widespread of the multi radio transmission mode wireless devices which are able to connect to a cellular base station CBS, and to a wireless access point WAP makes the daily used network an integration of different networks using different technologies. Much architecture has been proposed to integrate the wireless triangle [12]: Cellular, WLAN, and Mobile Ad Hoc networks. The proposed architecture consists of the following basic components (which are also summarized in Table 4): -Mobile nodes: contains all the nodes that are free to move, which could be consist of the following: 3G Cellular phones that support WiFi, and GPS. The cellular phones will represent the dual mode nodes; Dual mode nodes have two applicability, they are MANETs mobile nodes and mobile gateway at the same time. Internet access to MANET nodes is provided through mobile gateways. The concept of mobile gateways is presented in [13]. WLAN and Cellular network. Different battery life time is placed in each node. Laptops with WiFi and various battery life times. -Fixed nodes: nodes are connected to fixed infrastructure, which is LAN (IP network) or to a cellular fixed infrastructure in case of CBS: like Pc connected to LAN Fixed Internet Gateway (IG): IG works as an interface between two or more networks of different types. In this architecture, there will be one fixed gateway, and one of the cellular phones will act as the mobile gateway. Both gateways have dual interfaces. Cellular Base Station (CBS): in this architecture assumption is that the coverage of the cellular base station is larger than the wireless access points used in WLANs at all times. This scenario uses only one CBS that covers area with 10km radios. This CBS is connected to a central CBS; this connection does not exist in Fig. 10. Three Wireless Access Point (WAP): IEEE 802.11 access points with 100m radios coverage.

VI. IMPLEMENTATION We have developed an initial system prototype to proof of the proposed architecture and experimental development. The prototype gives us a practical environment to evaluate and demonstrate our proposed framework in actual wireless IP networks. This prototype employs off-the-shelf fixed/mobile equipments and also, standardized interfaces to ensure interoperability with existing network and systems for pervasive applications. The ad hoc network in this heterogeneous architecture incorporates on of the well-known MANET Reactive Protocols called AODV which is running on top of the IEEE 802.11 WLAN in ad hoc made to support multi-hop wireless communications. AODV-UU from Uppsala University is deployed as a user-space daemon that alters the kernel routing table dynamically as per the ever changing topology, and provides Internet connectivity through gateway support with tunnels [10]. A. EXPERIMENTAL TESTBED AND EXPERIMENTS We have developed an experimental work (Fig. 10) in our laboratory for performance evaluation of proposed architecture. Two main focuses in this experimental work are feasibility and usability. Without any infrastructure, using diverse terminal types (laptops, PDAs, and mobile phones) we were able to self-organize and form a community of group and/ private communications. Various wireless platforms, namely Mobile Ad Hoc networks (MANET), Cellular networks (2G, 2.5G, 3G)[11], Wireless Local Area Networks WLANs (WiFi), Wireless Personal Area Networks (WPANs) like Bluetooth technology,

38

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
TABLE IV. HETEREGENOUS NETWORK ARCHITECTURAL COMPONENTS

VII. CONCLUSION AND FUTURE WORK Diversity and complexity are the titles of the coming communication technologies. This situation caused by the increased production of the communication devices and systems without laying on one standardized concepts or common language. Most of the systems and devices nowadays concern on heterogeneous networks. That raised the intensive need to find one station to control and manage those networks, since controlling them separately brings a lot of difficulties and inconsistency. To conclude, we proposed network management tool architecture for managing heterogeneous networks with a web-based interface. The architecture is being currently designed for implementation to meet the criteria as for IS&NM, which are easy to use, easy to access, and capable of interpreting different devices. To proof the proposed NM Tool we have developed a Testbed in our laboratory. The result is an “Integrated Network Management Tool” targeting to enhance our future living environments. The initial implementations demonstrate that proposed network management architecture provides a solid foundation for developing future network management tools. Our future work will be providing QoS to this architecture. We anticipate that this will enable supporting many services and applications in heterogeneous networks such as multimedia applications. ACKNOWLEDGMENT This research is funded by Universiti Kebangsaan Malaysia (UKM) research projects UKM-GUP-NBT-08-29-116. The research group is known as Network Management Group. Please visit our website at http://www.ftsm.ukm.my/network for detail. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors. REFERENCES

Number of nodes 6

Device

Single mode/ dual mode Dual mode cellular and WLAN Single mode WLAN LAN Dual mode Single mode : cellular Single mode : WLAN

Specifications

Mobile Nodes

cellular

3G, movement speed 0m/S20m/s , random directions movement speed 0m/S20m/s , random directions Pentium III

5

laptop

1 1

PC Internet Gatewa y IG Cellular Base Station (CBS) Wireles s Access Point (WAP)

Fixed Nodes

1

10km radios coverage

3

100m radios coverage, IEEE 802.11

Cellular coverage area LAN

[1]

DM

Pc

IG

A. Gorod, R. Gove, B. Sauser, and J. Boardman, “System of Systems Management: A Network Management Approach”, Proceedings of the IEEE System of Systems Engineering, SoSE '07. IEEE International Conference on Year of Publication: 2007 ISBN: 1-4244-1160-2. IETF SNMP Extensions working group, Karl Auerbach, Epilogue Technology K. Ramesh Babu etc., www.ietf.org/rfc/rfc1157.txt ,last Data of Access Nov,2009s Information about the current status of Simple Network Management Protocol (SNMP) Version 3. www.ibr.cs.tu-bs.de/projects/snmpv3/ , last Data of Access Nov,2009 A. Tanenbaum, Computer Networks, 4th edition, New Jersey, USA, Prentice Hall, ISBN 0-13-038488-7, 2003. D. P. Griffin, A TMN system for VPC and routing management in ATM networks, Proceedings of the fourth international symposium on Integrated network management IV Elsevier Science, (Pages: 356 - 369 Year of Publication: 1995 ,ISBN:0-412-71570-8)

SM

WAP

CBS
WAP

[2]
DM

DM DM SM WAP SM DM

DM SM

SM

[3]
Wireless Hot spots

CMS: cellular Mobile Station LAN: Local Area Network infrastructure WAP: Wireless Access Point DM: Dual Mode Node SM: Single Mode Node IG: internet Gateway

[4] [5]

Figure 10. Heterogeneous Network Testbed

39

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

[6]

Klerer, S.M., The OSI management architecture: an overview, IEEE, Volume: 2, Issue: 2 , page(s): 20-29, Mar 1988, ISSN: 0890-8044

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

[7]

P. H Gross and N. Mercouroff. Integrated Network and Service Management for Mobile Networks, http://www1.alcatellucent.com/, last Data of Access Nov,2009 M. A. Baouer, P.J. Finnigan, J. W. Hong, J. Pachl and T. J. Teorey, An Integrated Distributed Systems Management Architecture, Proceedings of the Conference of the Centre for Advanced Studies on Collaborative Research, Software Engineering, (Volume 1, Toronto, Ontario, Canada, Pages 27-40, 1993). Netdisco (Network Discovery and Management) http://www.netdisco.org/, last Data of Access Nov,2009

[8]

Shima Mohseni was born 1984 in Iran. She received her B.Sc. in computer science and Engineering from the Islamic Azad University of Najafabad, Iran in 2005. She is presently a M.Sc. student in the Information Technology department from faculty of Computer Science at Universiti Kebangsaan Malaysia. Also, she was working as research assistant with “network management tool group” in 2009. Her fields of interest include computer networking, mobile computing and QoS routing. She can be connected at shima.mohsenids@gmail.com Zahian Ismail is a postgraduate student born in Malaysia. Her bachelor degree was in Computer Science and recently furthering her study in the same area at Faculty of Information Science and Technology in Universiti Kebangsaan Malaysia. She is a member of Network Management research group at her faculty. Her field of interest includes mobile ad hoc network, network communication, and QoS routing.

[9]

[10] E. Wedlund and C. Tschudin , “MANET Internet Connectivity with Half Tunnels”, presented at SNCNW, Arlandastad, 2003. [11] Mohammed F. Al-Hunaity, salam A. Najim, Ibrahiem M. ElEmary, Acomparative Study Between Various Protocols Of Manet Networks, American Journal of applied sciences, Volume 4 (9), pages 663-665, 2007,ISSN1546-9239. [12] Aymans Mansour Murad, Bassam Al-Mahadeen, Adding Quality Of Service Extensiois To The Enhanced Associativity Based Routing Protocol For Mobile Ad Hoc Networks(MANET), American Journal of applied sciences, volume 4 (11): pages 876881, 2007, ISSN 1546-9239. [13] H. Ammari and H. El-Rewini, “Integration of Mobile Ad Hoc Networks and the Internet Using Mobile Gateways,” Proceedings of the 4th International Workshop on Algorithms for Wireless, Mobile, Ad Hoc and Sensor Networks (WMAN04), Santa Fee, New Mexico, USA, April 26-30, 2004.

AUTHORS PROFILE Dr Rosilah Hassan was born in Malaysia. She received her first degree from Hanyang University, Seoul, Republic of Korea in Electronic Engineering (1996). She work as an Engineer with Samsung Electronics Malaysia, Seremban before joining Universiti Kebangsaan Malaysia (UKM) in 1997. She obtains her Master of Electrical Engineering (M.E.E) in Computer and Communication from UKM in 1999. In late 2003 she went to Glasgow, Scotland for her PhD. She received her PhD in Mobile Communication from University of Strathclyde in 2008. Her research interest is in mobile communication, networking, 3G, and QoS. She is a senior lecturer at UKM for more than 10 years

Dr Rozilawati Razali received her BSc. in Software Engineering from Sheffield Hallam University, United Kingdom in 1997. Prior to joining Universiti Kebangsaan Malaysia in 2003, she used to work in industry for about 6 years as a Software Engineer. She obtained her MSc. From Universiti Teknologi Mara in 2002 and PhD in Computer Science from University of Southampton, United Kingdom in 2008. Her research interests include software metrics, quality management and empirical software engineering.

40

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

A Topological derivative based image segmentation for sign language recognition system using isotropic filter
M.Krishnaveni,
Research Assistant, Department of Computer Science, Avinashilingam University for Women Coimbatore,India . Abstract-The need of sign language is increasing radically especially to hearing impaired community. Only few research groups try to automatically recognize sign language from video, colored gloves and etc. Their approach requires a valid segmentation of the data that is used for training and of the data that is used to be recognized. Recognition of a sign language image sequence is challenging because of the variety of hand shapes and hand motions. Here, this paper proposes to apply a combination of image segmentation with restoration using topological derivatives for achieving high recognition accuracy. Image quality measures are conceded here to differentiate the methods both subjectively as well as objectively. Experiments show that the additional use of the restoration before segmenting the postures significantly improves the correct rate of hand detection, and that the discrete derivatives yields a high rate of discrimination between different static hand postures as well as between hand postures and the scene background. Eventually, the research is to contribute to the implementation of automated sign language recognition system mainly established for the welfare purpose.
Key words: Sign Language, segmentation, restoration, .topological Derivates, Quality measures.

Dr.V.Radha,
Reader, Department of Computer Science, Avinashilingam University for Women Coimbatore,India . presents a segmentation approach to the automatic training and recognition of sign language. This work employs a enhancement method which uses filters with segmentation to locate the dominant hand. Both objective and subjective evaluation is measured by common parameters. The paper is organized as follows. In Section 2 we introduce the framework underlying the presented approach; Section 3 shortly introduces the restoration filters and Section 4 presents the segmentation method that is used in our approach. Section 5 presents the experimental results. Finally, the paper is summarized and concluded in section 7 with future work. II SYSTEM OVERVIEW An overview of the automatic sign language recognition system is given here. This allows the research work to adopt the image processing techniques according to the need of the system.
Scaling Image processing LDA Feature Combination

I. INTRODUCTION Currently, major research groups focus on the sign recognition problem. The recognition process includes segmentation as the chief step, but the segmentation results are not evaluated that good [18]. However, those recognition methods are based on several approaches that could also be used for sign segmentation [13]. The sign recognition methods can be classified into several categories according to the model of sign they refer to. The approaches are distinguished as one segment, multi segment and hidden model segment based. One-segment approach is used over this research work. In one-segment approach, each gesture is modeled as one single segment. This method has only been applied to gesture classification and was not employed to process real signs [2]. This kind of approach is mainly useful in sign language processing [14]. Here the research
Hand Tracking Geometric Features LDA

Recognition

Figure 1: System Overview

There are many segmentation methods approved over sign language system. However these methods lacked precision due to the limited degree of freedom between the different hand signs [7]. This drawback resulted in common incorrect decisions since most hand signs are very similar thus not

41

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

allowing the system to differentiate between them [16]. This research work over here is towards the enhancement strength before segmenting the hand sign which henceforth gives better performance in classifiers [8]. When the system runs the language model, transition and emission probabilities can be weighted by exponentiation with exponentsα, β andγ, respectively, the probability of the knowledge sources are estimated as in eqn (1):

A vector filter is isotropic eqn (3), if its kernel F(x) = F (x1,…..xd) : IRd →Md invariant by rotation and symmetries:

R.F ( RT x).R t = F ( x)∀R ∈ O(d )

---------------(3)

An Isotropic scalar filter f depends only on radial distance : f(x) = g (xTx).This is not true anymore for isotropic vector filters. Though the standard research proves that anisotropic filter is superior always, the scenario here states that the evaluation speaks positive only to isotropic filter, because i) ii) iii) It works better for Gaussian noise. The image taken is two dimensional Defined only on unbounded domain.

Pr( w 1 ) → p α ( w 1 ),
N N

Pr( s t | S t −1 , w 1 ) → p β ( S t | S t −1 , w 1 ), ----(1)
N N

Pr( x t | S t , w 1 ) → p γ ( x t | S t , W 1 ).
N N

The exponents used for scalingα, β and γ are named language model scale, time distortion penalty, and word penalty, respectively. The system overview is shown in Figure. 1. The system overview is explained in such a way that the image is prescreened for hand tracking and from that the geometric features is extracted [9]. Linear discriminant analysis will be used for selecting combination of features that are to be trained through classifiers[10]. This paper will describe the intermediate part of how to segment the hand postures through topological derivatives combined with image restoration (filters). III. IMAGE RESTORATION TECHNIQUES Noise occurs in all coherent imaging systems. To reduce the noise over the images two filters are used under continuum and discrete derivatives: isotropic and anisotropic filters. Here in this paper image restoration is handled with two filters which are shown in figure2. This approach before segmentation achieves the objective of detecting the number of open fingers using the concept of boundary tracing combined with finger tip detection [11]. It handles breaks, if any, during boundary tracing by rejoining the trace at an appropriate position. Here the restoration of the sign image is done using continuum and discrete topological derivative algorithm. The main idea behind this algorithm is to compute the topological derivative for an appropriate functional and a perturbation given by the introduction of cracks between pixels. This derivative is used as an indicator function to find the best pixels to introduce the cracks that, in the presence of diffusion, will most remove noise preserving relevant image characteristics[15]. Here, this paper shows the possibility to solve the image restoration problem using topological optimization. The basic idea is to adapt the topological gradient approach. A. Isotropic filtering is a filter that enhance the given noisy image in particular point if it looks the same in all directions [4]. It reduces the noise cause as the non filters do. Isotropic filter is scalar filter F(x) : IRd →IR is isotropic if it is invariant by rotations and symmetries ,i.e if eqn (2)

B. Anisotropic filtering is a method of enhancing the image quality of textures[6] on surfaces that are at oblique viewing angles with respect to the camera where the projection of the texture (not the polygon or other primitive on which it is rendered) appears to be non-orthogonal[3]. Like bilinear and trilinear filtering it eliminates aliasing effects, but improves on these other techniques by reducing blur and preserving detail at extreme viewing angles.

Figure 2 Image restoration (a) Original Image (b) Noise (Gaussian) (c) isotropic filter (d) anisotropic filter

f ( RT x) = f ( x)∀RεO(d ) ----------------(2)

42

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

Topological Derivative quantifies the sensitivity of a problem when the domain is perturbed by the introduction of heterogeneity (hole, inclusion, source term, etc.)[17][18]. Let the domain Ω under consideration is perturbed by the introduction of small holes (topology changes) in Ω as shown in fig4. Let us consider Ω be a bounded open set in RN (N=2,3) and γε be a crack of the ε centered at point xεΩ

ψ ε (Ωε ) = ψ (Ω) + f (ε ) DT ( x) + 0 f ((ε ))..................(4)
Where f(ε) is known positive function going to zero with ε,and Dτ (x) is the topological derivative at point x given in eqn (4).

^

Figure 3: Performance evaluation of the images objectively

The isotropic filter performs better than the anisotropic filter in sign images. The reason for the anisotropic filter performing less optimally at low SNR is because the filter kernel is derived from the data, which itself is very noisy. As the SNR was increased, the isotropic filters were more optimal and were less prone to blurring and other over filtering effects. The study also demonstrates that with data at moderately high SNR, filtering can easily introduce more errors than are removed. In conclusion, the results from this study demonstrate that isotropic filtering can effectively reduce errors in sign images as long as they are applied properly. IV. IMAGE SEGMENTATION TECHNIQUES There is no common solution to the segmentation problem in image processing domain. A priori knowledge about the objects present in the image, e.g., target, shadow, and background terrain, should be segmented in better way[14]. The visual relevance of the segmentation problems should be considered rather than simply their plurality; e.g. oversegmentation [5]. These techniques often have to be combined with preprocessing knowledge in order to effectively solve segmentation difficulties for a most needed application. One of the main purpose of proposed method is to precisely segment the image without the misplace of imperative information. Almost all image segmentation techniques are ad hoc in nature. Topological derivatives approach is used for image segmentation after best suited restoration process. Figure 3 demonstrates the experimental results of the segmentation work. A. Topological derivative

Figure 4: Concept of Topological Derivative

a. Discrete topological derivative The algorithm is based on discrete topological derivative concept proposed by larrabide in which the cost functional used for discrete approach is represented by the eqn (5)

Ψ (Ωt ) = ∑∑ k s , p ΔΩt .ΔΩt
s s, p s pen

s, p

-----------(5)

The Topological Derivative is given by the difference between perturbed cost function and original cost function. b. Continuum topological derivative Continuum (set theory), is known as the real line or the corresponding cardinal number. Continuum (theory), is nothing but, anything that goes through a gradual transition from one condition, to a different condition, without any abrupt changes. A single point (in the unique topology on a single point set) is a continuum ("is trivially a continuum", meaning that it satisfies the properties is easy and that the result is uninteresting); a continuum that contains more than one point. It is same as discrete but the design vector is b = φ2………………..φn}[18].These structure is {φ1, modeled/analyzed as a continuum. Analysis models can therefore be large and expensive.

43

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

PSNR difference and iteration. This reinforces the evaluation effort more suitably. The performance has its own uniqueness over the images taken. Figure 7 and 8 explain the objective evaluation of topological segmentation before and after filtering. Figure 9 demonstrates the performance evaluation of isotropic filter over topological derivative segmentation of sign images. It estimates the concert of segmentation methods reviewed for this work through iteration.

Figure 5: Image segmentation using topological derivative algorithm (a) Original image (b) continuum derivative (c) Discrete derivative

V. PROPOSED WORK Conventional problems of segmentation are known to be complicated and always in need of research as it implies mainly for object fragmentation. This research work is the expansion of enhancement and the need for the restoration filter before actual segmentation. Every joint framework outcome seems to be more prospective. This effort is validated with both subjective and objective experiments. The special significance of it is, it obtains promising results for isotropic filter.

Figure 8 Comparison of continuum topology before and after filtering

Figure 6: Image segmentation using topological derivative algorithm after enhancement filtering of image (a) Original image (b) continuum derivative (c) Discrete derivative

VI PERFORMANCE EVALUATION It is always vital when evaluating the performance of an algorithm when it is addressing a specific application. The relative performance of restoration and proposed method were evaluated and compared using the MSE criterion,

Figure 8 Comparison of Discrete topology before and after filtering

44

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

Figure 9: Performance of the topology using the values of iterations

VII. CONCLUSION The approach over here is to prove the significance of isotropic filter with topological derivative segmentation. The evaluation of image segmentation techniques is a key field of this study. The evaluation is categorized as objective and subjective. The result over the combination of isotropic filter with segmentation has given hopeful out come. The results indicate that the proposed approach is more robust and accurate than conventional segmentation methods mainly for foreground/ background segmentation evaluation problem. The computational time and number of iteration can be comparatively condensed by using optimization technique in further research. REFERENCES
[1] O. Friman, M. Borga, P. Lundberg, and H. Knutsson. Adaptive analysis of MRI data. NeuroImage, 19(3):837–845, 2003. [2] O. Friman, J. Carlsson, P. Lundberg, M. Borga, and H. Knutsson. Detection of neural activity in functional MRI using canonical correlation analysis. Magnetic Resonance in Medicine, 45(2):323–330, February 2001. [3] H. Knutsson, R. Wilson, and G. H. Granlund. Anisotropic nonstationary image estimation and its applications—Part I: Restoration of noisy images. IEEE Transactions on Communications, 31(3):388–397, March 1983. [4] R. R. Nandy, C. G. Green, and D. Cordes. Canonical correlation analysis and modified ROC methods for fMRI techniques.In Proceedings of the ISMRM Annual Meeting (ISMRM’02), Hawaii, USA, May 2002. ISMRM. [5] Marroquín, J. L., Arce, E., and Botello, S., 2003, Hidden Markov Measure Field Models for Image Segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 25, No. 11, pp. 1380-1387. [6]Orun, A. B., 2004, Automated Identification of Man-Made Textural Features on Satellite Imagery by Bayesian Networks. Photogrametric Engineering & Remote Sensing, Vol. 70, No. 2, pp. 211-216. [7] J.J. Kuch and T.S. Huang, “Vision-based hand modeling and tracking for virtual teleconferencing and telecollaboration,” in Proc. IEEE Int.Conf. Computer Vision, Cambridge, MA, pp. 666-671. June 1995. [8] J. Davis and M. Shah, “Visual gesture recognition,” Vision, Image, and Signal Processing, vol. 141, pp. 101-106, Apr. 1994.

[9] J. Rehg and T. Kanade, “DigitEyes: vision-based human hand tracking,” School of Computer Science Technical Paper CMU-CS-93220,Carnegie Mellon Univ., Dec.1993. [10] Y. Shirai, N. Tanibata, N. Shimada, “Extraction of hand features forrecognition of sign language words,” VI'2002,ComputerControlledMechanical Systems, Graduate School of Engineering, Osaka University, 2002. [11] C. Nölker, H. Ritter, “Detection of Fingertips in Human Hand Movement Sequences,” Gesture and Sign Language inHuman-Computer Interaction, I. Wachsmuth and M. FroÈhlich, eds., pp.209-218, 1997. [12] B. Bauer and H. Hienz, “Relevant features for video-based continuous sign language recognition,” in Proc. of Fourth IEEE InternationalConference on Automatic Face and Gesture Recognition, pp. 440-445,March 2000. [13] Y. Hamada, N. Shimada and Y. Shirai, “Hand Shape Estimation under Complex Backgrounds for Sign Language Recognition” , in Proc. of 6thInt. Conf. on Automatic Face and Gesture Recognition, pp. 589-594,May 2004. [14] V. Mezaris,1,2 I. Kompatsiaris2 and M. G. Strintzis1,2 ,“Still Image Objective Segmentation Evaluation using Ground Truth”, 1 Information Processing Laboratory, Electrical and computer Engineering Department,Aristotle University of Thessaloniki, Thessaloniki 54124, Greece 2 Informatics and Telematics Institute, 1st Km Thermi-Panorama Rd, Thessaloniki 57001, Greece, 5th COST 276 Workshop (2003), pp. 9– 14. [15] Auroux,D., Masmoudi,M., Belaid,L., 2006. ‘Image restoration and classification by topological asymptotic expansion,’ Variation Formulations in Mechanics: Theory and Applications- CIMNE, Barcelona, Spain (In press), (pp.1-16). [16]Dirk Selle, Wolf Spindler, Bernhard Preim, and Heinz-Otto Peitgen., 2002. ‘Mathematical Methods in Medical Imaging: Analysis of Vascular Structures for Liver Surgery Planning, (pp.1-21) [17]Larrabid,I., Novotny,A.A., Feijo’o,R.A., and Taroco,E., 2005. ‘A medical image enhancement algorithm based on topological derivative and anisotropic diffusion,’ Proceedings of the XXVI Iberian Latin-American Congress on Computational Methods in Engineering CILAMCE, (pp1-14). [18]Larrabide,I., Novotny,A.A., Feijo’o,R.A., Taraco,E., and Masmoudi,M., 2005 ‘An image segmentation method based on a discrete version of the topological derivative,’ World Congress Structural and Multidisciplinary Optimization 6, Rio de Janeiro., International Society for Structural and Multidisciplinary Optimization, (pp.1-11).

AUTHORS PROFILE

Ms. M. Krishnaveni, 3 Years of Research Experience Working as Research Assistant in Naval Research Board project, Area of Specialization: Image Processing, Pattern recognition, Neural Networks. Email id: krishnaveni.rd@gmail.com. She has 10 publications at national and International level journals and conferences.

Dr. V. Radha, more than 20 years of teaching experience as Reader. Area of Specialization: Image Processing, Optimization Techniques, Voice Recognition and Synthesis. Email id: radharesearch@yahoo.com. She has 20 publications at national and international level journals and conferences.

45

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No.3, 2009

A Framework for Validation of Object-Oriented Design Metrics
Devpriya Soni1
Department of Computer Applications Maulana Azad National Institute of Technology (A Deemed University) Bhopal 462007 India

Ritu Shrivastava2, M. Kumar3
SIRT, Bhopal (India)

Abstract: A large number of metrics have been proposed for the quality of object-oriented software. Many of these metrics have not been properly validated due to poor methods of validation and non acceptance of metrics on scientific grounds. In the literature, two types of validations namely internal (theoretical) and external (empirical) are recommended. In this study, the authors have used both theoretical as well as empirical validation for validating already proposed set of metrics for the five quality factors. These metrics were proposed by Kumar and Soni. Keywords- object-oriented software, metrics, validation.

2. The product metric is associated with some important external metric (such as measures of maintainability or reliability). 3. The product metric is an improvement over existing product metrics. An improvement can mean, for example, that it is easier to collect the metric or that it is a better predictor of faults. According to Fenton [4], there are two types of validation that are recognized: internal and external. Internal validation is a theoretical exercise that ensures that the metric is a proper numerical characterization of the property it claims to measure. Demonstrating that a metric measures what it purports to measure is a form of theoretical validation. External validation involves empirically demonstrating points (2) and (3) above. Internal and external validations are also commonly referred to as theoretical and empirical validation respectively [2]. Both types of validation are necessary. Theoretical validation requires that the software engineering community reach a consensus on what are the properties for common software maintainability metrics for object-oriented design. Software organizations can use validated product metrics in at least three ways: to identify high risk software components early, to construct design and programming guidelines, and to make system level predictions. The approaches used in two validations are shown in Figure 1.

I.

INTRODUCTION

Analyzing object-oriented software in order to evaluate its quality is becoming increasingly important as the paradigm continues to increase in popularity. A large number of software product metrics have been proposed in software engineering. While many of these metrics are based on good ideas about what is important to measure in software to capture its complexity, it is still necessary to systematically validate them. Recent software engineering literature has shown a concern for the quality of methods to validate software product metrics (e.g., see [1][2][3]). This concern is due to fact that: (i) common practices for the validation of software engineering metrics are not acceptable on scientific grounds, and (ii) valid measures are essential for effective software project management and sound empirical research. For example, Kitchenham et.al. [2] write: "Unless the software measurement community can agree on a valid, consistent, and comprehensive theory of measurement validation, we have no scientific basis for the discipline of software measurement, a situation potentially disastrous for both practice and research." Therefore, to have confidence in the utility of the many metrics those are proposed from research labs, it is crucial that they are validated. The validation of software product metrics means convincingly demonstrating that: 1. The product metric measures what it purports to measure. For example, that a coupling metric is really measuring coupling.

METRIC DEFINITION

THEORETICAL VALIDATION Property based approach Measure ment Theorybased

EMPIRICAL VALIDATION

Experiment

Case Stud ies

Surveys

Figure 1 Approaches to software metrics validation

46

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No.3, 2009

Recently, Kumar and Soni [5] have proposed a hierarchical model to evaluate quality of object-oriented software. This proposed model has been used for evaluation of maintainability assessment of object-oriented design quality, especially in design phase, by Soni and Kumar [6]. In this paper, the authors have attempted to validate the hierarchical model of objectoriented design quality metrics as given in [5]. The section II deals with theoretical validation of the model and the section III deals with empirical validation. II. THEORETICAL VALIDATION OF PROPOSED HIERARCHICAL MODEL OF METRICS

3) Quantifying distance between measurement abstraction: This activity requires the definition of a distance measure for the element of M. Basically this means that the distance defined in the previous activity are now quantified by representing i.e. measuring them as the number of elementary transformation by representing i.e. measuring them as the number of elementary transformations in the shortest sequence of elementary transformation between elements. Formally, the activity results in the definition of a metric MxM→R that can be used to map the distance between a pair of elements in M to a real number. 4) Finding a reference abstraction: This activity require a kind of thought experiment. We need to determine what the measurement abstraction for the object in P would look like if they were characterized by the theoretical lowest amount pty. If such a hypothetical measurement abstraction can be found, then this object is called the reference abstraction for P with respect to pty. 5) Defining a measure for the property: The final activity consists of defining a measure for pty. Since properties are formally defined as distances, and these distances are quantified with a metric function, the formal outcome of this activity is the definition of a function μ:P→R such that p Є P: μ(p)= δ(abs(p), ref(p)).

The main goal of theoretical validation is to assess whether a metric actually measures what it purports to measure [7]. In the context of an empirical study, the theoretical validation of metrics establishes their construct validity, i.e. it ‘proves’ that they are valid measures for the constructs that are used as variables in the study. There is not yet a standard, accepted way of theoretically validating software metric. Work on theoretical validation has followed two paths (see Fig 1): • Measurment-theory based approach such as those proposed by Whitmire[8], Zuse[9], and Poels and Dedene [10] Property-based approach (also called axiomatic approaches), such as proposed by Weyuker and Braind et al.[11]

•

For the theoretical validation DISTANCE framework proposed by Poels and Dedene[9], is a conceptual framework for software metric validation grounded in measurement theory. This is briefly described in the next section. A. The DISTANCE Measure Construction Procedure The measure construction procedure prescribes five activities. The procedure is triggered by a request to construct a measure for a property that characterizes the element of some set of objects. The activities of the DISTANCE procedure are given below. For notational convenience, let P be a set of objects that are characterized by some property pty for which a measure needs to be constructed. 1) Finding a measurement abstraction:The object of interest must be modeled in such a way that the property for which a measure is needed is emphasized. A suitable representation, called measurement abstraction hereafter, should allow to what extent an object is characterized by the property to be observed. By comparing measurement abstraction we should be able to tell whether an object is more, equally or less characterized by the property than other object. 2) Defining distance between measurement abstraction: This activity is based on a generic definition of distance that hold for elements in a set. To define distance between elements in a set, the concept of ‘elementary transformation function’ is used.

B. Metric Validation The proposed model of Kumar and Soni [5] is reproduced in Fig 2 for ready reference. We have used the five activities of DISTANCE measure procedure for metrics of the model and important metrics are summarized in Table 1 III. EMPIRICAL VALIDATION OF THE PROPOSED METRICS

We have seen that survey is also commonly used method to empirically validate defined metrics. To obtain the view of persons who have fair experience of the software design and development, a questionnaire was prepared to validate metrics defined in the Fig 2. The questionnaire used for views is given in the appendix A. The first and second column respectively contains metrics names and their definitions. The respondents were asked to solicit their opinion in the form of yes, no or partially depending upon the metric effects on the five main quality factors, namely functionality, effectiveness, understandability, reusability and maintainability. The questionnaire was sent generously to two groups of people, the professionals working in industry like Infosys, TCS, Wipro, Accenture and people from academic institutes. We received 52 responses of which nearly 70% are from industry professionals and the rest from academic institutes. The analysis of the responses is done using Excel 2007. The results are since significant at 95% confidence level, on the whole if represents the opinion fairly. The analysis is presented in the next section.

47

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No.3, 2009

1 Functionality 1.1 Design Size 1.1.1 Number of Classes (NOC) 1.2 Hierarchies 1.2.1 Number of Hierarchies (NOH) 1.3 Cohesion 1.3.1 Cohesion Among Methods of Class (CAM) 1.4 Polymorphism 1.4.1 Number of Polymorphic Methods (NOP) 1.5 Messaging 1.5.1 Class Interface Size (CIS) 2 Effectiveness 2.1 Abstraction 2.1.1 Number of Ancestors (NOA) 2.1.2 Number of Hierarchies (NOH) 2.1.3 Maximum number of Depth of Inheritance (MDIT) 2.2 Encapsulation 2.2.1 Data Access Ratio (DAR) 2.3 Composition 2.3.1 Number of aggregation relationships (NAR) 2.3.2 Number of aggregation hierarchies (NAH) 2.4 Inheritance 2.4.1 Functional Abstraction (FA) 2.5 Polymorphism 2.5.1 Number of Polymorphic Methods (NOP) 3 Understandability 3.1 Encapsulation 3.1.1 Data Access Ratio (DAR) 3.2 Cohesion 3.2.1 Cohesion Among Methods of Class (CAM) 3.3 Inheritance 3.3.1 Functional Abstraction (FA) 3.4 Polymorphism 3.4.1 Number of Polymorphic Methods (NOP)
Figure 2 Proposed hierarchical design quality model

4 Reusability 4.1 Design Size 4.1.1 Number of Classes (NOC) 4.2 Coupling 4.2.1 Direct Class Coupling (DCC) 4.3 Cohesion 4.3.1 Cohesion Among Methods of Class (CAM) 4.4 Messaging 4.4.1 Class Interface Size (CIS) 5 Maintainability 5.1 Design Size 5.1.1 Number of Classes (NOC) 5.2 Hierarchies 5.2.1 Number of Hierarchies (NOH) 5.3 Abstraction 5.3.1 Number of Ancestors (NOA) 5.4 Encapsulation 5.4.1 Data Access Ratio (DAR) 5.5 Coupling 5.5.1 Direct Class Coupling (DCC) 5.5.2 Number of Methods (NOM) 5.6 Composition 5.6.1 Number of aggregation relationships (NAR) 5.6.2 Number of aggregation hierarchies (NAH) 5.7 Polymorphism 5.7.1 Number of Polymorphic Methods (NOP) 5.8 Documentation 5.8.1 Extent of Documentation (EOD)

48

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No.3, 2009

TABLE I. DISTANCE BASED VALIDATION CRITERIA FOR METRICS Quality Attribute Metrics
Measurement Abstraction Object is more, equally or less characterized by the property than another object.

Validation criteria
Defining distance between two extreme abstractions A set Te of elementary transformation function, sufficient to change any element of M into any other element of M. Various Classes available in the design Various Classes available in the design Various Classes available in the design Various Classes in the hierarchy Various classes/ objects/attributes Class/methods/ parameters Quantifying Distance in extremes. M x M→R to map distance between a pair of elements in M to a real number. Hypothetical reference abstraction Reference abstraction as a reference point for measurement. EQ=1 if 8 or more classes EQ=1 hierarchy level is 5 or more EQ=1 if 6 or more ancestors EQ=1 depth is 6 level or more EQ=1 if aggregation hierarchy 5 or more EQ=1 if cohesion is between 5 or more classes EQ=1 if methods with polymorphic behavior 5 or more EQ=1 if ratio is 80% or more Defining a measure for pty µ:P → R such that pЄ P:μ(p)= δ (abs(p), ref(p))

Functionality
Number of Classes (NOC) Number of Hierarchies (NOH) Number of Ancestors (NOA) Maximum Depth of Inheritance (MDIT) Number of Aggregation Hierarchies (NAH) Cohesion Among Methods of Class (CAM) Number of Polymorphic Methods (NOP) Total number of classes in the design Number of class hierarchies in the design Number of classes along all paths from the root class (es) to all classes in an inheritance. Longest path from the class to the root of the hierarchy. Total number of aggregation hierarchies. Summation of the intersection of parameter of a method with the maximum independent set of all parameter types in the class. Total methods exhibiting polymorphic behavior. Ratio of the number of private (protected) attributes to the total number of attributes declared in the class. Ratio of the number of methods inherited by a class to the total number of methods accessible by member methods of the class. Count of classes that are directly related by attribute declarations and message passing (parameters) in methods. Number of public methods in a class. Number of methods defined in a class. Number of data declarations whose types are user-defined classes. Based on the documentation availability

EQ={1,.8,.6,.4,.2,0} EQ={1,.8,.6,.4,.2,0} EQ={1,.8,.6,.4,.2,0}

EQ=0 if no classes EQ=0 if no hierarchy EQ=0 if no ancestors EQ=0 if depth is 1 level EQ=0 if no aggregation hierarchy EQ=0 if no cohesion among methods EQ=0 if no methods with polymorphic behavior EQ=0 if ratio is less than 5%

EQ={1,.8,.6,.4,.2,0}

Effectiveness

EQ={1,.8,.6,.4,.2,0}

EQ={1,.8,.6,.4,.2,0}

Classes/methods

EQ={1,.8,.6,.4,.2,0}

Understandability

Data Access Ratio (DAR)

Private/Protected attributes and total attributes. Classes/methods

EQ={1,.8,.6,.4,.2,0}

Functional Abstraction (FA)

EQ={1,.5,0}

EQ=1 if ratio is 80% or more

EQ=0 if ratio is less than 5%

Direct Class Coupling (DCC)

methods/parameters passing mechanism

EQ={1,.5,0}

Reusability
Class Interface Size (CIS) Number of Methods (NOM) Number of Aggregation Relationship s (NAR) Extent of Documentati on (EOD)

Input / output parameter Classes/methods

EQ={1,.5,0}

EQ={1,.8,.6,.4,.2,0}

Maintainability

Various classes/ object attributes Data dictionary present or not

EQ={1,.8,.6,.4,.2,0}

EQ=1 if message passing is upto 5 or more classes EQ=1 if public methods present are more than 5 EQ=1 if methods per class are 6 or more EQ=1 if number is more than 6 EQ=1 if documentation is upto 100%

EQ=0 if no. of classes is 1 or less EQ=0 if public method absent EQ=0 if no methods EQ=0 if no aggregation relationship EQ=0 if Documentation is upto 5%

EQ={1,.8,.6,.4,.2,0}

49

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No.3, 2009

A. Observations 1) Number of Classes (NOC): The Figure 3 illustrates that number of classes affects various quality factors in one way or other. 92.31% respondents agree that functionality gets affected by NOC. 90.38% have opinioned that maintainability gets affected by NOC and over 76.92% respondents agree that reusability gets affected by NOC.
Number of classes 100 80 60 Data 40 20 0
y E ffe ct iv en es s U nd er st an da bi lit y ty M ai nt ai na bi lit y Fu nc tio na lit R eu sa bi li

Yes No Partial

Qality Factors

Figure 3 Impact of NOC on quality factors

2) Number of Hierarchies (NOH): The Figure 4 illustrates that number of hierarchies affects various quality factor in one way or other. 90.38% respondents agree that functionality gets affected by NOH. While 88.46% believed that effectiveness gets influenced by NOH. 78.85% have opinioned that maintainability gets affected by NOH.
100 90 80 70 60 Data 50 40 30 20 10 0

Number of Hierarchies

6) Number of Ancestors (NOA): 88.46% respondents agree that effectiveness gets affected by NOA. While 78.85% believed that maintainability gets influenced by NOA. 7) Maximum Depth of Inheritance (MDIT): 90.39% respondents agree that effectiveness gets affected by MDIT. 8) Data Access Ratio (DAR): 86.54% believed that understandability gets influenced by DAR. While 84.62% respondents agree that effectiveness gets affected by DAR. and over 76.92% respondents agree that maintainability gets affected by DAR. 9) Number of Aggregation Relationships (NAR):84.62% respondents agree that maintainability gets affected by NAR. While 78.85% believed that effectiveness gets influenced by NAR. 10) Number of Aggregation Hierarchies (NAH): 82.69% respondents agree that effectiveness gets affected by NAH. While 80.77% believed that maintainability gets influenced by NAH. 11) Functional Abstraction (FA): 80.77% respondents agree that understandability gets affected by FA. While 78.85% believed that effectiveness gets influenced by FA. 12) Direct Class Coupling (DCC): 84.62% respondents agree that reusability gets affected by DCC. While 80.77% believed that maintainability gets influenced by DCC. 13) Number of Methods (NOM): 82.69% respondents agree that maintainability gets affected by NOM. 14) Extent of Documentation (EOD): 75% respondents agree that maintainability gets affected by EOD.

Yes No Partial

Data

Quality Factors

Figure 4 Impact of NOH on quality factors

3) Cohesion Among Methods of Class (CAM): 90.38% believed that understandability gets influenced by CAM. 84.62% have opinioned that reusability gets affected by CAM. 82.69% respondents agree that functionality gets affected by CAM. 4) Number of Polymorphic Methods (NOP): 86.54% respondents agree that understandability gets affected by NOP. 80.77% have opinioned that functionality gets affected by NOP. While 78.84% believed that maintainability gets influenced by NOP and over 75% respondents agree that effectiveness gets affected by NOP. 5) Class Interface Size (CIS): 90.38% respondents agree that functionality gets affected by CIS. While 82.69% believed that reusability gets influenced by CIS.

94 92 90 88 86 84 82 80 78 76 74 NOC NOH CAM Functionality NOP CIS

lit y

ct iv en

da bi li t y

sa bi lit y

na

tio

er st an

R eu

Fu nc

Ef fe

U nd

M

ai nt a

in a

bi lity

es s

Impact

Figure 5 Impact of metrics on functionality

50

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No.3, 2009
92 90 88 Data 86 Impact 84 82 80 78 NOH NOA MDIT Effectiveness DAR NAH

IV.

CONCLUSION

Figure 6 Impact of metrics on effectiveness

92 90 88 86 Data 84 82 80 78 76 74 CAM NOP DAR FA Understandability Impact

A majority of respondents have opinioned that the metric NOC impacts three quality factors, functionality, maintainability and reusability and hence placement of NOC at these factors is justified. Further, majority respondents have opinioned that the metric NOH impacts three quality factors functionality, effectiveness and maintainability and hence placement of NOH at these factors is justified. Similar interpretations can be provided to other metrics. It is further observed that functionality is critically affected by the metric NOC followed by NOH (see Fig 5). Effectiveness is much affected by MDIT followed by NOH and NOA (see Fig 6). Understandability is much affected by CAM followed by NOP and DAR (see Fig 7). Reusability is much affected by CAM and DCC (see Fig 8). Similarly maintainability is much affected by NOC followed by NAR (see Fig 9). We have considered only five metrics in maintainability, however respondents opinioned that it is also affected by metrics NOH, NOP and NOA.

REFERENCES
N. Fenton and B. Kitchenham, "Validating Software Measures," Journal of Software Testing, Verification and Reliability, vol. 1, no. 2, pp. 2742, 1990. [2] B. Kitchenham, S-L Pfleeger, and N. Fenton, "Towards a Framework for Software Measurement Validation," IEEE Transactions on Software Engineering, vol. 21, no. 12, pp. 929-944, 1995. [3] N. Schneidewind, "Methodology for Validating Software Metrics," IEEE Transactions on Software Engineering, vol. 18, no. 5, pp. 410422, 1992. [4] N. Fenton, "Software Metrics: Theory, Tools and Validation," Software Engineering Journal, pp. 65-78, January, 1990. [5] M.Kumar and D. Soni, “Observations on Object-Oriented Design Assessment and Evolving New Model”, Proc of The National Conference on Software Engineering, pp. 161-164, 2007. [6] D. Soni and M.Kumar “Maintainability assessment of object-oriented design quality” International Journal on Computer Engineering and Information Technology, Vol 2, no 2, pp. 35-41, 2009. [7] R. Van Solingen and E. Berghout, “The Goal/Question/Metric Method: A practical guide for quality improvement of software development”. McGraw-Hill. 1999 [8] J. Whitmire, “Correctly assessing the “ilities” requires more than marketing hype. IT Professional” , Volume 2, Number 6, pp 65-67, 2000. [9] H. Zuse, “A Framework of Software Measurment, Walter de Gruyter Berlin”, 1998. [10] G. Poels and G. Dedene, “Distance-based software measurement: Necessary and sufficient properties for software measures. Information and Software Technology” , Volume 42, Number 1, pp 35-46, 2000 [11] L. Briand, S. Morasca and V. Basili, “An operational process for goaldriven definition of measures”. IEEE Transaction on Software Engineering Volume 30, Number 2, pp 120-140, 2002. [1]

Figure 7 Impact of metrics on understandability

86 84 82 Data 80 Impact 78 76 74 72 NOC CAM CIS DCC Reusability

Figure 8 Impact of metrics on reusability

92 90 88 86 Data 84 82 80 78 76 74 NOC NAR NAH Maintainability DCC NOM Impact

Figure 9 Impact of metrics on maintainability

51

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No.3, 2009

Appendix A The following questionnaire was sent to respondents.
Factor Metric Name Number of Classes (NOC) Number of Hierarchies (NOH) Cohesion Among Methods of Class (CAM) Definitions
Total number of classes in the design Number of class hierarchies in the design Summation of the intersection of parameter of a method with the maximum independent set of all parameter types in the class. Total methods exhibiting polymorphic behavior. Number of public methods in a class. Number of classes along all paths from the root class (es) to all classes in an inheritance. Longest path from the class to the root of the hierarchy. Ratio of the number of private (protected) attributes to the total number of attributes declared in the class. Number of data declarations whose types are user-defined classes. Total number of aggregation hierarchies. Ratio of the number of methods inherited by a class to the total number of methods accessible by member methods of the class. Count of classes that are directly related by attribute declarations and message passing (parameters) in methods. Number of methods defined in a class. Based on the documentation availability

Functionality

Effectiveness

Understandability

Reusability

Maintainability

Number of Polymorphic Methods (NOP) Class Interface Size (CIS) Number of Ancestors (NOA) Maximum Depth of Inheritance (MDIT) Data Access Ratio(DAR)

Number of aggregation relationships (NAR) Number of aggregation hierarchies (NAH) Functional Abstraction (FA)

Direct Class Coupling (DCC)

Number of Methods (NOM) Extent of Documentation (EOD)

AUTHORS PROFILE

Devpriya Soni has seven years of teaching experience to post graduate classes at MANIT and four years of research experience. She is pursuing her PhD at Department of Computer Applications, MANIT, Bhopal. Her research interest is object-oriented metrics and object-oriented databases. Ritu Shrivastava has 12 years of teaching experience to graduate classes at MANIT and 2 years at Amity university at Delhi. She is pursuing research in object-oriented software engineering.

Dr Mahendra Kumar is presently Prof. & Dean of Computer Science at S I R T. Bhopal. He was Professor and Head Computer applications at M A N I T. Bhopal. He has 42 years of teaching and research experience. He has published more than 90 papers in National and International journals. He has written two books and guided 12 candidates for Ph D degree and 3 more are currently working. His current research interests are software engineering, cross language information retrieval, data mining, and knowledge management.

52

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

A New Image Steganography Based On First Component Alteration Technique
Amanpreet Kaur1, Renu Dhir2, and Geeta Sikka3
1,2,3

Department of Computer Science and Engineering. National Institute of Technology, Jalandhar, India

Abstract—In this paper, A new image steganography scheme is proposed which is a kind of spatial domain technique. In order to hide secret data in cover-image, the first component alteration technique is used. Techniques used so far focuses only on the two or four bits of a pixel in a image (at the most five bits at the edge of an image) which results in less peak to signal noise ratio and high root mean square error. In this technique, 8 bits of blue components of pixels are replaced with secret data bits. Proposed scheme can embed more data than previous schemes and shows better image quality. To prove this scheme, several experiments are performed, and are compared the experimental results with the related previous works. Keywords—image; mean square error; Peak signal to moise ratio; steganography.

I. INTRODUCTION Due to rapid development in both computer technologies and Internet, the security of information is regarded as one of the most important factors of Information Technology and communication. Accordingly, we need to take measures which protect the secret information. Generally, secret information may be hidden in one of two ways, such as cryptography and steganography. The methods of cryptography makes the data unintelligible to outsiders by various transformations, whereas the methods of steganography conceal the existence of messages. The word steganography is derived from Greek words meaning “covered writing” and as such it centers on the concept of hiding a message. As defined by Cachin [3], steganography is the art and science of communicating in such a way that the presence of message is detected. Steganography is very old method used. Around 440 B.C., Histiaeus shaved the head of his most trusted slave and tattooed it with a message which disappeared after the hair had regrown. The purpose of this message was to instigate a revolt against the Persians. Another slave could be used to send a reply. During the American Revolution, invisible ink which would glow over a flame was used by both the British and Americans to communicate secretly. Steganography was also used in both World Wars. German spies hid text by using invisible ink to print small dots above or below letters and by changing the heights of letter-strokes in cover texts.


Among the methods of steganography, the most common thing is to use images for steganography. This is called image steganography. In this image hiding method, the pixels of images are changed in order to hide the secret data so as not to be visible to users, and the changes applied in the image are not tangible. The image used to camouflage the secret data is called the cover-image while the cover-image with the secret data embedded in it is called the stego-image. Image steganographic techniques can be divided into two groups [7]: the Spatial Domain technique group, and the Transform Domain technique group. The Spatial domain technique embeds information in the intensity of the pixels directly, while the Transform domain technique embeds information in frequency domain of previously transformed image. Our proposed scheme is a kind of the spatial domain techniques.
II. RELATED WORKS

A. Least Significant Bit Hiding (LSB) Scheme This method is probably the easiest way of hiding information in an image. In the LSB technique, the LSB of the pixels is replaced by the message to be sent. The message bits are permuted Before embedding, which has the effect of distributing the bits evenly, thus on average only half of the LSB’s will be modified. B. Pixel-Value Differencing (PVD) scheme The alteration of edge areas in the human visual system cannot be distinguished well, but the alteration of smooth areas can be distinguished well. That is, an edge area can hide more secret data than a smooth area. With this concept, Wu and Tsai proposed a novel steganography technique using the pixel-value differencing (PVD) method to distinguish edge and smooth areas. The PVD technique can embed more data in the edge area which guarantees high imperceptibility. C. Lie-Chang’s scheme The Steganographic technique has to possess two important properties. These are good imperceptibility and sufficient data capacity. A scheme which satisfied both properties was proposed by Lie-Chang [5]. The scheme is an Adaptive LSB technique using Human Visual System (HVS). HVS has the following characteristics: Just Noticable Difference (JND), Contrast Sensitivity Function (CSF), Masking and Spectral
53 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Sensitivity. The characteristic of HVS used by Lie-Chang is JND (also known as the visual increment threshold or the luminance difference threshold). In this scheme, JND is defined as the amount of light ΔI necessary to add to a visual field of intensity I such that it can be distinguished from the background. In HVS, the curve for ΔI versus I can be analytically and mathematically modeled. The JND technique is simple and has a higher embedding capacity than other schemes. Also, this technique has high embedding capacity about overall bright images and has high distortion of a cover image when the embedding capacity is increased, but does not concern overall dark images. D. MSB3 edge-detection Generally, the human eye is highly sensitive to overall pictures of the field of view, while having low senstivity to fine details. Such characteristic of HVS is called the CSF. One of several computational models which explain the CSF is proposed by Mannos-Sakrison [1]. According to MannosSakrison scheme, if additional data is embedded in the pixels of high spatial frequency, one is able to satisfy both the increment of hiding capacity and good imperceptibility. In order to judge whether any pixel has the high spatial frequency or the low spatial frequency in a digital image, the edge detection algorithm is generally used. The GAP algorithm is one of the edge detection algorithms. For the input value of the GAP algorithm, we use the technique of the using three bits from the MSB. Due to this, the pixels that are selected from an embedding phase must be equal to pixels that are selected from an extracting phase. Three bits are embedded in a pixel if the pixel-value is smaller than the first threshold value (intensity 88) and is judged with the edgeregion. MSB3 Edge-Detection is summarized through the following steps: Step 1: Execute MSB3 Edge-Detection at a cover image, in order to sort out edge- regions in the cover image. Step 2: If any pixel value is smaller than the first threshold value and exists on the edgeregion, embed three bits of secret data in the pixel. E. Image Steganography Based on 2k Correction and Edge- Detection Scheme In this method, author used the just noticeable difference (JND) technique and method of contrast sensitivity function (CSF). This is an MSB3 edge-detection which uses part information of each pixel-value. In order to have better imperceptibility, a mathematical method which is the 2k correction is used. If one supposes the secret data is hidden at a pixel of cover image, some differences occurred between cover-pixel and stego-pixel. Because of these differences, the cover image is distorted and the quality of cover image is dropped. 2k correction corrects each pixel-value as 2k. That is, supposing that k-bits are embedded in a pixel value, the method adds or subtracts 2k to each pixel-value, and finally the corrected pixel value becomes closer to the original-pixel. Hence, the secret data in the stego-pixel is not changed.
54

This scheme can embed more data than previous schemes, and shows better imperceptibility. The method is an edgedetection which only uses 3-bits from MSB of each pixelvalue. In this method, data embedment depends on the value received from each pixel value whether it is on the edge or on other part of an image. If it is on the edge it embed data in the cover image based on the value of k, value of k is decided on the pixel position whether it is on the edge or not. This method modifies the stego pixel value near to the cover pixel using 2k correction mathematical formula. III. PROPOSED IMAGE STEGANOGRAPHY SCHEME In the proposed scheme, a new image steganography scheme based on first componenet alteration technique is introduced. In a computer, images are represented as arrays of values. These values represent the intensities of the three colors R (Red), G (Green) and B (Blue), where a value for each of three colors describes a pixel. Each pixel is combination of three components(R, G, and B). In this scheme, the bits of first component (blue component) of pixels of image have been replaced with data bits, which are applied only when valid key is used. Blue channel is selected because a research was conducted by Hecht, which reveals that the visual perception of intensely blue objects is less distinct that the perception of objects of red and green. For example, suppose one can hide a message in three pixels of an image (24-bit colors). Suppose the original 3 pixels are: (00100111 11101001 11001000) (00100111 11001000 11101001) (11001000 00100111 11101001) A steganographic program could hide the letter "A" which has a position 65 into ASCII character set and have a binary representation "01000001", by altering the blue channel bits of pixels. (01000001 11101001 11001000) (00100111 11001000 11101000) (11001000 00100111 11101001) A. Embedding phase The embedding process is as follows. Inputs: Image file and the text file Output: Text embedded image Procedure: Step 1: Extract all the pixels in the given image and store it in the array called Pixel-Array. Step 2: Extract all the characters in the given text file and store it in the array called Character- Array. Step 3: Extract all the characters from the Stego key and store it in the array called Key- Array. Step 4: Choose first pixel and pick characters from KeyArray and place it in first component of pixel. If there are more characters in Key- Array, then place rest in the first component of next pixels, otherwise follow Step (e). Step 5: Place some terminating symbol to indicate end of the key. ‘0’ has been used as a terminating symbol in this algorithm.

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Step 6: Place characters of Character- Array in each first component (blue channel) of next pixels by replacing it. Step 7: Repeat step 6 till all the characters has been embedded. Step 8: Again place some terminating symbol to indicate end of data. Step 9: Obtained image will hide all the characters that we input.

PSNR = Where MAXPIX is the maximum pixel value and RMSE is the Root Mean Square Error of the image (it quantifies the average sum of distortion in each pixel of the encrypted image i.e. average change in pixel caused by encryption algorithm)

B. Extraction phase The extraction process is as follows. RMSE = Inputs: Embedded image file Output: Secret text message Procedure: The results are then compared with various steganography Step 1: Consider three arrays. Let they be Character-Array, methods as shown in the following table. Key-Array and Pixel-Array. Step 2: Extract all the pixels in the given image and store it Lena LSB3PVD Lie Jae GilFirst in the array called Pixel-Array. Step 3: Now, start scanning pixels from first pixel and image Chang’s Yu Component extract key characters from first (blue) component of the alteration pixels and place it in Key-Array. Follow Step 3 till we get technique terminating symbol, otherwise follow step 4. 37.92 41.48 37.53 38.98 46.11 PSNR Step 4: If this extracted key matches with the key entered by the receiver, then follow Step 5, otherwise terminate the program by displaying message “Key is not matching”. Step 5: If the key is valid, then again start scanning next pixels and extract secret message characters from first (blue) component of next pixels and place it in Character Array. Follow Step 5 till we get terminating symbol, otherwise follow step 6. Step 6: Extract secret message from Character-Array.

IV. EXPERIMENTAL RESULTS To evaluate the performance of the proposed scheme, the Image Steganography is firstly applied to Lena’s image as a test image. Different results have been observed with RGB components by changing first component to embed data in it and to measure image quality of the proposed scheme, we used Peak Signal-to-Noise Ratio (PSNR) and the MSE (Mean Square Error) for an Encrypted Image. The results are then compared with various Encryption Method as shown in the table. The PSNR computes the peak signal-to-noise ratio, in decibels, between two images. This ratio is often used as a quality measurement between the original and a compressed image. The higher the PSNR, the better the quality of the compressed or reconstructed image. The MSE (Mean Square Error) represents the cumulative squared error between the compressed and the original image, the lower the value of MSE, the lower the error. To compute the PSNR, the block first calculates the meansquared error using the following equation: MSE =

Figure 2(a) Original Lena image (512 X 512)

Figure 2(b) Stego Lena Image (512 X 512) V. CONCLUSIONS

Where x (m, n) and y (m, n) are the two images of size m*n. In this case x is the original image and y is the encrypted image.

In this paper, a new image steganography scheme which is a
kind of spatial domain technique. In order to hide secret data in cover-image, the first component alteration technique is used. Techniques used so far focuses only on the two or four
55 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

bits of a pixel in a image ,(at most five bits at the edge of an image.) which results less peak to signal noise ratio and high root mean square error i.e. less than 45 PSNR value. Proposed work is concentrated on 8 bits of a pixel (8 bits of blue component of a randomly selected pixel in a 24 bit image), resulting better image quality. Proposed technique has also used contrast sensitivity function (CSF) and just noticeable difference (JND) Model. Proposed scheme can embed more data than previous schemes [7, 5, 10], and shows better imperceptibility. To prove this scheme, several experiments are performed, and the experimental results are compared with the related previous works. Consequently, the experimental results proved that the proposed scheme is superior to the related previous works. The future work is to extend proposed technique for videos and to modify given scheme to improve image quality by increasing PSNR value and lowering MSE value. REFERENCES [1]
J. L. Mannos and D. J. Sakrison. The effects of a visual fidelity criterion on the encoding of images. IEEE Trans. On Information Theory, pages 525–536, 1974. [2] Herodotus, The Hisories, and chap. 5 - The fifth book entitled Terpsichore, 7 - The seventh book entitled Polymnia, J. M. Dent & Sons, Ltd, 1992 [3] C. Cachin, “An Information-Theoretic Model for Steganography”, Proceedings of 2nd Workshops on Information Hiding, MIT Laboratory for Computer Science, May 1998 [4] Joshua R. Smith and Chris Dodge, Developments in Steganography. Proceedings of the Third International Workshop on Information Hiding Pages: 77 – 87, 1999 [5] W. N. Lie and L. C. Chang. Data hiding in images with adaptive number of least significant bits based on the human visual system. Proc. ICIP ’99, 1:286–290, 1999 [6] W. Stallings. Cryptography and Network Security – principles and practices. Pearson Education, Inc., 2003 [7] D. C. Wu and W. H. Tsai. A steganographic method for images by pixelvalue differencing. Pattern Recognition Letters, 24:1613–1626, 2003 [8] Niels Provo and Peter honeyman university of Michigan, Hide and seek – An introduction to steganography, 2003 [9] Jonathan Cummins, Patrick Diskin, Samuel Lau and Robert Parlett, School of Computer Science, The University of Birmingham, Steganography and Digital watermarking, 2004 [10] Jae-Gil Yu, Eun-Joon Yoon, Sang-Ho Shin and Kee-Young Yoo. A New Image Steganography Based on 2k Correction and Edge-Detection. ITNG Proceedings of the Fifth International Conference on Information Technology: New Generations Pages 563-568 , 2008 [11] White Paper on the contributions of cryptography & steganography to internet security. [12] Francesco Queirolo, Steganography in images: Final communication report

56

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Evaluating Effectiveness of Tamper-proofing on Dynamic Graph Software Watermarks
Prof. Dr. Malik Sikandar Hayat Khiyal
Department of Computer Science and Software Engineering Fatima Jinnah Women University Rawalpindi, Pakistan .

Aihab Khan
Department of Computer Science Fatima Jinnah Women University Rawalpindi, Pakistan .

Dr. M. Shahid Khalil
Department of Mechanical Engineering University of Engineering and Technology

Sehrish Amjad
Graduate, Department of Computer Science Fatima Jinnah Women University

Texila, Pakistan
. .

Rawalpindi, Pakistan

Abstract—For enhancing the protection level of dynamic graph software watermarks and for the purpose of conducting the analysis which evaluates the effect of integrating two software protection techniques such as software watermarking and tamper-proofing, constant encoding technique along with the enhancement through the idea of constant splitting is proposed. In this paper Thomborson technique has been implemented with the scheme of breaking constants which enables to encode all constants without having any consideration about their values with respect to the value of watermark tree. Experimental analysis which have been conducted and provided in this paper concludes that the constant encoding process significantly increases the code size, heap space usage, and execution time, while making the tamper-proofed code resilient to variety of semantic preserving program transformation attacks. (Abstract) Keywords-component; contsant watermarking; tamper-proofing; encoding; software

constant encoding technique of tamper-proofing dynamic graph software watermarks was first proposed by Yong He in [17] and then with few modification again proposed by Clark Thomborosn in [16].This novel tamper-proofing method is based on encoding constant of software programs into data structure of watermark tree by the means of protecting against program transformation attacks. Section I of the paper comprised on basic introduction of the problem found in different already proposed constant encoding techniques. In Section II, tamper-proofing techniques of dynamic graph software watermarks along with their limitations are discussed. While Section III provides the understanding of the structure. Section IV deals with the implemented technique. In Section V, outcome of this paper i.e. experimental results are mentioned and several suggestions for further enhancement are described in Section VI. A. Contribution This paper provides the details about the practical experience with the idea of constant encoding techniques, and then also measures the effect of incorporating two software protection techniques which are watermarking and tamperproofing. The purpose of the underlying effort is to find out the proportion of change due to combining two software protection techniques on different parameters such as code size, execution time, heap space usage and resilience. II. RELATED WORK

I.

INTRODUCTION

The most significant property involved in digital information is that, it is in principle very easy to produce and distribute unlimited number of its copies. This might undermine the music, film, book and software industries and therefore it brings a variety of important problems. In this paper the dilemma under consideration is of software piracy, concerning the protection of the intellectual and production rights that badly need to be solved. To overwhelm the problem of software piracy through watermarking, the emerging technique is dynamic graph software watermarking. But the major drawback involve in this technique is the lack of stealthiness. To protect dynamic graph software watermarks against attacks due to lack of resemblance between the watermark code and source program,

To tamper-proof the dynamic and static software watermarks, different methods are employed. But in this paper, our main focus is on the tamper-proofing of Dynamic Graph Watermarks. In our literature search, we find only a few

57

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

publications which describe the tamper-proofing method of Dynamic Graph Watermarks. The only method which is described through out the publications is of constant encoding. Constant encoding method evaluates from the idea of Palsberg described in [21]. In constant encoding, the main concept which is followed is to replace the constant values with function calls. In [17], Yong He described the constant encoding technique for tamper-proofing the DGW. The basic idea behind this technique is the replacement of constant values with function calls. The value return by the function calls is dependent upon the values of pointer variables in the dynamic data structure of same shape as the watermark tree. The graph structure utilized by this technique is of planted plane cubic tree (PPCT) shape. The procedure utilizes by the Yong technique for encoding the constants successfully helps in creating the false dependencies from the watermarked program to the constant tree structure. The distinguishing property of Yong’s technique is handling of large constant due to not having the constraint on the size of the constant tree. But unfortunately the effectiveness of this protection technique depends upon the assumption that the attacker is unable to differentiate between the watermark building code and constant tree building code. In [16], Clark Thomborson proposed the constant encoding technique as the modification of Yong He technique. To overcome the weakness of existence of false dependency between the watermark tree and the constant tree, the idea of Thomborson is based on utilization of watermark tree for finding substructure instead of separately creating the constant tree. The concept provided by the Thomborson reinforce the Yong’s constant encoding technique in this way that even if the attacker successfully locates the watermark, he will not able to remove or modify it without taking risk of changing the constant values. III. FRAMEWORK OVERWIEW

After watermarking, decompilation is performed to convert generated file of byte code into java source code. Constant encoding technique implemented in this paper handles constant of only one type which is of numeric.

A. Framework/Model of Constant Encoding Technique in Collaboration of Constant Splitting Technique Procedure of constant encoding is performed on the watermarked program generated by the SandMark system by selecting the subsequent options. • • • • • • • • • • Watermark Type: Numeric Storage Method: Hash Storage Policy: Formal Protection Method: if: safe: try GraphType:SandMark.util.newgraph.codec.PlantedPla neCubicTree Use Cycle Graph: No Sub graph Count: 1 Inline Code: No Replace Watermark Class: No Dump Intermediate Code: No

Figure 1. Constant Encoding Technique.

The scope of constant encoding technique is limited to the single tree structure which is of planted plane cubic tree shape. B. Framework/Model of Constant Splitting Technique Block diagram mentioned in this section describes the procedure of constant splitting which is essential for handling the encoding of those constants which cannot be encoded due to mismatch of shape.

58

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

In the reverse situation, record the location of root of constant substructure in watermark tree • Step 7: Construct function to retrieve constant substructure in the watermark tree at runtime • Step 8: Generate decoding function to retrieve constant value back from the constant substructure • Step 9: Modify the watermarked source program according to generated decoding function B. Algorithm of Constant Encoding Technique For the understanding of the algorithm, suppose that two variables namely as even and odd are taken having values ‘1’ and ‘0’ respectively • Step 1: Input constant value. • Step 2: Check that constant value is greater than ‘1’ • Step 3: Check whether the constant is even or odd • Step 4: If constant is even, divide it by ‘2’, and move to Step 6. • Step 5: If constant is odd, then first subtract ‘1’ from it and then move to Step 7. • Step 6: Multiply the value of even by ‘2’ and store the result again in even • Step 7: Add ‘1’ in the value of odd and assign it again to odd, and then go to Step 4.
Figure 2. Constant Splitting Technique.

Constant splitting method is recursive in nature and applies until the constant substructure is not found in the watermark tree. IV. TECHNIQUE

A. Algorithm of Constant Encoding Technique • Step 1: Input watermarked source program • Step 2: Select constants for encoding from the source program • Step 3: Build constant graph of the selected constants • Step 4: Search constant substructure of same shape as constant graph from watermark tree • Step 5: If constant substructure is not found in the watermark tree, then apply the constant splitting technique • Step 6:

Our Technique is based on Thomborson’s technique of constant encoding. The implementation and experimental results mentioned in next sections show that constant encoding can be done effectively with dynamic graph watermarking but with reasonable increase in code size, heap space usage and execution time. Our system is built in java language and its target language is also Java. The system takes the watermarked program as input and successfully outputs the constant encoded programs. For handling the encoding of large constant, constant splitting technique is used. The main purpose of constant splitting technique is that even if the whole constant could not encode due to dissimilarity of shapes, then some part of that of that will definitely encoded. Our implemented technique mainly does access the information of watermark tree provided in the watermark class of watermarked program. For referencing the constant substructure at the runtime in watermark tree, different mechanisms are employed. V. EPERIMENTAL RESULTS

We run the experiments under windows XP professionals with 512MB of RAM. The processor used in the experiments is

59

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Pentium 4. We use the J2SDK1.4.2_17 as backend java tool and Realj version 3.1 as front end tool. The system used for watermarking is SandMark v3.4.0, and decompilation process is done through Front End Plus v1.04. We tested our system on different medium sized programs. Each program is watermarked with the same three digit number. The number of encoded constants of watermarked program varied. Further description of the programs used for testing and analysis are given below in table 1 After constant encoding procedure, parameters which include code size, execution time, heap space usage and resilience are evaluated. A. Evaluation of Parameters For the measurement of execution time, code size and heap space usage, each program is executed ‘n’ times. In evaluation of parameters, the comparison is made between the watermarked (WM) and tamper-proofed (TP) applications. Tables and Figures which are mentioned below specify the difference between the values of different parameters of watermarked and tamper-proofed applications along with the specified units used in measurement. 1) Heap space usage: a) Tabular representation: The table I given below distinguishes the usage of heap space after the constant encoding process of watermarked application.
TABLE I. Program 1 2 3 n DIFFERENCE IN HEAP SPACE USAGE 1 Heap Space Usage
Watermark Tamper-proof Difference

Figure 3. Difference in Heap Space Usage.

may vary due to dependency upon the number of constants encoded in each application. So less number of constants to be encoded will yield less increase in the size of heap space. 2) Execution Time: a) Tabular representation: Table II which is provided below illustrates the variation in execution time subsequent to the constant encoding process.
TABLE II. Program DIFFERENCE IN EXECUTION TIME Execution Time
Watermark Tamper-proof Difference

339.405 60.550 323.485 119.840

451.900 69.175 329.755 127.115

12.405 8.625 6.27 7.275

2 3 n

359589.4 145648.0 278706.5 268397.2

418581.2 200507.0 343980.9 254565.6
a.

58991.8 54859.0 68274.4 46168.4

a. Sample of a Table footnote. (Table footnote)

Sample of a Table footnote. (Table footnote)

As the table indicates that no one application has the same incremented difference in heap space usage after constant encoding process. b) Graphical representation: The graphical representation of Table 1 demonstrated in Figure 3 provides the conclusion that after constant encoding process, heap space usage of all watermarked application will definitely increase, but the ratio of increment.

As the table reveals that after constant encoding process, execution time of each watermarked application is increased but proportion of change also differ here. b) Graphical representation: Figure 4 depicts the conclusion that the dissimilarity in execution time of ‘n’ number of programs is due to the values of constants based on which constant graph is generated and the number of times the constant splitting technique is applied.

60

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Figure 4. Difference in Execution Time.

Figure 5. Difference in Code Size.

Thus it can be concluded that constant encoded applications will always have the more execution time then the watermarked application, but the percentage of increment may vary by reason of having total dependency upon the value of constant which will encode in the watermark tree. 3) Code Size: a) Tabular representation: Table III gives an idea about the effect of constant encoding process on the code size of each watermarked program.
TABLE III. Program 1 2 3 n DIFFERENCE IN CODE SIZE Code Size
Watermark Tamper-proof Difference

So number of constants and values of constants to be encoded will always have no effect on the code size of each application. 4) Resilience: Resilience basically measures at what extent the watermarked and tamper-proofed application is unfeasible and invulnerable against transformation attacks such as semantic preserving transformations which include code obfuscation and code optimization. For evaluating the resilience level of watermarked and tamper-proofed application, techniques applied are class splitter, reorder instruction, duplicating the register, field assignment and variable reassigner a) Reorder Instruction Program Transformation Attack: Reorder instruction attack tries to distort the program by reordering the instruction within each basic block of a method. Table IV indicates the consequences of reorder instruction attack and reveals that constant encoding process does not affect the resilience of watermarked application.
TABLE IV. Program 1 2 3 EFFECT OF REORDER INSTRUCTION ATTACK Reorder Instruction Attack
Watermark Tamper-proof

07.53 02.61 07.18 04.52

18.53 13.61 18.18 15.52

11.0 11.0 11.0 11.0

Not affected Not affected Not affected Not affected

Not affected Not affected Not affected Not affected
a. Sample of a Table footnote. (Table footnote)

a. Sample of a Table footnote. (Table footnote)

As it is obvious from the table that after constant encoding process, all the applications has uniform increase in the code size. b) Graphical representation: Figure 5 provides the graphical representation of disparity in code size which is mentioned in Table III. Analysis which is conducted on the basis of Table III and Figure 5 concludes that code size of each watermarked program is incremented by ‘11’ kb and the cause of this augmentation is the amount of code inserted in watermarked program after the constant encoding process.

n

61

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

b) Class Splitter Program Transformatioon Attack: This attack utilizes the technique of splitting class in half by placing some methods and fields to the super class. Table V given below illustrates that constant encoded applications are susceptible to class splitter attack in contradictory to watermarked applications.
TABLE V. Program 1 2 3 n EFFECT OF CLASS SPLITTER ATTACK

TABLE VII. Program 1 2 3 -

EFFECT OF FIELD ASSIGNMENT ATTACK Field Assignment Attack
Watermark Tamper-proof

Not affected Not affected Not affected Not affected

Not affected Not affected Not affected Not affected
a. Sample of a Table footnote. (Table footnote)

Class Splitter Attack
Watermark Tamper-proof

n

Not affected Not affected Not affected Not affected

Affected Affected Affected Affected
a. Sample of a Table footnote. (Table footnote)

c) Duplicate Register Program Transformation Attack: Procedure of duplicate register involves in taking a local variable in a method and then split reference to it with a new variable. Table VI provides the detail about the successful behavior of the watermarked programs after constant encoding process after the attack.
TABLE VI. Program 1 2 3 n EFFECT OF DUPLICATE REGISTER ATTACK Duplicate Register Attack
Watermark Tamper-proof

e) Varaiable Reassigner Program Transformation Attack: Variable reassigner functions in program by reallocating the local variable in order to minimize the use of number of local variable slots. Table VIII demonstrates that variable reassigner attack makes all the watermarked and tamper-proofed applications no more to executable and it also indicates no improvement in resilience level of watermarked application even after the constant encoding process against this attack.
TABLE VIII. Program 1 2 3 n EFFECT OF VARIABLE REASSIGNER ATTACK Variable Reassigner Attack
Watermark Tamper-proof

Affected Affected Affected Affected

Affected Affected Affected Affected
a. Sample of a Table footnote. (Table footnote)

Not affected Not affected Not affected Not affected

Not affected Not affected Not affected Not affected
a. Sample of a Table footnote. (Table footnote)

The analysis which is conducted helps in concluding that the only thing that can affect the resilience of tamper-proofed application is large code size. VI. CONCLUSION AND FUTURE WORK

d) Field Assignment Program Transformation Attack: This attack performs obfuscation by inserting bogus field into class and then making assignment to that field in different locations of the program Table VII depicts the effect of field assignment attack and shows that this attack has no influence on the performance of the tamper-proofed applications.

The implementation of constant encoding tamper-proofing process provided in this paper is promising step for further research. The analysis which has been performed for evaluating the effectiveness of dynamic graph software watermarks concludes with considerable effect of tamper-

62

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

proofing process on dynamic graph software watermarks. It is summarized from the evaluation of parameters that after constant encoding process code size of the application is always increased, but the incremental change in heap space usage and execution time may vary due to dependence upon various other factors. After constant encoding, the level of resilience of watermarked application can also be degraded due to large code size. To make the tamper-proofed code resilient against all types of attacks, one needs to improve the constant encoding process by inserting opaque predicated or through some other obfuscation techniques. In this paper, constant splitting technique is adopted in combination with Clark Thomborson’s technique to handle all the constants which are greater or smaller than the watermark value. Another better approach which can improve this integration of software protection techniques to great extent is that, instead of breaking up the constant value, splits the constant graph into subgraphs in situation where constant substructure is not found in watermark tree, though it would results in difficulty of tracking all the references related to the roots of subgraphs, but this method can provide high level of protection against attacks. ACKNOWLEDGMENT

[7]

[8]

[9]

[10]

[11] [12]

[13]

[14]

[15]

[16]

REFERENCES
[17] [1] [2] Erin Joyce “Software Piracy Losses Add Up to $29B.”July 8, 2004. http://www.enterpriseitplanet.com/security/news/article.php/3378251 Diego Bendersky, Ariel Futoransky, Luciano Notarfrancesco, Carlos Sarraute, and Ariel Waissbein, “Advanced Software Protection Now”. Corelabs, Core Security Technologies; Departamento de Computaci´on, FCEyN, Departamento de Matem´atica, FCEyN, UBA Argentina. Mohannad Ahmad AbdulAziz Al Dharrab, “ Benchmarking Framework for Software Watermarking”. Deanship of Gradute Studies, King Fahad University of Petroleum and Minerals, Dhahran, Saudi Arabia. Dr. Mikhail Atallah, Eric D. Bryant, and Dr. Martin R. Stytz. “A survey of anti-tamper technologies” Gareth Cronin, “A Taxonomy of Methods for Software Piracy Prevention”,Department of Computer Science, University of Auckland, New Zealand gareth@cronin.co.nz Maña, A., Pimentel, E. (2001). “An Efficient Software Protection Scheme”,Retrieved Jun 2002 from http://polaris.lcc.uma.es/~amg/papers/IFIPSEC01-SoftProt.pdf. [18]

[19] [20]

[3]

[4] [5]

[21]

[22] [23]

[6]

Chang, H. & Atallah, M. “Protecting Software Code by Guards,”. Retrieved Apr 2002, also available at, http://www.starlab.com/sander/spdrm/papers.html. Anckaert, Bjorn De Sutter and Koen De Bosschere. “Software Piracy Prevention through Diversity”, Proceedings of the 4th ACM workshop on Digital rights management, p.63-71, 2004. T. Premkumar Devanbu and Stuart Stubblebine. “Software engineering for security a roadmap”. In Proceedings of the conference on The future of Software engineering, pages 227-239. ACM Press, 2000. Ori Dvir, Maurice Herlihy, Nir N. Shavit,“Virtual Leashing: InternetBased Software Piracy Protection” , computer Science Department, Tel Aviv University, Brown University oridvir@hotmail.com, herlihy@cs.brown.edu, shanir@cs.tau.ac.il Gene Tyler, IATAC Director, “SAMATE’s Contribution to Information Assurance”. Volume 9 Number 2 • Fall 2006. William Feng Zhu, “Concepts and Techniques in Software Watermarking and Obfuscation”. The Department of Computer Sciences, The University of Auckland New Zealand, August 2007 P. Cousot and Radhia Cousot.. “An Abstract Interpretation-Based Framework for Software Watermarking”, In Conference Record of the 31st ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Programming Languages, Venice, Italy, January 14-16, 2004. ACM Press, New York, U.S.A. pp. 173—185. Peter A. Jamieson, “Intellectual Property Protection via watermarking”. A Survey in Winter Semester 2000 Department of Electrical and Computer Engineering University of Toronto Toronto, Ontario M5S 3G4, Canada, April 17, 2000 Collberg, Clark Thomborson, and Gregg Townsend. “Dynamic graphbased software watermarking”. Technical report, Dept. of Computer Science, Univ. of Arizona, 2004. R. S. Clark Thomborson, Jasvir Nagra and C. He, “Tamper-proofing Software Water-marks” in Proc. Second Australasian Information Security Workshop (AISW2004), volume 32, pages 27–36, 2004. Yong He, “Tamperproofing a Software Watermark by Encoding Constants”, University of Auckland, Master thesis, 2002. C. Collberg and C. Thomborson, Software Watermarking: “Models and Dynamic Embeddings, in Proceedings of Symposium on Principles of Programming Languages”, POPL'99, pages 311-324, 1999. Lei Wang, “A Constant Encoding Algorithm Which Tamper-proofs the CT-Watermark”,September 2006. D. Aucsmith, Tamper Resistant Software: An Implementation, in Proceedings of the First International Workshop on Information Hiding, pages 317{333, London, UK, 1996, Springer-Verlag. J. Palsberg, S. Krishnaswamy, M. Kwon, D. Ma, and Q. Shao, “Experience with Software Watermarking”, in 16th Annual Computer Security Applications Conference, 2000. www.siia.net/piracy/whatis.asp 68k C. Collberg, “SandMArk a tool for software protection” [Collberg Page] C. Collberg, "Sandmark homepage," <http://www.cs.arizona.edu/sandmark/>, undated, (Accessed: 20 Jan 2002).

63

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

A novel trigon-based dual authentication protocol for enhancing security in grid environment
V. Ruckmani
Senior lecturer Department of Computer Applications Sri Ramakrishna Engineering College, India .

Dr G Sudha Sadasivam
Professor Department of Computer Science and Engineering PSG College of Technology, Coimbatore, India .

Abstract— In recent times, a necessity has been raised in order to distribute computing applications often across grids. These applications are dependent on the services like data transfer or data portal services as well as submission of jobs. Security is of utmost importance in grid computing applications as grid resources are heterogeneous, dynamic, and multi-domain. Authentication remains as the significant security challenge in grid environment. In traditional authentication protocol a single server stores the sensitive user credentials, like username and password. When such a server is compromised, a large number of user passwords, will be exposed. Our proposed approach uses a dual authentication protocol in order to improve the authentication service in grid environment. The protocol utilizes the fundamental concepts of trigon and based on the parameters of the trigon the user authentication will be performed. In the proposed protocol, the password is interpreted and alienated into more than one unit and these units are stored in two different servers, namely, Authentication Server and Backend Server. Only when the combined authentication scheme from both the servers authenticates the user, the privilege of accessing the requested resources is obtained by the user. The main advantage of utilizing the dual authentication protocol in grid computing is that an adversary user cannot attain the access privilege by compromising a single consolidated server because of the fact that the split password is stored in different servers. Keywords-Dual authentication; authentication protocol; trigon parameters; authentication code; grid computing; grid security.

The necessity for secure communication between entities on the Grid has motivated the development of the Grid Security Infrastructure (GSI). GSI provides integrity, protection, confidentiality and authentication for sensitive information transferred over the network in addition to the facilities to securely traverse the distinct organizations that are part of collaboration [6]. Secure invocation of Grid services brings out the need for a security model that reflects the security components that require to be recognized and defined based on the Grid security requirements [8]. Security requirements within the Grid environment are motivated by the requirement to support scalable, dynamic, distributed virtual organizations (VOs) [3]—collections of various and distributed individuals that are looking to share and utilize different resources in a synchronized fashion [7]. A general scenario within Grid computing involves the formation of dynamic “virtual organizations” (VOs) [3] including groups of individuals as well as associated resources and services combined by a general purpose but not located inside a single administrative domain [9]. The concept of Virtual Organization (VO) [3] has been launched to define the relationships between a set of grid components comprising computing resources, data, applications and users [10]. For a VO to operate successfully participants must have control over resource sharing policies via a secure infrastructure [11]. To avoid the illegal users from visiting the grid resources, it ought to be guaranteed that strong mutual authentication is necessary for users and server [5]. Users require to know if they are interacting with the “right” piece of software or human, and that their messages will not be altered or stolen as they traverse the virtual organization (if the users have such a requirement). Users will frequently need the ability to prevent others from reading data that they have stored in the virtual organization [12]. Grid systems and applications require standard security services comprising of authentication, access control, integrity, privacy [13]. Security plays a most important role in providing the confidentiality of the communication, the integrity of data and resources, and the privacy of the user information for large scale deployment of Grid [14]. The sensitive information and resources in information systems are shielded from illegitimate access by means of the access control that is widely employed as a security mechanism [15].

I. INTRODUCTION Enhanced network bandwidth, powerful computers, and the acceptance of the Internet have motivated the constant necessity for latest and enhanced ways to compute [1]. The growing complexity of computations, improved processing power of the personal computers and the constantly rising speed of the Internet have laid down the path for grid computing [2]. “Grid” computing has emerged as a significant new field, distinguished from conventional distributed computing by its concentration on large-scale resource sharing, innovative applications, and, in some cases, highperformance orientation [3]. Grid computing is concentrating on large-scale resource sharing and collaboration over enterprises and virtual organizations boundaries [4]. As the goal of grid computing is to only provide secure grid service resources to legal users, the security issue becomes a significant concern of grid computing [5]

64

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

At the base of any grid environment, there must be mechanisms to provide security including authentication, authorization, data encryption, and so on [33]. Authentication is the basis of security in grid [34]. Basically, authentication between two entities on remote grid nodes means that each party sets up a level of trust in the identity of the other party. In practical use, an authentication protocol sets up a secure communication channel between the authenticated parties, so that successive messages can be sent devoid of repeated authentication steps, even though it is possible to authenticate every message. The identity of an entity is typically some token or name that exclusively identifies the entity [16]. Grid technologies have adopted the use of X.509 identity certificates to support user authentication. An X.509 Certificate with its corresponding private key forms a unique credential, termed as grid credential, within the Grid. Grid credentials are utilized to authenticate both users and services [17]. In order to get the authentication from the server users and services are required to provide credentials. A credential is nothing but a piece of information that is utilized to prove the identity of a subject. Security frequently depends on the strength of the protections guarding a user’s credentials. The secure storage as well as the management of these credentials is the user’s responsibility. Usability, user mobility, and insufficient protection of workstations can cause major problems that often weaken the security of user credentials [18]. Passwords and certificates are some of the instances of credentials. Password-based authentication is still the most extensively used authentication mechanism, mainly due to the ease with which it can be understood by end users and implemented [19]. Password authentication is considered as one of the simplest and most convenient authentication mechanisms [22]. On the other hand, password authentication protocols are very subject to replay, password guessing and stolen-verifier attacks [20]. (1) Replay attack: A replay attack is an offensive action in which an adversary impersonates or deceives another legitimate participant via the reuse of information obtained in a protocol. (2) Guessing attack: A guessing attack involves an adversary simply (randomly or systematically) trying passwords, one at a time, in hope that the correct password is set up. Ensuring passwords selected from an adequately large space can resist exhaustive password searches. However, the majority of the users choose passwords from a small subset of the full password space. Such weak passwords with low entropy are easily guessed by means of the so-called dictionary attack. (3) Stolen-verifier attack: In the majority of the applications, the server stores verifiers of users’ passwords (e.g., hashed passwords) instead of the clear text of passwords. The stolen-verifier attack means that an adversary who steals the password-verifier from the server can use it directly to masquerade as a legitimate user in a user authentication execution. [21].

Clearly untraceable on-line password guessing attacks and off-line password guessing attacks are the most significant considerations in designing a password-based authentication scheme [22]. A great part of protocols for password-based authenticated key exchange system are intended for a single server environment where all the information about legitimate users is stored in one server. For that reason, a credential weakness is existed in this approach due to the fact that the user’s password is exposed if this server is ever compromised. A natural solution includes splitting the password between two or more servers which provides concrete security proofs for authentication protocol [23]. The dual-server model that includes two servers at the server side, one of which is a public server exposing itself to users and the other of which is a back-end server staying behind the scene; users contact only the public server, but the two servers work jointly to authenticate users [24]. This paper proposes a novel dual authentication protocol which utilizes dual servers for authentication to enhance the grid security. The novelty of the protocol is the usage of the fundamental concepts and basic elements of the trigon to authenticate. With these trigon parameters, the user credential is interpreted and then stored in two servers which provide solid security evidences for authentication protocol. The dual authentication protocol gives authentication to the grid user if and only if both the servers are mutually involved in the authentication mechanism. It is not possible to obtain the password by hacking a single server. Moreover, our protocol offers effective security against the attacks like replay attack, guessing attack and stolen-verifier attack as the user authentication is a combined mechanism of two servers. Also, it provides the security to the valid users as well as securing the user credentials, as an additional feature. Succinctly, the protocol provides secured environment while the grid user entered into the VO and the services access from the grid. The remaining of the paper is organized as follows: Section II deals with some of the existing research works which have been done so far and Section III is constituted by the proposed dual authentication protocol, user registration process and design of the authentication code with required illustrations and mathematical formulations. Section IV discusses about the implementation results and Section V concludes the paper. II. RELATED WORKS Wei Jiea et al. [25] have proposed a scalable GIS architecture for information management in a large scale Grid Virtual Organization (VO). The architecture was comprised of the VO layer, site layer and resource layer: at the resource layer, information agents and pluggable information sensors were deployed on every resource monitored. The information agent and sensor approach provided a flexible framework that facilitated particular information to be captured; at the site layer, a site information service component with caching capability aggregates and maintained up-to-date information of all the resources monitored inside an administrative domain; at the VO layer, a peer-to-peer approach was utilized to construct a virtual network of site information services for information discovery and query in a large scale Grid VO. In addition to that, they proposed a security framework for the GIS, which provided security policies for authentication and

65

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

authorization control of the GIS at both the site and the VO layers. Their GIS has been implemented based on the Globus Toolkit 4 as Web services compliant to Web Services Resource Framework (WSRF) specifications. The experimental results showed that the GIS presented satisfactory scalability in maintaining information for large scale Grids. Haibo Chena et al. [26] have presented the work of Daonity which was their effort to strengthening grid security. They identified that a security service which they named behavior conformity be desirable for grid computing. Behavior conformity for grid computing was an assurance that ad hoc related principals (users, platforms or instruments) forming a grid VO should each act in conformity with the rules for the VO constitution. They applied trusted computing technologies in order to attain two levels of virtualization: resource virtualization and platform virtualization. The former was about behavior conformity in a grid VO and the latter, that in an operating system. With those two levels of virtualization working together it was possible to construct a grid of truly unbounded scale by VO together with servers from commercial organizations. Yuri Demchenko [27] has provided insight into one of the key concepts of Open Grid Services Architecture (OGSA) and Virtual Organizations (VO). They have analyzed problems related to Identity management in VOs and their possible solution on the basis of utilizing WS-Federation and related WS-Security standards. The paper provided basic information about OGSA, OGSA Security Architecture and analyses VO security services. A detailed description was provided for WSFederation Federated Identity Model and operation of basic services for instance Security Token Service or Identity Provider, Attribute and Pseudonym services for typical usage scenarios. G. Laccetti and G. Schmid [28] have introduced a sort of unified approach, an overall architectural framework for access control to grid resources, and one that adhered as much as possible to security principles. Grid security implementations were viewed in the light of the model, their main drawbacks were described, and they showed how their proposal was able to prevent them. They believed that a main strategy could be to adopt both PKI (Public Key Infrastructure) and PMI (Privilege Management Infrastructure) infrastructures at the grid layer, ensured that a sufficient transfer of authentication and authorization made between the Virtual Organization and Resource Provider layers. That can be attained by expanding those features at the OS layer as system applications and services. Xukai Zoua et al. [29] have proposed an elegant DualLevel Key Management (DLKM) mechanism by means of an innovative concept/construction of Access Control Polynomial (ACP) and one-way functions. The first level provided a flexible and secure group communication technology whereas the second level offered hierarchical access control. Complexity analysis and Simulation demonstrated the efficiency and effectiveness of the proposed DLKM in the computational grid as well as the data grid. An example was demonstrated.

Li Hongweia et al. [30] have proposed an identity-based authentication protocol for grid on the basis of the identitybased architecture for grid (IBAG) and corresponding encryption and signature schemes. Commonly, grid authentication frameworks were attained by means of applying the standard SSL authentication protocol (SAP). The authentication process was very complex, and therefore, the grid user was in a heavily loaded point both in computation and in communication. Being certificate-free, the authentication protocol aligned well with the demands of grid computing. By means of simulation testing, it was seen that the authentication protocol was more lightweight and effective than SAP, in particular the more lightweight user side. That contributed to the better grid scalability. Yan Zhenga et al [31] have aimed at designing a secure and effective method for grid authentication by means of employing identity-based cryptography (IBC). Nevertheless, the most extensively accepted and applied grid authentication was on the basis of the public key infrastructure (PKI) and X.509 certificates, which made the system, have lesser processing efficiency and poor anti-attack capability. An identity-based signature (IBS) scheme was first proposed for the generation of private key during grid authentication. On the basis of the proposed IBS and the IBE schemes, the structure of a grid authentication model was given, followed by a grid authentication protocol explained in detail. According to the theoretical analysis of the model and the protocol, it could be discussed that the system has enhanced both the security and efficiency of the grid authentication when compared with the traditional PKI-based and some IBCbased models. Hai-yan Wanga. C and Ru-chuan Wanga [32] have proposed a grid authentication mechanism, which was on the basis of combined public key (CPK) employing elliptic curve cryptography (ECC). Property analysis of the mechanism in comparison to the globus security infrastructure (GSI) authentications, showed that CPK-based grid authentication, might be applied as an optimized approach towards efficient and effective grid authentication. Our proposed work on a novel dual authentication protocol utilizes dual servers for authentication to enhance the grid security. The novelty of the protocol is the usage of the fundamental concepts and basic elements of the trigon to authenticate. III. PROPOSED USER REGISTRATION PROCESS FOR TRIGONBASED DUAL AUTHENTICATION

To achieve the dual authentication, it necessitates the user to register with the Authentication server. The procedures followed in the Authentication server and the Backend server during registration of the user is as follows. The users have to register with the Authentication server, so that it can hold a part of the interpreted password with itself and another part in the Backend server. The block diagram illustrating the registration process of the users is depicted in the Figure 1.

66

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

With the parameters a , a ' and a ' ' as the sides of trigon and Pi be the angle between the sides a and a ' the generated trigon will be assumed as in the Figure 2.

Figure 2. A sample trigon generated using the parameters a , a ' , a ' ' and

Pi
Figure 1. Flow chart explaining the registration process of user

As illustrated in Figure 1, the users who require services from the VO have to register initially with the Authentication server using their username and password. The Authentication server calculates the Pi as given in (1). Along with this, the Authentication server also generates two large prime numbers, namely, a and a ' , which are considered as the two sides of a trigon. It is difficult to hack the values of a and a ' as they are large prime numbers (as per RSA Factoring Challenge). Here, Pi is taken as the angle between the two a and a ' . Now, the Authentication server can easily determine the opposite side of the angle Pi , termed as a ' ' . With these trigon parameters, the user determines α , Vaa ' and Paa ' as follows

Vaa ' = a − a ' Paa ' = a * a '

(1) (2) (3)

After the calculation of α , Vaa ' and Paa ' , the authentication server stores the α value and its corresponding username in a database and forwards Vaa ' and Paa ' to the Backend server along with the username. The Backend server maintains the Vaa ' and Paa ' for the corresponding username in a database. Hence, the password is interpreted and alienated into two units and stored in two separate servers, thereby achieving the concept of dual authentication. The process is repeated for all the users who wish to register in the server, so that both the servers can maintain all the users’ account. When any of the users try to access the VO, they will be validated by these servers using the account information and then allowed to access the VO by providing TVO . If the user is an adversary and if it tries to use wrong password or username, the server can validate effectively, asserts the user as invalid and sends a warning to the adversary. The dual authentication code proposed here is designed based on the fundamentals of trigon and the design steps are discussed in the section below. A. The proposed Trigon-based dual authentication protocol Taking the security as the main constraint in grid computing environment, we are proposing a dual authentication protocol, which will authenticate the user by a combined mechanism of two servers, namely, authentication server and backend server and then allows the user to access the VO for services. Here, the public server is mentioned as the authentication server as it performs the major authentication mechanism. The authentication procedure we have developed is on the basis of the fundamental concepts of a trigon. The Figure 3 depicts the activity diagram of the proposed dual authentication protocol.

α = 2 Paa ' − a ' ' 2
where,

a , a ' and a ' ' are the three sides of the trigon,

α is a strengthening parameter used as the index to represent user credentials,
Vaa ' and Paa ' are the Variance and the product of the sides a and a ' respectively.

67

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Figure 3. Activity Diagram of the proposed trigon-based dual authentication protocol

As described in the Figure 3, initially, the user who wants the services of VO has to login to the Authentication server using the username and password. Here, u i and pwi refers

th to username and password of i user. The Authentication server calculates the Password index ( Pi ) from the password as

PAI is the ASCII-interpreted value of the given password pwi , n is the total number of digits in PAI and PAI ( j ) represents the first j digits of PAI . The PAI can be
In (4), calculated by the following steps. • • Change the

pwi into its corresponding ASCII value.

 PAI  10 n − 2 ; if PAI ( j ) ≥ 180  Pi =   PAI ; else  10 n − 3 

(4) •

Calculate the three-fourth of total digits of the ASCII value modulo 180, which results the first three digits of PAI . Append the remaining one-fourth of the ASCII digits to PAI .

68

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Then, from Pi , the Authentication Server determines the (i ) Authentication index ( AI ) for u i as

validates the user and allows the user for resource sharing in the grid environment. B. Design of the authentication code The authentication code provided in (7) takes the eventual decision of whether the user who logins is valid or adversary. The steps by which the authentication code is developed are described elaborately as follows. The semi-perimeter determined as

P ( AI i ) = i 2

(5)

Then, the Authentication Server searches for the username index α i for the corresponding u i which has already been stored in the server database during the process of the registration. Subsequently, α i is sent to the backend server along with u i . When the Backend server receives the index

S P of the trigon depicted above is

SP =

a + a '+ a ' ' 2

(8)

α i and

the username from the Authentication server, it

But it is known that,

Vaa ' and Paa ' the Variance and the product of the sides a and a ' respectively, which have been saved in the
searches for backend server database during the process of registration. From these values, the Backend server calculates an Authentication Token AT and sends it to the Authentication
i

 ( S − a) − (S P − a ' )  Sin 2 ( AI ) =  P  a . a'  

1/ 2
(9)

Square of the RHS value of (9) takes the form,
 a + a '+ a' '    a + a '+ a' '    − a    − a '  2 2 ( S P − a ) − ( S P − a ' )       (10) = a . a' a . a'

server to authenticate the u i . The

ATi can be calculated as

αi + V 2 '
ATi =
In (6),

aa i
i

2 Paa '

(6) Applying (9), (10) can be reorganized as follows

Vaa ' and Paa ' are pre-calculated values computed during individual user registration. After the retrieval of AT i
from the Backend server, the Authentication server authenticates the user based on the token from the Backend server and the index calculated at the Authentication server. The authentication code (or) condition which authenticates the u i is given by (proved in section III.B)

Sin 2 ( AI ) =
As given in (3),

2 a a '  a 2 + a ' 2 −a ' ' 2   −   4 a a'  4 a a' ' 

(11)

2 a a ' = a ' ' 2 +α
Using (10), (9) can be written as follows,

(12)

 1 − ATi ( Sin( AI i ) ) =   2 

   

1/ 2
(7)

Sin 2 ( AI ) =

1  a 2 + a ' 2 −2 a a '+α   −  2  4 a a'  

(13)

The authentication process is performed by the authentication condition given in (7). When the condition is satisfied, the user is decided to be valid and the Server sends a token called Token for VO access TVO to the user. Using the Token TVO the user can contact the VO and accomplish its tasks and access the resources in the VO. If the condition is not satisfied, then a word of warning is given to the user. As a consequence, the user has no TVO to contact the VO and hence no resource sharing. Thus, the proposed dual authentication protocol based on two servers effectively

1  (a − a' ) 2 + α   Sin ( AI ) = −   2  4 a a'  
2
Substituting (1) and (2), in (14) gives

(14)

1  Vaa ' 2 + α   Sin ( AI ) = −  2  4 Paa '   
2

(15)

69

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

The re-arranged format of the above equation is given by

is interpreted and then alienated into two modules and stored in the two servers. V. RESULTS AND DISCUSSIONS

Sin 2 ( AI ) =
Substituting get

1  Vaa ' 2 + α  1 −   2 2 Paa '  

(16)

AT which is given in (6) into (16), we can
1/ 2

 1 − AT  Sin 2 ( AI ) =    2 

(17)

From the above steps, the authentication code utilized for the proposed dual authentication protocol is designed and can also serve as a proof for the effectiveness of the protocol. The protocol devised is based on the trigon parameters and effectively provides an enhanced security, because both the authentication server and the backend server have been involved in the authentication mechanism. So, compromising a single server and enjoying the VO services is impossible by any means. IV. SECURITY ANALYSIS Replay attack: Usually replay attack is called as ‘man in the middle’ attack. Adversary stays in between the user and the server and hacks the user credentials when the user contacts server. Normally, to overcome this, the user has to change the credential randomly. But it is less probable to do that. Our protocol is robust when the replay attack happens in between the two servers as the credentials are interpreted and alienated into two parts. Guessing attack: Guessing attack is nothing but the adversaries just contacts the servers by randomly guessed credentials. The effective possibility to overcome this attack is to choose the password by maximum possible characters, so that the probability of guessing the correct password can be reduced. As the proposed uses random generation of prime numbers for the calculation of the sides of the trigon, it is more difficult to guess the password. Stolen-verifier attack: Instead of storing the original password, the server is normally storing the verifier of the password. If the password steals the verifier from the server, then it will masquerade as the legitimate user. But this not happens in any two server protocol, as the password is alienated into two modules. Hence, we can justify that our protocol is also more robust against the attack, as the password

The proposed dual authentication protocol has been implemented in the platform of JAVA (version 1.6). The protocol is tested with five valid and five invalid users. Each of the five valid users has their own username and password. Initially, they have created their user account by registering with their username and password, making them valid in the VO. The usernames, passwords and the corresponding trigon parameters of the five valid users are given in the Table II. The trigon parameters have been determined during the registration process as stated earlier and they have been stored in the database maintained at the servers.
TABLE I. USERNAMES, PASSWORDS AND THE TRIGON PARAMETERS BASED ON
THE USER PASSWORDS PROVIDED AT THE TIME OF REGISTRATION

Sl. No 1

User name user1

Pass word admin

α

Vaa '

Paa '
1.2020119316 9E11 1.9450880549 E10 6.6052326655 1E11 2.1752646107 E10 3.377365493E 9

665840.0 3.8067649159674 07E11 2.2186644627851 108052.0 135E10 1.2148984151865 -300790.0 493E12 4.2139671410780 29146.0 15E10 452092.0 1.9786419670597 998E11

2 3 4 5

user2 user3 user4 user5

ascii test5 test8 test10

The α values for the five valid users mentioned in the Table I have been stored in the authentication server database and Vaa ' and Paa ' have been stored in the database of Backend server for the corresponding usernames. Instead of keeping the actual passwords, the servers maintain the interpreted passwords derived from the trigon parameters. When the servers authenticate any user, the servers determine some authentication elements based on the values which have been stored in the database and the login credential provided by the user. Using such authentication elements, the servers generate an authentication code and validate the user. The Table II shows the authentication elements generated by the servers when valid and invalid users contact the authentication server for authentication. The parameters contributed in the authentication and the authentication outcome for five valid and invalid users are given in the Table II.

70

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
TABLE II. THE AUTHENTICATION PARAMETERS DERIVED FROM THE TRIGON PARAMETERS, THE AUTHENTICATION CODE STATUS AND THE OUTCOME OBTAINED FROM AUTHENTICATION OF THE USER. Sl. No 1 2 3 4 5 6 7 8 9 10 User name user1 user2 user3 user4 user5 user1 user2 user3 user4 user5 Pass word admin ascii test5 test8 test10 admins asci test4 good user10

Pi
105.11 150.5105 171.1653 171.1656 164.948 151.10115 59.105 171.1652 91.1 114.4948

AT
0.26067301143 654953 0.87044592265 49521 0.98813554752 03079 0.98813635167 23201 0.96569053179 7593 0.26067301143 654953 0.87044592265 49521 0.98813554752 03079 0.98813635167 23201 0.96569053179 7593

AI
52.555 75.2552 5 85.5826 5 85.5828 82.474 75.5505 75 29.5525 85.5826 45.55 57.2474

Sin 2 (AI )
0.3696634942 817 0.0647770386 725 0.0059322262 398 0.0059318241 638 0.0171547341 012 0.0622628865 386 0.2432668150 931 0.0059323602 682 0.4904012788 002 0.2926946722 584

1 − AT 2
0.369663494281 7 0.064777038672 5 0.005932226239 8 0.005931824163 8 0.017154734101 2 0.369663494281 7 0.064777038672 5 0.005932226239 8 0.005931824163 8 0.017154734101 2

Authentication code Authentication balanced? outcome Yes Yes Yes Yes Yes No No No No No Valid Valid Valid Valid Valid Invalid Invalid Invalid Invalid Invalid

The AI for each user as illustrated in the Table II, has been calculated by the authentication server and the AT for each user has been calculated by the Backend server. Based on these values, the authentication server generated the authentication code and checked whether it has been satisfied or not. When the authentication code has been satisfied by any of the user, the servers asserted that the user is valid and permits users to utilize the services offered by the VO. The status of the authentication code and the outcome of the server for valid and invalid cases are clearly tabulated in the Table II. This shows the effective performance of the protocol in enhancing the security of the grid environment by identifying valid and adversary users. Each user was provided a separate TVO if and only the user credential supplied by the concerned user satisfied the authentication code. The user credential that did not satisfy the authentication code was declared as invalid credential and the concerned user was asserted as an adversary. This is because that the authentication code will be satisfied if and only if the user credentials submitted for authentication are properly registered. Hence, the protocol effectively pinpointed the adversary and denied the services for that adversary user. VI. CONCLUSION The authentication protocol, proposed here, enhanced the grid security as the authentication mechanism utilized two servers for authentication. As the servers kept the interpreted and distinct form of user credentials, there is very less chance to reveal the user credentials to the adversary. Moreover, the protocol utilized the fundamental properties of the trigon and the trigon parameters, made the grid more secure as the alienated passwords had been derived from these trigon parameters. This simple trigon concept utilization in the authentication protocol introduced a novel and revolutionary

idea in the authentication mechanism as well as in grid environment. The implementation of our dual authentication protocol showed its effective performance in pinpointing the adversaries and paving the way to valid users for access with the VO for resource sharing. When the protocol identified any of the adversaries while authentication, it strictly prohibited those invalids from accessing with VO, which satisfies the essential pre-requisite for any authentication protocol. The development procedures of the authentication code discussed in our paper is a proof which shows the effectiveness of the protocol. So the utilization of this protocol will make the grid environment more secure. VII. ACKNOWLEDGEMENT We authors would like to thank Mr. K. V. Chidambaram, Director, Data Infrastructure & Cloud Computing Group, Yahoo Software Development India Pvt Ltd., and Dr. R. Rudramoorthy, Principal, PSG College of Technology, Coimbatore for their support in carrying out the research work. We also thank the management of Sri Ramakrishna College of engineering for their support. REFERENCES
[1] V.Vijayakumar and R.S.D.Wahida Banu, “Security for Resource Selection in Grid Computing Based on Trust and Reputation Responsiveness”, IJCSNS International Journal of Computer Science and Network Security, Vol.8, no.11, November 2008. Wenliang Du, Jing Jia, Manish Mangal, and Mummoorthy Murugesan, “Uncheatable Grid Computing”, in Proceedings of the 24th International Conference on Distributed Computing Systems (ICDCS'04), pp. 4 - 11, 2004. Foster. I., Kesselman. C. and Tuecke. S, “The Anatomy of the Grid: Enabling Scalable Virtual Organizations”, International Journal of High Performance Computing Applications”, vol. 15, no.3, pp. 200-222, 2001. Li Wang, Wenli Wu, YingJie Li and XueLi Yu, “Content-aware Trust Statement for semantic Grid”, in proceedings of the Second International

[2]

[3]

[4]

71

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
Conference on Semantics, Knowledge and Grid, pp.95 - 95, November 2006. Rongxing Lu, Zhenfu Cao, Zhenchuan Chai, and Xiaohui Liang, "A Simple User Authentication Scheme for Grid Computing, International Journal of Network Security, vol.7, no.2, Pp.202–206, September 2008. Ionut Constandache, Daniel Olmedilla, Frank Siebenlist and Wolfgang Nejdl, "Policy-driven Negotiation for Authorization in the Semantic Grid", Technical report, L3S Research Center, October 2005. Von Welch, Frank Siebenlist, Ian Foster, John Bresnahan, Karl Czajkowski, Jarek Gawor, Carl Kesselman, Sam Meder, Laura Pearlman and Steven Tuecke, “Security for Grid Services”, in proceedings of the 12th IEEE International Symposium on High Performance Distributed Computing, pp.48- 57, June 2003. N. Nagaratnam, P. Janson, J. Dayka, A. Nadalin, F. Siebenlist, V. Welch, I. Foster, S. Tuecke, The security architecture for open grid services, OGSA-SEC-WG document,http://www.cs.virginia.edu/~humphrey/ogsa-sec-wg/OGSASecArch-v1-07192002.pdf, July 17 2008. Foster, I., Kesselman, C., Tsudik, G. and Tuecke, S, “A Security Architecture for Computational Grids” in proceedings of the ACM Conference on Computers and Security, pp. 83-91, 1998. Thawan Kooburat and Veera Muangsin, "Centralized Grid Hosting System for Multiple Virtual Organizations", 10th Annual National Symposium on Computational Science and Engineering (ANSCSE10), Chiangmai, March 2006. David W. O Callaghan, Brian A. Coghlan, "Bridging Secure WebCom and European DataGrid Security for Multiple VOs over Multiple Grids", in proceedings of the Third International Symposium on Parallel and Distributed Computing/Third International Workshop on Algorithms, Models and Tools for Parallel Computing on Heterogeneous Networks (ISPDC/HeteroPar'04),ispdc, pp.225-231, 2004. Marty Humphrey, Mary R. Thompson and Keith R. Jackson, "Security for Grids", in Proceedings of the IEEE ,vol. 93, no. 3, pp.644 - 652, March 2005. Alexander Kemalov, “A Security Policy in GRID Architecture", International Conference on Computer Systems and Technologies, 2005. Shashi Bhanwar, and Seema Bawa, “Securing a Grid”, in Proceedings of World Academy of Science, Engineering and Technology, vol.32, August 2008. M.Nithya and R.S.D.Wahida Banu, "Towards Novel And Efficient Security Architecture For Role Based Access Control In Grid Computing", IJCSNS International Journal of Computer Science and Network Security, vol. 9, no.3, March 2009. Mary R. Thompson, Doug Olson, Robert Cowles, Shawn Mullen and Mike Helm," CA-based Trust Model for Grid Authentication and Identity Delegation", Global Grid Forum CA Operations WG Community Practices Document, Oct 2002. Stephen Langella, Scott Oster, Shannon Hastings, Frank Siebenlist, Joshua Phillips,David Ervin,Justin Permar, Tahsin Kurc and Joel Saltz, "The Cancer Biomedical Informatics Grid (caBIG) Security Infrastructure", in Proceedings of 2007 AMIA Annual Symposium, Chicago, Illinois, 2007. Dr. Dennis Kafura and Dr. Markus Lorch , “A security architecture to enable user collaboration in computational grids”, CISC Research Report 04-05. J. Crampton, H.W.Lim, K.G.Paterson and G.Price, "A Certificate-Free Grid Security Infrastructure Supporting Password-Based User Authentication" In Proceedings of the 6th Annual PKI R&D Workshop 2007, pp. 103-118, Gaithersburg, Maryland, USA, 2007. Lin, C.L., and T. Hwang, “A password authentication scheme with secure password updating”, Computer & Security, vol.22, no.1, pp.68– 72, 2003. Eun-Jun Yoon, Eun-Kyung Ryu and Kee-Young Yoo, “ Attacks and Solutions of Yang et al.’s Protected Password Changing Scheme”, Informatica, vol.16 , no. 2, pp. 285-294, April 2005. Her-Tyan Yeh, Hung-Min Sun and Tzonelih Hwang, “Efficient ThreeParty Authentication and Key Agreement Protocols Resistant to Password Guessing Attacks”, Journal of Information Science and Engineering, vol.19, no.6, pp. 1059-1070, 2003. [23] Michael Szydlo and Burton Kaliski , “Proofs for Two-Server Password Authentication” , In proceedings of the Cryptographer’s Track at the RSA(CT-RSA 2005) Conference, pp. 227-244, San Francisco, CA, USA, 2005. [24] Yanjiang Yang, Robert H. Deng and Feng Bao, "A Practical PasswordBased Two-Server Authentication and Key Exchange System", IEEE Transactions on Dependable and Secure Computing, vol. 3, no. 2, AprilJune 2006. [25] Wei Jiea,Wentong Caib, Lizhe Wangc and Rob Proctera, "A secure information service for monitoring large scale grids",Parallel Computing, Vol.33, no. 7-8, pp. 572-591, August 2007. [26] Haibo Chena, Jieyun Chenb, Wenbo Maoc and Fei Yand, "Daonity – Grid security from two levels of virtualization",Information Security Technical Report, Vol.12, no.3, pp. 123-138, 2007. [27] Yuri Demchenko, "Virtual organisations in computer grids and identity management", Information Security Technical Report, vol.9, no. 1, pp.59-76, January-March 2004. [28] G. Laccetti and G. Schmid, "A framework model for grid security”, Future Generation Computer Systems, vol. 23, no. 5, pp.702-713,June 2007. [29] Xukai Zoua, Yuan-Shun Dai and Xiang Rana, "Dual-Level Key Management for secure grid communication in dynamic and hierarchical groups", Future Generation Computer Systems,Vol. 23, no. 6,pp. 776786,July 2007. [30] Li Hongweia, Sun Shixina and Yang Haomiaoa, "Identity-based authentication protocol for grid", Journal of Systems Engineering and Electronics, Vol. 19, no. 4, pp.860-865, August 2008. [31] Yan Zhenga, Hai-yan Wanga and Ru-chuan Wang, "Grid authentication from identity-based cryptography without random oracles", The Journal of China Universities of Posts and Telecommunications, Vol.15, no. 4, pp.55-59, December 2008. [32] Hai-yan Wanga. C and Ru-chuan Wanga,"CPK-based grid authentication: a step forward", The Journal of China Universities of Posts and Telecommunications, Vol.14, no. 1, pp.26-31, March 2007. [33] Yuanbo Guo, Jianfeng Ma and Yadi Wang, "An Intrusion-Resilient Authorization and Authentication Framework for Grid Computing Infrastructure",in proceedings of the Workshop on Grid Computing Security and Resource Management, Springer Berlin / Heidelberg, Vol.3516, pp.229-236, 2005. [34] Shushan Zhao Aggarwal. A and Kent. R.D, "PKI-Based Authentication Mechanisms in Grid Systems", in proceedings of the International Conference on Networking, Architecture and Storage, pp.83-90, Guilin, July 2007.

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13] [14]

[15]

[16]

[17]

[18]

[19]

V Ruckmani received B. Sc, MCA and M. Phil degrees from the department of computer science, Bharathiar University, India in 1994, 1997 and 2003 respectively. She is currently pursuing the Ph. D degree, working closely with Prof. G. Sudha Sadasivam. From 1997 to 2000 she worked at PSG College of Arts and Science in the department of Computer Science. Since December 2000 she is working as a senior lecturer in Department of Computer Applications in Sri Ramakrishna Engineering College, India. She works in the field of Grid Computing specializing in the area of security. You may contact her at ruckmaniv@yahoo.com
Dr G Sudha Sadasivam is working as a Professor in Department of Computer Science and Engineering in PSG College of Technology, India. Her areas of interest include, Distributed Systems, Distributed Object Technology, Grid and Cloud Computing. She has published 20 papers in referred journals and 32 papers in National and International Conferences. She has coordinated two AICTE – RPS projects in Distributed and Grid Computing areas. She is also the coordinator for PSG-Yahoo Research on Grid and Cloud computing. You may contact her at sudhasadhasivam@yahoo.com

[20]

[21]

[22]

72

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

Design and Analysis of a Spurious Switching Suppression Technique Equipped Low Power Multiplier with Hybrid Encoding Scheme
S.Saravanan Department of ECE, M.Madheswaran Department of ECE

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

K.S.R.College of Technology
Tiruchengode-637215, India. .

Muthayammal Engineering College
Rasipuram-647408, India

Abstract— Multiplication is an arithmetic operation that is mostly used in Digital Signal Processing (DSP) and communication applications. Efficient implementation of the multipliers is required in many applications. The design and analysis of Spurious Switching Suppression Technique (SSST) equipped low power multiplier with hybrid encoding is presented in this paper. The proposed encoding technique reduces the number of switching activity and dynamic power consumption by analyzing the bit patterns in the input data. In this proposed encoding scheme, the operation is executed depends upon the number of 1’s and its position in the multiplier data. The architecture of the proposed multiplier is designed using a low power full adder which consumes less power than the other adder architectures. The switching activity of the proposed multiplier has been reduced by 86% and 46% compared with conventional and Booth multiplier respectively. It is observed from the device level simulation using TANNER 12.6 EDA that the power consumption of the proposed multiplier has been reduced by 87% and 26% compared with conventional and Booth multiplier Keywords-component; Low power Multiplier, Hybrid encoding. VLSI Design, Booth

I.

INTRODUCTION

Multiplication is one of the most critical operations in many computational systems. The growing popularity of portable and multimedia devices such as video phones, note books in which multipliers play the important role. This has motivated the research in the recent years to design low power VLSI circuits. Application specific integrated circuits rely on efficient implementation of various arithmetic circuits for executing the specified algorithms. It is well known that if the density of transistor increases, the complexity of arithmetic circuits also increases and consumes more energy. This has further motivated the new concepts of designing low power VLSI circuits. It is also clear that the reduction in power consumption and enhancement in the circuit design are expected to pose challenges in implementing wireless multimedia and digital image processing system, in which multiplication and multiplication-accumulation are the key computations. In the recent past, the researchers proposed various design methodologies on dynamic power reduction using minimizing the switching activities. Choi et al [1]

proposed Partially Guarded Computation (PGC) which divides the arithmetic units into two parts and turns off the unused part to minimize the power consumption. The reported results show that the PGC can reduce power consumption by 10% to 44% in an array multiplier with 30% to 36% area overhead in speech related applications. A 32-bit 2’s complement adder equipping a dynamic-range determination (DRD) unit and a sign-extension unit was reported by Chen et al [2]. This design tends to reduce the power dissipation of conventional adders for multimedia applications. Later, Chen et al [3] presented a multiplier using the DRD unit to select the input operand with a smaller effective dynamic range that yield the Booth codes which reduces 30% power dissipation than conventional method. Benini et al [4] reported the technique for glitching power minimization by replacing few existing gates with functionally equivalent ones that can be “frozen” by asserting a control signal. This method operates in the layout level environment which is tightly restricted and hence it reduces 6.3% of total power dissipation. The doubleswitch circuit-block scheme was proposed by Henzler et al [5] is capable of reducing power dissipation by shortening the settling time during down time. Huang and Ercegovac [6] presented the arithmetic details about the signal gating schemes and showed 10% to 45% power reduction for adders. The combination of the signal flow optimization (SFO), leftto-right leapfrog (LRLF) structure, and upper/lower split structure was incorporated in the design to optimize the array multipliers by Huang and Ercegovac [7] and it is reported that the new approach can save about 20% power dissipation. Wen et al [8] reported that for the known output, some columns in the multiplier can be turned off and reduce 10% power consumption for random inputs. Chen and Chu [9] later, reported that the spurious power suppression technique has been applied on both compression tree and modified Booth decoder to enlarge the power reduction. Ko et al [10] and Song and Micheli [11] investigated full adder as the core element of complex arithmetic units like adder, multiplier, division, exponentiation and MAC units. Several combinations of static CMOS logic styles have been used to implement low-power one bit adder cells. In general, the logic styles were broadly divided into two major categories such as the complementary CMOS and the pass-transistor logic

73

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

circuits. The complementary CMOS logic style uses the power lines as input where the pass transistor logic uses separate input signals.

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

The complementary CMOS full adder is based on the regular CMOS structure with pMOS pull-up and nMOS pulldown transistors [12]. The authors reported that the series transistors in the output stage form a weak driver and additional buffers at the last stage is required for providing the necessary driving power to the cascaded cells. Chandrakasan and Brodersen [13] reported that the Complementary Pass transistor Logic (CPL) full adder with swing restoration structure utilizes 32 transistors. A Transmission Function Full Adder (TFA) based on the transmission function theory was presented by Zhuang and Hu [14]. Later, Weste and Eshraghian [15] presented a Transmission Gate Adder (TGA) using CMOS transmission gates circuit which is a special kind of pass-transistor logic circuit. The transmission gate logic requires double the number of transistors of the standard passtransistor logic or more to implement the same circuit. Hence the research has been focused by various researchers on smaller transistor count adder circuits, most of which exploit the non full swing pass transistors with swing restored transmission gate technique. This is exemplified by the stateof-the-art design of 14T and 10T which was reported by Vesterbacka [16] and Bui et al [17]. Chang et al [18] proposed a hybrid style full adder circuit in which the sum and carry generation circuits are designed using hybrid logic styles.

Full adders are used in a tree structure for high performance arithmetic circuits and a cascaded simulation structure is introduced to evaluate the full adders in real time applications. Keeping the above facts, it is proposed to improve the performance of the multiplier unit using spurious switching suppression technique and hybrid encoded scheme. In this research paper a novel design method has been proposed to reduce the number of switching activities and power consumption of multiplier. II. PROPOSED HYBRID ENCODED LOW POWER MULTIPLIER

Hybrid encoding rule In general, multiplication process consists of two parts as multiplicand and multiplier. According to the conventional shift and add multiplication, the number of partial products (PP) are equal to the number of bits in the multiplier. The number of partial products can be reduced by half using Booth recoding. In the proposed encoding technique, the partial products can still be reduced which in turn reduces the switching activity and power consumption. The operation can be defined according to the number of 1’s and its position in the multiplier. The proposed hybrid encoding rule is demonstrated and provided in Fig 1. The operation of proposed hybrid encoding rule is stated in Table 1 with details of operation.

A.

Multiplier Data

Whether Number of 1’s ≤ 3

No

Split Multiplier into two parts

Yes

Yes

Whether Number of 1’s ≤ 3 No Booth Multiplication

Proposed Multiplication

Whether Number of 1’s=2 Yes Check for Position of 1 Category C Category D

No

Whether Number of 1’s=1 Yes Check for Position of 1

No

Whether Number of 1’s=3 Yes Check for Position of 1 Category E Category F

Category A

Category B

Figure 1. Flow chart of the proposed multiplier

74

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

If the number of 1’s in the multiplier is less than or equal to 3, the control goes to proposed multiplication technique, otherwise the control split the multiplier in to two parts. Again the number of 1’s in the part of the multiplier is verified. If the number of 1’s is more than three, the control goes to Booth multiplication. Otherwise the control goes to proposed multiplication technique. If the number of 1’s in the multiplier is one and depends upon its position, the control goes to execute the operation in category A or B. If the number of 1’s in the multiplier is two and depends upon its position, the control goes to execute the operation in category C or D. Otherwise the number of 1’s in the multiplier is three and depends upon its position, the control goes to execute the operation in category E or F. The proposed multiplication technique is explained with the example shown in Fig. 2.
TABLE I. Number of 1’s in the Multiplier 1 1 2 2 3 3 HYBRID ENCODING SCHEME

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Multiplicand

Multiplier

Hybrid Encoder with Spurious Switching Suppression Technique

Detection Unit Control Signal

+ +

+

+ + +

+

Position of the 1 1st bit ith bit 1st and ith bit i th and i+j th bit th i =1st, jth and k th bit ith , jth and kth bit

Product
Category A B C D E F Operation Figure 3. SSST equipped multiplier Add 0 to multiplicand (M) Shift M left by i-1 and add 0 Shift M left by i-1 and add M Shift M left by j , add M and shift the result left by i-1 Shift M by k-j , add M and shift the result left by j-i, add M and shift the result left by i-1 Shift M by k-j , add M and shift the result left by j-i, add M and shift the result left by i

According to category E, the proposed encoding rule needs one partial product P1 with two additions. The remaining partial products P2 to P5 are zero, so the addition operation in this area can be neglected, which reduces the switching activity and power consumption. This spurious switching activity can be reduced by freezing the adders which perform this unwanted addition. In Fig 3 dashed adders operation can be blocked and shadowed adders only functioning to avoid the unwanted switching operation and power. B. Block diagram of proposed hybrid encoded low power multiplier The block diagram of the proposed hybrid encoded low power multiplier is shown in Fig 4. The process of the proposed multiplier can be divided into hybrid encoding, multiplication and controlling. The proposed encoder works as per the method explained in Fig 1. In the partial product compression the partial products are added without carry propagation and row bypassing can be used when the entire row of the PP is zero. This is expected to reduce the switching activity and power consumption. In the final adder unit a column bypassing provision is available to avoid the unwanted addition operation. The detection logic circuit is used to detect the effective data range. If the part of the input data does not make any impact in the final computing results then the data controlling circuit freezes that portion to avoid unnecessary switching transitions. A glue circuit can be used to control the carry and sign extension unit which will manage the sign. III. RESULTS AND DISCUSSIONS

(2AC9H) 0010101011001001 (0029H) X 00000000001010010 Booth recoding (8PP) Proposed recoding (1PP) +0+0+0+0+1-1-2+1 Category E PP1 PP2 PP3 PP4 PP5

001101101101000101001 000000000000000 0000000000000000 0000000000000000 0000000000000000 (6DA29H) 001101101101000101001

Figure 2. Proposed hybrid encoded multiplication

For the above multiplication, conventional multiplication scheme needs 16 partial products and 15 addition operations, Booth multiplication needs 8 partial products and 7 addition operation.

The low power multiplier circuit which is a part of MAC unit has been simulated using TANNER 12.6 EDA schematic editor.

75

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

ENCODING Multiplier

MULTIPLICATION Multiplicand

CONTROLING NG Multiplicand Multiplier

PP Generator Proposed Encoder PP Compression using Row Bypassing Detection Logic Asserting Circuit

Final Adder with Column Bypassing

Glue Circuit Sign Extension

PRODUCT
Figure 4. Block diagram of proposed hybrid encoded low power multiplier

The output of the EDA editor is shown in Fig 5. The simulated adder is shown in enlarged version for understanding. The power analysis of the proposed multiplier-adder circuit has been estimated with example. For multiplying 65, which is the pixel value in Multiply and Accumulate unit (MAC) with another pixel value 34, the proposed procedure shown in Fig 6 may be adopted.

Multiplicand (65) Multiplier (34) Booth recoding (4PP)

01000001 X 001000100 +1 -2 +1 -2 Category D

Proposed recoding (1PP) Result (2210)

100010100010

Figure 6. Hybrid encoded multiplication scheme for the pixel value multiplication

Figure 5. Architecture of multiplier with proposed adders

For the above multiplication, it needs 8 partial products for normal multiplication and 4 partial products for Booth recoding but only one partial product is enough for the proposed hybrid encoding method. Moreover, the proposed technique doesn’t need the 2’s complement process and virtual 0 which is to be placed as a first bit of Booth recoding. Table II shows the power and delay analysis of different multipliers for the stated multiplication. The simulation results have been taken for different voltage ranges from 0.8V to 2.4V. The power consumption of the proposed multiplier has been reduced by 87% and 26% compared with conventional and Booth multiplier.

76

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
TABLE II. Multiplier type Conventional multiplier (8 PP) [6] Booth multiplier (4 PP) [5] Proposed multiplier (1 PP) Parameter Power (µw) Delay (ns) Power (µw) Delay (ns) Power (µw) Delay (ns) POWER AND DELAY ANALYSIS OF DIFFERENT MULTIPLIERS VDD (volts)
0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4

31.98 11.20 13.71 4.800 4.569 1.600

84.56 5.138 36.24 2.200 12.08 0.734

122.5 4.165 52.50 1.790 17.50 0.595

167.8 3.213 69.75 1.380 23.25 0.459

246.9 2.765 105.8 1.190 35.27 0.395

413.7 2.443 177.3 1.050 59.10 0.349

525.0 2.296 225.0 0.980 75.00 0.328

625.59 2.1910 268.11 0.9400 89.370 0.3130

662.2 1.932 283.8 0.830 94.60 0.276

The full adder cell which is the important sub module of the proposed multiplier architecture is designed according to the following equations.

C S

AB BC CA A B C A BC AB C

ABC

In full adder, four inverters can be used to provide inverted inputs, the sum and carry circuits are joined together. A pull down nMOS transistor is connected near the carry output to provide the undistorted output. The output wave form of the full adder without and with the pull down transistor is shown in Fig 7 and Fig 8 respectively. Here 0.13µm TSMC technology files were used, for simulating in TANNER 12.6 EDA tool. The various adder circuits have been simulated using the TSPICE TANNER 12.6 EDA tool for supply voltages range from 0.8V to 2.4 V. The operating frequency is set at 100 MHz. The power consumption variation with various voltages is shown in Fig 9.
Figure 8. Output waveform of the full adder with pull down mechanism

Figure 7. Output waveform of the full adder without pull down mechanism.

Figure 9.Variation of power consumption with input for different adders.

77

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

It is seen from the figure that the 14T design consumes more power beyond the supply voltage range 0.8V. All other designs C-CMOS, TGA, TFA, CPL, hybrid and the proposed method are working better at the input supply voltage ranges from 0.8V to 2.4V. Even the number of transistors required to design TGA and TFA is less, they require additional buffers at the output. This additional buffer increases the short circuit power and also switching power because of less driving capacity. CPL adder design consumes more power than hybrid and C-CMOS due to its dual-rail structure and the large number of internal nodes. Even though the transistor count of the proposed adder design is more than the 10T and 14T, the proposed adder cell consumes less power than other design which is shown in the comparison. IV. CONCLUSION

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
[9]

[10]

[11]

[12]

[13] [14] [15] [16]

The performance of the SSST equipped low power multiplier with hybrid encoding has been estimated and compared with existing multipliers. The proposed encoding technique reduces the number of spurious switching activity and dynamic power consumption by analyzing the bit patterns in the input data. A low power full adder cell, which consumes less power than the other adders has been used to design the proposed multiplier. The switching activity of the proposed multiplier has been reduced by 86% and 46% compared with conventional and Booth multiplier respectively. It is observed from the device level simulation using TANNER 12.6 EDA that the power consumption of the proposed multiplier can be reduced by 87% and 26% compared with conventional and Booth multiplier. REFERENCES
[1] Choi.J, Jeon.J and Choi.K, “Power minimization of functional units by partially guarded computation,” in Proceeding of IEEE International Symposium Low Power Electron Devices, pp. 131–136, 2000. Chen.O, Sheen.R, and Wang.S, “A low power adder operating on effective dynamic data ranges,” IEEE Transaction on Very Large Scale Integration (VLSI) System, vol. 10, no.4, pp.435–453, 2002. Chen.O, Wang.S, and Wu.Y.W, “Minimization of switching activities of partial products for designing low-power multipliers,” IEEE Transaction on Very Large Scale Integration (VLSI) System, vol. 11, no.3, pp. 418–433, 2003. Benini.L, Micheli.G.D, Macii.A, Macii.E, Poncino.M, and Scarsi.R, “Glitching power minimization by selective gate freezing,” IEEE Transaction on Very Large Scale Integration. (VLSI) System, vol. 8, no. 3, pp. 287–297, 2000. Henzler.S, Georgakos.G, Berthold.J, and Schmitt-Landsiedel.D, “Fast power-efficient circuit-block switch off scheme,” Electronics Letter. vol. 40, no. 2, pp. 103–104, 2004. Huang.Z and Ercegovac.M.D, “On signal-gating schemes for low power adders,” Proceeding of 35 th Asilomar Conference on Signal, Systems & Computer.pp.867-871, 2001. Huang.Z and Ercegovac.M.D, “High performance low power left-toright array multiplier design,” IEEE Transaction on Computer., vol. 54, no. 3, pp. 272-283, 2005. Wen.M.C, Wang.S.J and Lin.Y.N, “Low-power parallel multiplier with column by passing,” Electronic Letter. vol. 41, no. 12, pp. 581– 583, 2005.

[17]

[18]

Chen.K.H and Chu.Y.S, “A low power multiplier with spurious power suppression technique,” IEEE Transaction on Very Large Scale Integration (VLSI) System.vol. 15, no.7, pp. 846–850, 2007. Ko.U, Balsara.P and Lee.W, “Low-power design techniques for high-performance CMOS adders,” IEEE Transaction on Very Large Scale Integration (VLSI) System. Volume. 3, no.2, pp.327-333, 1995. Song.P.J and De Micheli.G, “Circuit and architecture trade-offs for high-speed multiplication,” IEEE Journal on Solid-State Circuits, Vol.26, no.9, pp.1184-1198, 1991. Shams.A, Darwish.T and Bayoumi.M, “Performance analysis of low power 1-bit CMOS full adder cells,” IEEE Transaction on Very Large Scale Integration (VLSI) System, vol. 10, no. 1, pp. 20–29, 2002. Chandrakasan.A.P and Brodersen.R.W, Low Power Digital CMOS Design. Norwell, MA: Kluwer, 1995. Zhuang.N and Hu.H, “A new design of the CMOS full adder,” IEEE Journal on Solid-State Circuits, vol. 27, no. 5, pp. 840–844, 1992. Weste.N and Eshraghian.K, Principles of CMOSVLSI Design, A System Perspective. Reading, MA: Addison-Wesley, 1993. Vesterbacka.M, “A 14-transistor CMOS full adder with full voltage swing nodes,” in Proceedings of IEEE Workshop Signal Processing Systems, pp. 713–722, 1999. Bui.H.T, Wang.Y, and Jiang.Y, “Design and analysis of low-power 10-transistor full adders using novel XOR-XNOR gates,” IEEE Transaction on Circuits & System II, Analog Digital Signal Processing vol. 49, no. 1, pp. 25–30, Jan 2002. Chang.C.H, Gu.J, Zhang.M, “A review of 0.18-µm full adder performances for tree structured arithmetic circuits,” IEEE Transaction on Very Large Scale Integration (VLSI) System vol. 13, no.6, pp. 686–695, 2005.

AUTHORS PROFILE S.Saravanan received his B.E. Degree in Electrical and Electronics Engineering from Madras University, Tamilnadu, India in 2000 and M.E. Degree in Applied Electronics from Anna University, Tamilnadu, India in 2005. He is currently working towards the Ph.D degree in Information and Communication Engineering in Anna University, Chennai. Tamilnadu, India and working as an Assistant Professor in ECE Department, K.S.Rangasamy college of Technology, Tamilnadu, India. Dr. M. Madheswaran has obtained his Ph.D. degree in Electronics Engineering from Institute of Technology, Banaras Hindu University, Varanasi in 1999 and M.E degree in Microwave Engineering from Birla Institute of Technology, Ranchi, India. He has started his teaching profession in the year 1991 to serve his parent Institution Mohd. Sathak Engineering College, Kilakarai where he obtained his Bachelor Degree in ECE. He has served KSR college of Technology from 1999 to 2001 and PSNA College of Engineering and Technology, Dindigul from 2001 to 2006. He has been awarded Young Scientist Fellowship by the Tamil Nadu State Council for Science and Technology and Senior Research Fellowship by Council for Scientific and Industrial Research, New Delhi in the year 1994 and 1996 respectively. His research project entitled “Analysis and simulation of OEIC receivers for tera optical networks” has been funded by the SERC Division, Department of Science and Technology, Ministry of Information Technology under the Fast track proposal for Young Scientist in 2004. He has published 120 research papers in International and National Journals as well as conferences. He has been the IEEE student branch counselor at Mohamed Sathak Engineering College, Kilakarai during 1993-1998 and PSNA College of Engineering and Technology, Dindigul during 2003-2006. He has been awarded Best Citizen of India award in the year 2005 and his name is included in the Marquis Who's Who in Science and Engineering, 2006-2007 which distinguishes him as one of the leading professionals in the world. His field of interest includes semiconductor devices, microwave electronics, optoelectronics and signal processing. He is a member of IEEE, SPIE, IETE, ISTE, VLSI Society of India and Institution of Engineers (India).

[2]

[3]

[4]

[5]

[6]

[7]

[8]

78

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Using Sloane Rulers for Optimal Recovery Schemes in Distributed Computing
R. Delhi Babu
Department of Computer Science and Engineering SSN College of Engineering Chennai - 603 110. INDIA Email: rdelhibabu@gmail.com

P. Sakthivel
Department of Electronics and Communication Engineering Anna University Chennai Chennai - 600 025. INDIA Email: psv@annauniv.edu

Abstract—Clusters and distributed systems offer fault tolerance and high performance through load sharing, and are thus attractive in real-time applications. When all computers are up and running, we would like the load to be evenly distributed among the computers. When one or more computers fail, the load must be redistributed. The redistribution is determined by the recovery scheme. The recovery scheme should keep the load as evenly distributed as possible even when the most unfavorable combinations of computers break down, i.e., we want to optimize the worst-case behavior. In this paper we compare the worst-case behavior of schemes such as Modulo ruler, Golomb ruler, Greedy sequence and Log sequence with worst-case behavior of Sloane sequence. Finally we observe that Sloane scheme performs better than all the other schemes. Keywords: Fault tolerance, High performance computing, Cluster technique, Recovery schemes, Sloane sequence.

HACMP (IBM) [4], and node preference list in Windows Server 2003 cluster (Microsoft, earlier called MSCS) [9]. We consider that the computers are connected in a ring topology in this work. A recovery scheme specifies where to transfer a process when the computer in which it is running goes down. We call a transfer of a process from one computer to another as a jump. A jump is specified by a number, that gives which computer to resume the process. The jump is the number of computers to pass by in the ring. Hence, the jumps are the same wherever in the ring we start. The jump is only dependent on the number of previous jumps of the process: i.e. on the number of transfers for the process. If a process is transferred from computer A to computer B, and also computer B is down, the next jump in the recovery scheme is used, counting from computer B. We use the term ”wraparound” when the total sum of jumps for a process exceeds the number of computers. We would like to do an optimal recovery process. Here, optimal means that the maximal number of processes on the same computer after k crashes is BV(k) (Bound Vector). The function BV(k) provides a lower bound for any recovery scheme [8]. ‘Greedy’, ‘Golomb’ and ‘Modulo’ schemes [5] [6] are optimal for a larger number of computers than ‘Log’. The ‘Modulo’ rule gives better optimal result for a larger number of computers down than the Golomb schemes and Greedy scheme. Both Golomb and Greedy recovery schemes consider the formulation where wrap-around is not taken into account whereas in this paper we use it as in ‘Modulo’ scheme. In this paper we use a sharper mathematical formulation of the Sequence Optimization problem, and give a new recovery schemes called Sloane scheme. These are optimal for a larger number of computers down in the original Sequence Optimization problem. i.e. these results represent state-of-the-art in the field. This paper is organized as follows: The problem is define and explained in In Section II. We briefly review the existing works of Optimal Recovery Schemes and general results in Section III. In section IV, the features of Sloane scheme is explained. The General Theorems are defined and proved in Section V. The performance analysis of all the other schemes are given in Section VI. In Section VII the paper is concluded with its results.

I. I NTRODUCTION One way of obtaining high availability and fault tolerance is to execute an application on a cluster or distributed system. There is a primary computer that executes the application under normal conditions and a secondary computer that takes over when the primary computer breaks down. There may also be a third computer that takes over when both the primary and secondary computers are down, and so on. Another advantage of using distributed system or cluster, besides fault tolerance, is load sharing between the computers. When all computers are up and running, we would like the load to be evenly distributed. The load on some computer will, however, increase when one or more computers are down. But under these condition, we would like to distribute the load as evenly as possible on the remaining computers. The distribution of the load when a computer goes down is decided by the recovery list of the processes running on the faulty computer. The set of all recovery lists is referred to as the recovery schemes. Load balancing and availability are specially important in fault tolerant distributed system, where it is difficult to predict on which computer the process should be executed. This problem is NP-complete for the large number of computers. Most cluster vendors support this kind of error recovery, e.g. the node list in Sun Cluster [12], the priority list in MC/ServiceGuard (HP) [3], the placement policy in TruCluster (DEC) [2] [13], Cascading resource groups in

79

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

II. P ROBLEM DEFINITION We consider a cluster with n identical computers with one process on each computer. The work is evenly split between these n processes. There is a recovery list associated with each process. This list determines where the process should be restarted if the current computer breaks down. The set of all recovery lists is referred to as the recovery scheme. Fig. 1 shows such a system for n = 4. We assume that processes are moved back as soon as a computer comes back up again. In most cluster systems this can be configured by the user [3] [12] [13], i.e. in some cases one may not want automatic relocation of processes when a faulty computer comes back up again. The left side of the fig. 1 shows the system under normal conditions. In this case, there is one process on each computer. The recovery lists are also shown; one list for each process. The set of all recovery lists is referred to as the recovery scheme. The right side of Fig. 1 shows the scenario when computer zero breaks down. The recovery list for process zero shows that it should be restarted on computer one when computer zero breaks down. If computer one also breaks down, process zero will be restarted on computer two, which is the second computer in the recovery list. The first computer in the recovery list for process one is computer zero. However, since computer zero is down, process one will be restarted on computer three. Consequently, if computers zero and one are down, there are two processes on computer two (processes zero and two) and two processes on computer three (processes one and three). If computers zero and one break down the maximum load on each of the remaining computers is twice the normal load. This is a good result, since the load is as evenly distributed as possible. However, if computers zero and two break down, there are three processes on computer one (processes zero, one and two), i.e. the maximum load on the most heavily loaded computer is three times the normal load. Consequently, for the recovery scheme in Fig. 1 the combination of computers zero and two being down is more unfavorable than the combination of computers zero and one being down. Here, we are interested in the worst-case behavior. Our results are also valid when there are n external systems feeding data into the cluster, e.g. one telecommunication switching center feeding data into each computer in the cluster. If a computer breaks down, the switching center must send its data to some other computer in the cluster, i.e. there has to be a ”recovery list” associated with each switching center. The fail-over order can alternatively be handled by recovery lists at the communication protocol level, e.g. IP takeover [10]. In that case, redirecting the communication to another computer is transparent to the switching center. We assume that the work performed by each of the n computers must be moved as one atomic unit. Examples are systems where all the works performed by a computer is generated from one external

system or when all the works are performed by one process, or systems where the external communication is handled by IP takeover. III. O PTIMAL R ECOVERY S CHEMES Here we review previous works in which algorithms that give recovery schemes for a number of crashed computers. In [8] the problem of finding a recovery scheme that can guarantee optimal worst-case load distribution when at most x computers are down is presented for the first time. The schemes should have as large k as possible. The Log algorithm generates the recovery schemes that guarantee optimality when at most log2 n computers go down. Here optimal means that the maximal number of processes on the same computer after k crashes is BV(k), where the function BV (k) provides a lower bound for any static recovery scheme. BV is by definition increasing and contains exactly k entries that equals x for all x ≤ 2. The j-th entry in the 2(j + 1) + 1/2 . Hence, BV(k) = vector BV(k) equals 2, 2, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 5, 6, .... . Klonowska et al. [5] presents two other algorithms, the Greedy algorithm and the Golomb scheme. These algorithms generate the recovery schemes that give better optimality than the Log algorithm [7], (i.e. better load balancing). The Greedy algorithm is based on the mathematical problem of finding the sequence of positive integers such that all sums of subsequences are unique and minimal. It is easy to calculate the Greedy algorithm even for large n. The Golomb scheme is a special case of the Greedy algorithm. However, finding (and proving) optimal Golomb schemes becomes exponentially more difficult as the number of computers(n) increases. Therefore, for large n one can easily calculate a sequence with distinct partial sums with the Greedy algorithm [13] [15]. In [6] the problem is optimized by taking into account the wrap-around scenario: process being sent backwards or passing by the initial computer. This corresponds to the new mathematical problem of finding the longest sequence of positive integers for which the sum of all subsequences are unique modulo n. This mathematical formulation of the Sequence Optimization problem gives new and more powerful recovery schemes, called modulo schemes. In our scheme [1], the number of cluster computers n=140. ”In worst case scenario our scheme gives better results”. The problem of finding recovery schemes for any number of crashed computers by an exhaustive search, where brute force testing is avoided by a mathematical reformulation of the problem and a branch-and-bound algorithm is given in [7]. The search nevertheless has a high complexity. Optimal sequences, and thus a corresponding optimal bound, are presented for a maximum of twenty one computers in the distributed system or cluster.

80

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Fig. 1.

An application execution on a cluster with four computers.

Although these schemes are optimal, the worst case scenario needs to be improved. We provide an alternative scheme based on Sloane sequence which gives better performance when compared with the above mentioned schemes on worst case scenario. IV. S LOANE S CHEMES An intuitive Recovery Scheme (RS) is Ri = { (i+1) mod n,(i+2) mod n,(i+3) mod n, . . . ,(i+n-1) mod n} ,where Ri is the recovery list for process i. If n=4 we get: R0 = {1, 2, 3} , R1 = {2, 3, 0}, R2 = {3, 0, 1}, and R3 = {0, 1, 2}. The intuitive scheme is not optimal for n=4 ( in that case BV(2)=2 but Vector Length VL(4,2,{0, 1},RS) = 3). For n = 8 we know that BV(5) but VL(8,5,{0, 1, 2, 3, 4},RS) = 6. Consequently, the worst-case behavior of the intuitive scheme is twice as bad as an optimal scheme for n = 8. It was previously shown that when at most x = log2 n computers break down VL(i) = BV(i)(i ≤ x). If we (for some integer y >1) have x = Σy i ≤ log2 n , the i=2 difference between the intuitive scheme and optimal scheme can be Σy i/y (i.e. VL(x) = y) but the load on the heavily i=1 loaded computer in the intuitive scheme may be as high as: x + 1 = Σy i + 1 = Σy i . i=2 i=1 This occurs when the computers 0,1,2,....,x-1 are down, e.g. for y = 3 and thus x = 2+3 =5 we get a difference of (1+2+3)/3 = 2 when the computers 0,1,2 and 3 are down. We will now define a recovery scheme called the Sloane recovery scheme which is based on [11] that can guarantee optimal behavior and also when significantly more than log2 n

computers break down. We will show that this recovery scheme is optimal for number of cases. We start by defining R0 , i.e. the recovery list for process zero (n = {1, 2, 3, ...}), and x number of jumps to the r-th computer by r(x).

s1 = set of prime numbers s2 = {x|x + 1 ∈ s1 } s3 = {Σa∈{1,2,..x}\a|x a | x ∈ s2 }

(1) (2) (3)

s4 is required where in the element of s3 are arranged in required order. For example, n = 7 we get s4 = {1, 3, 7, 2, 4, 5, 6}. From the definition above, The sequence of n that is s(n) begin as follows for n ≥ 1: 1,3,7,12,18,28,31,39. In Sloane’s online encyclopedia of integer sequence [14] this is sequence number A008332.We see that for large values of n (i.e. n >140) the first 9 elements in R0 for the Sloane recovery schemes are: 0,1,4,11,23,41,69,100,139. We now define the new vector of length x. This vector is called the reduced step length vector (r ) and it consist of the first x entries in the step length vector r, where x = max(i), such that R0 (i) < n. From the definition of Sloane recovery scheme we know that all sums of subsequent in r are unique. Table I illustrates this for n = 140 resulting in a reduced step length of vector 9 (9 = x = max(i), such that R0 (i) < n = 140).

81

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
TABLE I A LL SUM OF SUBSEQUENCE OF THE REDUCED STEP LENGTH VECTOR FOR n = 140. Entry No(i) Sum of subseq Sum of subseq Sum of subseq Sum of subseq Sum of subseq Sum of subseq Sum of subseq Sum of subseq Sum of subseq of of of of of of of of of length length length length length length length length length 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9 0 1 4 11 23 41 69 100 139 1 3 7 12 18 25 31 39 4 10 19 30 46 59 70 11 22 32 53 77 98 23 40 65 39 116 41 68 96 128 69 99 135 100 138 139

n, ..., (y − Σx r (i) + n) mod n. i=x From the definition of r we know that all the sums of subsequences Σb r(l), (1 ≤ a ≤ b ≤ x) are unique which i=a are shown in table I. This means that no computer is included in more than one of the list. Thus, the worst-case is clear when the computers in the shortest list have crashed, e.g. for x = 3 the worst-case occurs when computer, (y−Σ1 r (i)+n) mod i=1 n have crashed, and then when computers (y − Σ2 r (i) + n) i=1 mod n, (y−Σ2 r (i)+n) mod n have crashed. By comparing i=2 this proof of Theorem 2 we see that the first x entries in B and VL(n,RS)) are identical (RS = Sloane Recovery Scheme, x = max(i) such that R0 (i) < n).

V. G ENERAL T HEOREMS Theorem 1. V L(i) ≤ n/(n − i) ,V L(i) is entry number i in VL [8] Proof: If i computers are down, there are i processes which must be allocated to the remaining n − i computers. The best one can hope for is obviously to obtain a load of n/(n − i) processes on the most heavily loaded computer. Theorem 2. BV is consecutively smaller than V L.[8] Proof: Because of Theorem 1 we know that BV(n-1) < n ≤ V L(n − 1). Based on theorems 1 and 2, we define B(i)=max(BV(i), n/(n − i) ). Next we prove that our proposed scheme is optimal. Theorem 3. The Sloane recovery scheme is optimal as long as x computes or less have crashed, where x = max(i), such that R0 (i) < n) be the heaviest loaded computer when x computers have crashed, where x = max(i), such that R0 (i) < n.[1] Proof: When x computers have crashed, process z(0 ≤ z < n) will in the ith step end on computer (z + Σi r(j)) j=1 mod n(1 ≤ i ≤ x).This means that a process ends up on computer y after i steps was originally allocated to computer (y − Σi r (j) + n) mod n; and this process has passed j=1 computers (y−Σi r (j)+n) mod n, (y−Σi r (j)+n) mod j=2 j=3 n, ...., (y − Σi r (j) + n) mod n before it reaches computer j=i y. The list below show the x possible sequence of computers need to be down in order for extra process to end up on computer y, i.e. if computer (y − r (1) + n) mod n is down one extra process will end up on computer y, if computers (y − r (1) − r (2) + n) mod n and (y − r (2) + n) mod n are down another extra process will end on computer y and so on. 1) (y − Σ1 r (i) + n) mod n i=1 2) (y − Σ2 r (i) + n) mod n, (y − Σ2 r (i) + n) mod n i=1 i=2 and so on. In general for xth x) (y − Σx r (i) + n) mod n, (y − Σx r (i) + n) mod i=1 i=2

VI. P ERFORMANCE WITH ALL OTHER S CHEMES In fig. 2 Sloane schemes is compare with all the other schemes. For example, in the case of n = 100 computers in a cluster, Modulo-m sequences guarantee optimal behavior in the case of 11 crashes, while Golomb rulers only guarantee optimality for 10 crashes, Greedy scheme for 9 crashes and Sloane scheme for 6 crashes. The advantage with the Sloane and Greedy algorithm compared to other schemes is that we can easily calculate a sequence with distinct partial sum. Apart from this, our scheme has the following advantages:
• •

Overall our scheme produced best results even under worst case scenario. When n > 45 our scheme performs better than all the other schemes (see Fig. 2 for n = 100).

In many cluster and distributed systems, the designer must provide a recovery scheme. Such schemes define how the workload should be redistributed when one or more computers break down. The goal is to keep the load as evenly distributed as possible, even when the most unfavorable combinations of computers break down, i.e. we want to optimize the worstcase behavior which is particularly important in real-time systems. We consider n identical computers, which under normal conditions execute one process each. All processes perform the same amount of work. Recovery schemes that guarantee optimal worst-case load distribution, when x computers have crashed are referred to as optimal recovery schemes for the values n and x. The contribution in this work is that we have shown that the problem of finding optimal recovery schemes for a system with n computers corresponds to the mathematical problem of finding the longest sequence of positive integers such that the sum and the sums of all subsequences number are unique. No efficient algorithm that finds the longest sequence with these properties is known. We have previously obtained recovery schemes that are optimal when a larger number of computers are down and they cover or do not cover load balancing when wrap-around occurs.

82

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Fig. 2.

Performance with Sloane Schemes

VII. C ONCLUSION In this paper, we present the Sloane sequences that minimize the maximum load. Modulo sequences allow optimal behavior for a larger number of crashed computers than Golomb, Greedy and Sloane sequences. For example, in the case of n = 100 computers in a cluster, Modulo-m sequences guarantee optimal behavior in the case of 11 crashes, while Golomb rulers only guarantee optimality for 10 crashes, Greedy scheme for 9 crashes and Sloane scheme for 6 crashes. Golomb rules are known for lengths upto 41912 ( with 211 marks). Of these the first 373 ( with 23 marks) are known to be optimal while Modulo rulers are known only for 13 marks. Modulo sequence are known up to 92 computers in the cluster and the Sloane schemes are known only up to 140 computer in a cluster. Our recovery schemes can be immediately used in commercial cluster systems, e.g. when defining the list in Sun Cluster using the scconf command. The results can also be used when a number of external systems, e.g. telecommunication switching centers, send data to different nodes in a distributed system (or a cluster where the nodes have individual network addresses). In that case, the recovery lists are either implemented as alternative destinations in the external systems or at the communication protocol level. R EFERENCES
[1] R. Delhi Babu, ”Optimal Recovery Schemes in Distributed Computing a Short Survey”, International Conference on Mathematics and Computer Science (ICMCS) 2009, Vol. 2, pp. 458 - 461. [2] ”TruCluster Server - Cluster Highly Available Applications”, HewlettPackard Company, September 2002. [3] ”Managing MC / ServiceGuard”, Hewlett- Packard, March 2002. [4] ”Concepts and Facilities Guide”, IBM, July 2002.

[5] K. Klonowska, L. Lundberg and H. Lennerstad, ”Using Golomb Rulers for Optimal Recovery Schemes in Fault Tolerant Distributed Computing”, Proc. of the 17th International Symposium on Parallel & Distributed Processing (IPDPS) 2003, pp. 213 - 218. [6] K. Klonowska, L. Lundberg, H. Lennerstad and C. Svahnberg, ”Using modulo rulers for optimal recovery schemes in distributed computing”, 10th International Symposium (PRDC) 2004, pp. 133-142. [7] K. Klonowska, L. Lundberg, H. Lennerstad and C. Svahnberg, ”Optimal recovery schemes in fault tolerant distributed computing”, Acta infomatica, Vol. 41 (3) pp. 314 - 365. [8] L. Lundberg and C. Svahnberg, ”Optimal Recovery Schemes for HighAvailability Cluster and Distributed Computing”, Journal of Parallel and Distributed Computing, Vol. 61(11), pp. 1680 - 1691. [9] ”Server Clusters: Architecture Overview for Windows Server 2003”, Microsoft Corporation, March 2003. [10] G. F. Pfister, ”In Search of Clusters”, Prentice-Hall, 1998. [11] N. J. A. Sloane, and S. Plouffe, ”The encyclopaedia of interger sequence”, Academic Press, 1995. [12] ”Sun Cluster 3.0 Data Services Installation and Configuration Guide”, Sun Microsystems, 2000. [13] ”TruCluster, Systems Administration Guide”, Digital Equipment Corporation, http://www.unix.digital.com/faqs/publications /cluster.doc. [14] http://www.distributed.net/ogr/index.html [15] http://www.research.ibm.com/people/s/shearer /grtab.html

83

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

ICD 10 Based Medical Expert System Using Fuzzy Temporal Logic
*P.Chinniah, **Dr.S.Muttan
Research Scholar, Department of ECE, CEG, Anna University, Chennai, INDIA. ** Professor, Centre for Medical Electronics, CEG, Anna University, Chennai, India .
Abstract-Medical diagnosis process involves many levels and considerable amount of time and money are invariably spent for the first level of diagnosis usually made by the physician for all the patients every time. Hence there is a need for a computer based system which not only asks relevant questions to the patients but also aids the physician by giving a set of possible diseases from the symptoms obtained using logic at inference. In this work, an ICD10 based Medical Expert System that provides advice, information and recommendation to the physician using fuzzy temporal logic. The knowledge base used in this system consists of facts of symptoms and rules on diseases. It also provides fuzzy severity scale and weight factor for symptom and disease and can vary with respect to time. The system generates the possible disease conditions based on modified Euclidean metric using Elder’s algorithm for effective clustering. The minimum similarity value is used as the decision parameter to identify a disease. Keywords -Fuzzy clustering, symptoms, fuzzy severity scale, weight factor, Minkowski distance, ICD, WHO, Rules Base, TSQL
*

determining the appropriate therapeutic actions become increasingly difficult. A single disease may manifest itself quite differently in different patients at different disease stages. Further, a single symptom may be indicative of several different diseases and the presence of several diseases in a single patient may disrupt the expected symptom pattern of any one of them. Although medical knowledge concerning the symptom – disease relationship constitutes one source of imprecision and uncertainty in the diagnostic process, the knowledge concerning the state of the patient constitute another. The physician gathers knowledge about the patient from past history, physical examination, laboratory test results and other investigative procedures such as x-ray and ultrasonic. Since the knowledge provided by each of these sources carries with it varying degrees of uncertainty, here we use a fuzzy temporal logic since the state and symptoms of the patient can be known by the physician with only a limited degree of precision. In this paper, we discuss about a medical expert system in which we use fuzzy logic to identify the diseases form the symptoms which helps to develop Fuzzy rules that can be stored in the knowledge base and can be fired during further decision process. Coding of medical reports to ICD is a difficult, multilevel process, in which various kinds of errors may occur (Gergely Heja, 2002). The Fuzzy medical expert system proposed in this paper uses ICD coding to represent data and also clustering algorithm in order to generate the most possible diseases for the given symptoms. Since, Knowledge base plays a vital role in developing a medical expert system (Hudson and Cohen, 1989), we have developed a knowledge base consisting of symptoms, diseases, question sets, fuzzy severity scale, weight factor and also ICD 10 data. This expert system provides an effective user interface which consists of pages for both doctors and patients. The doctor page displays part related symptoms and diseases with provisions for giving upper bound, lower bound values and weight factor for both selected symptom and disease. In this system, an interactive patient window is provided that generates the symptoms present in the particular part of the body when the user supplies data on body parts. For this, the whole body is divided into eight parts namely head, neck, chest, abdomen, pelvic, leg, arm and back. Each part is further subdivided into many subparts. A list of possible symptoms for each subpart is displayed by the system after user inputs are received by the system so that the user can specify the location of the symptom exactly. By selecting one or more symptoms, the associated conditions are displayed by the inference engine. After the user answers

1.0 INTRODUCTION There are three main relevant classes of information to be accessed by physicians when trying to reach a decision concerning a medical case namely expert’s opinion, colleague’s opinion and medical literature (WilliamW.Melek et al 2000). The expert opinion is necessary in medic-al decision making, since there are wide variations in clinical practices. Moreover, the growing need to assess and improve quality of health care has brought to light the possibility of developing and implementing clinical practice guidelines based on expert opinions. Even though the colleague’s opinion helps in accessing information about real cases which is another important source of information, an important goal to reach when dealing with real medical cases is to have simultaneous access to the expert’s opinion about the same indications of the real case being treated. The increase of the information volume in each medical field, due to the emergence of new discoveries, treatments, medicines and technologies, leads to a frequent need of consulting medical literature and in particular specialized revues and journals. Certainly, due to the huge volume of this information, a classified, targeted, access is necessary. In the field of medicine, Imprecision and Uncertainty play a large role in the process of diagnosis of disease that has most frequently been the focus of these applications. With the increased volume of information available to physicians from new medical technologies, the process of classifying different sets of symptoms under a single name and

84

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

the questions, the most possible diseases are generated by excluding the least possible diseases using fuzzy temporal logic based decision making. For this purpose, we use a temporal data base that stores past and current data of patients in separate tables In order to query the temporal data base effectively, we have developed a special query language called temporal structured query language (TSQL). The remainder of this paper is organized as follows: Section 2 provides a survey of related works and compares them with the work presented in this paper. Section 3 depicts the architecture of the medical expert system proposed and implemented in this work. This section explains the various components of the expert system. Section 4 explains the implementation techniques. Section 5 shows few results obtained in this work using graphs. Section 6 gives a conclusion on this work and suggests some possible future enhancements. 2.0 RELATED WORKS There are many works in the literature that explains about the design and implementation of medical expert systems. Shusaku Tsumoto(1999) had proposed a web based medical expert system in which the web server provides an interface between hospital information systems and home doctors. According to them, the recent advances in computer resources have strengthened the performance of decision making process and the implementation of knowledge base (Shusaku Tsumoto 2006) operations. Moreover, the recent advances in web technologies are used in many medical expert systems for providing efficient interface to such systems. Moreover, many such systems are put on the Internet to provide an intelligent decision support in telemedicine and are now being evaluated by regional medical home doctors. Vladimir Androuchko et al (2006) proposed an expert system called Medoctor, which is a webbased system and has a powerful engine to perform all necessary operations. The system architecture presented by them is highly scalable, modular, and accountable and most importantly enables the incorporation of new features to be economically installed in future versions. The user interface module of that system presents a series of questions in layman's language for knowledge acquisition and also to show the top three possible diseases or conditions. However this system lacks in accuracy in decisions and also it is not following the coding of diseases as per the standards. Hence, there is a need for proposing a system with increased accuracy and standard. Viviane et al (2004) proposed a medical expert system in which platform independency plays a vital role while developing their medical expert system. Time-factor in medical diagnoses is a challenging area of research since a formalization of timevarying situations for computer-assisted diagnosis systems presents a lot of open problems for investigators (Tatiana Kiseliova, et al 2001). In this paper, we propose a medical expert system that separates rules into non-temporal and temporal components in which both components can be used by the inference engine to make predictions and decisions using fuzzy and temporal rules. The major

advantages of this proposed medical expert system in comparison with the existing medical expert systems are the provision of temporal data base for storing the past and present data, a knowledge Base for inference using fuzzy temporal rules, ICD coding and a user interface for knowledge acquisition and querying. 3.0 SYSTEM ARCHITECTURE The architecture of the system proposed and implemented in this research work is shown in Fig.1. This system consists of seven major components namely user interface, ICD coding module, Inference Engine, Temporal Information Manager, Temporal Fuzzy Decision Manager and Knowledge Base. The user interface accepts details regarding symptoms, vital signs, diseases and stores them in the knowledge base during knowledge acquisition. Moreover this user interface is used to create and manipulate a temporal data base for maintaining the patient history through the temporal information manager.

ICD Coding Module

User Interface

Temporal Information Manager

Inference Engine

Temporal Data Base

Knowledge Base

Fuzzy Decision Manager
Figure1. System Architecture

The temporal information manager manages the temporal data base using TSQL commands which process queries by applying instant comparison operators and interval comparison operators. This system uses time series forecasting method to predict the future using past and present data .The temporal data base maintains two tables for each entity set to store the current data and historical data separately. The inference engine has two components namely a scheduler for scheduling the rules to be fired and an interpreter that fires the rules using forward chaining inference technique. The knowledge base is used to store

85

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

rules related to patient’s symptoms and diseases. The fuzzy decision manager applies fuzzy rules and coordinates with inference engine to make decision on diseases. The ICD coding module is used to store the ICD codes of diseases and symptoms, so that it is possible to make decisions using standards. 3.1 Temporal Database Design The database plays a vital role in decision making since it provides the necessary facts. The temporal data base consists of symptoms, diseases, question sets from which the user can interact with system for the possible diseases (Ahamd Hansnah, 1994), fuzzy severity scale, weight factor and also ICD 10 data. There are totally 3 tables in the data base which are: (i)Tables listing all possible conditions for symptom along with ICD 10 code. (ii) Tables listing all possible symptoms for every condition along with ICD 10 code. (iii) Tables of severity scale with upper bound, lower bound values and weights for every disease and symptoms present in the human body. Here, both the severity scale and the weight vary between zero and one for any given symptom. In this work, for the classification of diseases, the whole body is divided into eight anatomical parts and each anatomical part has many subparts. For every subpart, a symptom and disease table and a relationship between them are created. A unique code following the specification of ICD10 given by World Health Organization has been assigned to every disease and every symptom. Here, the diseases and symptoms are grouped subpart wise for better understanding and identification of diseases and symptoms in a particular part of the body. Here we have compared the number of symptoms and the number of diseases for each sub part of the body to present information about complexities that will arise if proper care is not provided to check parts like Head and Back which have large number of symptoms and diseases. The number of symptoms and diseases has been worked out and it varies for each subpart of the human body. Table 3.1 lists the possible number of symptoms and diseases for every subpart. There are totally 839 symptoms and 4210 diseases available for all the subparts. In this work, tables are created for every subpart. Each table has symptoms and corresponding ICD 10 code with number of symptoms ranging from 10 to 80 depending on the part selected by the user. Moreover, there are 44 tables that have been created for different types of diseases. Each table has symptom and their corresponding conditions with ICD code. The total number of conditions in each table varies from 40 to 230. Along with it, there are tables for common and general symptoms and conditions. The database also contains a table for skin symptoms and diseases. For every symptom provided by the user this system generates the possible conditions using the respective tables for performing disease diagnosis.

TABLE 3.1 TABLES OF SUBPARTS

Part Subparts Head Head Ears Eyes Nose Mouth Face Neck Chest Neck Chest Side of chest Sternum. Abdomen Upperabdomen Lowerabdomen. Pelvic Inguinal Pelvis Genital Hip Arm Fingers Palm Wrist Forearm Elbow Upper arm Shoulder Leg Foot Ankle Shin knee Thigh Toe Back Sole Calf Hamstring Back Upper spine Lower spine

No of symptoms 84 16 75 20 66 21 38 34 11 16 22 27 14 23 35 20 32 23 11 16 20 14 13 21 13 18 19 16 18 18 17 19 16 14 18

No of diseases 543 76 327 101 248 88 221 218 46 94 166 158 56 83 160 79 149 102 66 56 89 59 75 94 83 69 97 69 97 86 75 68 626 64 70

86

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

TABLE 3.2 DATABASES OF ALL SUB PARTS OF HUMAN BODY

Ankle Ankle Con Armpit Armpit Con Back Back Con Back of knee Back of knee Con Buttock Buttock Con Calf Calf Con Chest Chest Con Common Con Common skin Con Common skin symptom Common symptom Doctor Ears Ears Con Elbow Elbow Con Eyes

Eyes Con Face Face Con Fingers Fingers Con Foot Foot Con Forearm Fore arm Con General Con General skin symptom General symptom Genital Female Genital Female Con Genital Male Genital Male Con Hamstring Hamstring Con Head Head Con HipFe male Hip Female Con Hip Male Hip Male Con

Inguinal Female Inguinal Female Con Inguinal Male Inguinal MaleCon Jaw Jaw Con Knee Knee Con Lower abdomen Lower abdomen Con Lower spine Lower spine Con Main parts Mouth Mouth Con Neck Neck Con Nose Nose Con Palm Palm Con Patient Pelvis Female Pelvis Female Con Pelvis Male

Pelvis Male Con Register Shin Shin Con Shoulder Shoulder Con Side of chest Side of chest Con Skin Con Sole Sole Con Sternum Sternum Con Thigh Thigh Con Toes Toes Con Upper abdomen Upper abdomen Con Upper arm Upper arm Con Upper spine Upper spine Con Wrist Wrist Con

3.2. Knowledge Base Creation The system has been designed in such a way that it has a strong knowledge base. The knowledge base consists of rules for 44 symptoms in which each rule has symptoms and corresponding ICD 10 code with the number of symptoms ranging from 10 to 80 depending on the part selected by the user. Moreover, there are many rules for the 44 symptoms that are created for making decisions. Each rule has symptom and their corresponding conditions with ICD code. The total number of conditions in each rule base varies from 40 to 230. Along with it, there are separate rules for common and general symptoms and conditions. It also contains a rule base for skin symptoms and diseases. For every symptom the user is selecting, the rule base generates the possible conditions from the above said table for disease diagnosis. 3.3 ICD10 The International Classification of Disea-ses is published by the World Health Organization (WHO). The International Statistical Classification of Diseases and Related Health Problems (most commonly known by the abbreviation ICD) provides codes to classify diseases and a wide variety of signs, symptoms, abnormal findings, complaints, social circumstances and external causes of injury or disease. Every health condition can be assigned to a unique category and given a code, up to six characters long. Such categories can include a set of similar diseases. The ICD is used world-wide for morbidity and mortality statistics, reimbursement systems and automated decision support in medicine (William Melek and Alireza Sadeghian, 2000). This system has been designed to promote international compatibility in the collection, processing, classification, and presentation of these statistics. The ICD is a core classification of the WHO. The ICD is revised periodically by WHO and is currently in its tenth edition. The ICD is a core classification of the WHO-FIC. The ICD10, as it is therefore known, was developed in 1992 to track mortality statistics. ICD-11 is planned for 2011 and has become the most widely used statistical classification system in the world. 4.0 DECISION MAKING PROCESS. The decision making process is initiated by the inference engine whenever it receives user queries in the form of TSQL SELECT statement (Snodgrass). The TSQL SELECT statement provides an additional WHEN condition in addition to the WHERE condition provided in the SQL syntax. Using this WHEN condition, the user can ask the system to perform instant comparison, interval comparison, explanation of the past and prediction of the future. Other feature of the TSQL are automatic addition of temporal attributes with instant stamping of tuples for transaction time and interval stamping of tuples for valid time. Moreover, it provides options for history maintenance in the UPDATE and DELTE statements by extending the corresponding statements provided in SQL. In this work,

The database consisting of list of all diseases is shown in Table 3.2. The temporal data base for symptoms, signs and diseases has been created using TSQL and also by using ICD 10 code as shown in Table 3.3. Hence the user can access any of the symptoms or signs and diseases with their corresponding ICD code for their reference.
TABLE 3.3 SYMPTOM TABLE WITH ICD10

87

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

decisions are made by applying the K-means clustering algorithm (Mac Queen.J.B, 1967) on the symptoms and patient histories. Then instant comparison operations such as BEFORE, AFTER and AT (Sarda.N.L, 1990)as well as Allen’s interval algebra operators (James F Allen, 1983) are used for analysis of the past. For predicting the future, least square method is used in which a curve is fit based on the past and present data. From this curve, interpolation and extrapolation techniques are used to make a first level decision. A set of 5 to 10 decisions are made using temporal information manager and inference engine. They are sent to the fuzzy logic decision manager for providing weights. Based on the fuzzy scores, top 3 decisions are obtained by the inference engine. The decisions are communicated to the user interface and a user interaction is carried out. Based on this user interaction and temporal data, a final decision is derived by the inference engine. The system developed in the present work is a user friendly and platform independent and hence it is helpful for the end user and also the medical physician as well as the knowledge engineer. On receiving the input symptoms the system generates the possible disease conditions based on modified Euclidean metric using Elder’s algorithm (Elder R.C.,Esogbue A.O 1984). After login, if the patient selects the symptoms STRANGE SMELL, SNEEZING, NASAL CONGESTION, RUNNY NOSE with their severity scales as Patient X = [0.1, 0.7, 0.4, and 0.6]. With grade of membership in fuzzy set (Rudolf Seising, 2006). Then the system infers the diseases using fuzzy temporal logic by generating minimum distances as 0.39, 0.19 and 0.54 for the possible diseases namely DUST EXPOSURE, COMMON COLD and FOREIGN OBJECT IN NOSE respectively. In this case, the most likely disease corresponding to minimum value (0.19) of the distance in the similarity measure is COMMON COLD. 5.0 RESULTS AND DISCUSSIONS The decision accuracy has been improved in this work as a result of introducing fuzzy temporal logic. Figure 2 shows the comparison of decision accuracy between conventional fuzzy logic based decision making and the decision making using fuzzy temporal logic. The figure 3 shows the existence of difference in the False Positive Rate by fuzzy logic method and the False Positive Rate obtained by using Fuzzy Temporal logic method. The False Negative Rate for fuzzy logic and fuzzy temporal logic is shown in Figure.3. It is clear that the decision accuracy increases with respect to increase in the number of samples and the false positive and false negative rate decrease with increase in the number of samples.

100 90 80 Correct Inference(%) 70 60 50 40 30 20 10 0 100 200 300 400 500 600 700 800 900 1000 Fuzzy logic Fuzzy Temporal logic

No. of patients

Figure 2. Accuracy of Decision making

100 90 80 False Positive Rate(%) 70 60 Fuzzy Logic 50 40 30 20 10 0 100 300 500 700 No. of patients 900 Fuzzy Temporal Logic

Figure 3. False Positive Analysis

88

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

100 90 False Negative Rate(%) 80 70 60 50 40 30 20 10 0
100 300 500 700 900

Fuzzy Logic Fuzzy Temporal Logic

No. of patients

Figure 4. False Negative Analysis

6.0 CONCLUSION AND FUTURE ENHANCEMENTS The expert system developed in this work is helpful to the physician, since this system not only saves time by assisting the physician but also avoids the situation where the physician needs to ask same set of questions repeatedly to the patients for the years to diagnose a disease. By this system, computerized and some intelligence added to the system in assisting the physician to take a decision. This system uses temporal logic and fuzzy logic for improving the inference process and hence inference is improved by 10% and false positive and false negative are also decreased by 7% and 9%. The major contribution of this paper is the use of ICD code, temporal rule into management and fuzzy logic based decision making. This system finds only the elementary and routine diseases in all parts of the body. In future, this system can be divided into separate expert systems that focus on only one area of human body so that the accuracy of decision making can be aided with a single domain expert REFRENCES
[1] Ahmad Hasnah(1994),‘Knowled-ge Acquisition for Computer Based Medical Systems, Comp-uter Science Department Illinois Institute of Technology Chica-go, 1994 IEEE. [2] ElderR.C.,EsogbueA.O(1984), ‘Fuzzy sets and systems: Theory and Applications’, Academic Press, New Yark, pp.223-242 [3] GergelyHeja(2002),‘Semi Auto-matic Classification of Clinical Diagnoses with Hybrid Appro-ach’,Compute Based Medical Sy-stems, 2002.(CBMS 2002). Proceedings of the 15th IEEE Symposium on 4-7 June 2002 Page(s):347 – 352. [4] Hudson D.L, ph.D, M.E.Cohen , Ph.D (1989), ‘Human-computer interaction in a medical decision support system’, University of California san Francisco, 1989 IEEE

[5] James F Allen(1983),‘Maintaining knowledge about Temporal Intervals‘, communications of ACM, Vol. 2, No.11, PP 832-843 [6] Mac Queen (1967), “Some method for classification and Analysis of multivariate observa-tions, proceedings of 5th Berkely symposiumonmathematicalstatisticsandprobability.”Berkeley,Uni-versity of California press 1:281-297. [7] Rudolf Seising (2006), ‘From vagueness in medical thought to theFoundations-of-fuzzy reason-ing in medical diagnosis’, Section on Medical Expert and Knowledge Based Systems, Aus-tria Artificial Intelligence in Medicine (2006), Elsevier. [8] Sarda.N.L(1990), ‘Extension to SQL for Historical Databases’, IEEE Transation on Knowledge and Data Engineering’, Vol.2 No.2 pp.220230 [9] Shusaku Tsumoto (1999), ‘Web based medical decision support system: Application of Internet to telemedicine’, Applic-ations and the Internet Worksh-ops, 2003. Proceedings. 2003 Symposium on27-31 Jan. 2003 Page(s):288 - 293. [10] Shusaku Tsumoto (2003), ‘Web based Medical Decision Support System for Neurological Diseas-es’, WebIntelligence, 2003. WI 2003. Proceedings.IEEE/WICInternation-al Conference on 13-17 Oct. 2003 Page(s):629 - 632. [11] Snodgrass.R.T and I.Ahn.‘Temp-oral Database’, IEEE compyter, Vol.19, No.9, pp.35-45 [12] Tatiana Kiseliova, Claudio Mora-ga (2003),‘Modeling TemporalDistribution of Symptoms and Diseases with Fuzzy Logic’, Proceedings of the IEEE/WIC International Conference on Web Intelligence (WI’03),2001.Vol.3, 25-28 July 2001 Page(s):1637 - 1641 [13] Viviane Van Hoofa, Arno Wor-mekb, Sylvia Schleutermannb, Theo Schumacherc, Olivier Loth-airec(2004),‘Medical Systems De veloped in j. MD, a Java Based System Shell: Application in Clinical Laboratories’.Medinfo, (2004), p. 89-93. [14] Vladimir Androuchko, Charlie Kelly (2006),‘Intelligent Medical Diagnostic System’, eHealth Net-working, Applications and Serv-ices, 2006. HEALTHCOM 2006. 8th International Conference on17-19 Aug. 2006 Page(s):196 – 197 [15] William W.Melek, Alireza Sade-ghian (2000), ‘A Neuro Fuzzy-based System for Disease Diagnosis’, Dept. of Mechanical Engineering, Department of Computer, Canada

P.Chinniah was born in India in 1966. He received the B.Tech degree in electronics engineering and ME degree in Medical Electronics from Anna University, Chennai, India in 1992 and 1998 respectively. He is currently pursuing Ph.D. in Medical informatics Anna University, Chennai, India. He has 14 years of Teaching Experience. His research interests are in medical informatics, medical standards, e-Health and medical expert systems. S.Muttan received the bachelor’s degree in ECE in1985, M.E degree with honors in Applied Electronics in 1991and the Ph. D. degree in evolution and design of integrated cardiac information system in multimedia at Anna University, India in 2001. He is currently working as a professor in Centre for Medical electronics, Department of electronics and communication engineering, college of engineering Guindy, Anna University, India. His research interests are in medical informatics, pattern recognition, e-health service and biometrics.

89

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

DNA-MATRIX: a tool for constructing transcription factor binding sites Weight matrix

Chandra Prakash Singh1 Research Scholar, Uttrakhand Technical Univ. Dehradun, india . Dr. Feroz Khan2 MSB Division, Central Institute of Medicinal & Aromatic Plants (CSIR), Lucknow (India)

Sanjay Kumar Singh3 Reader, Department of Computer Sciences, R.S.M.T., U.P. College, Varanasi (India). Prof. Durg Singh Chauhan4 Vice-Chancellor, Uttrakhand Technical Univ. Dehradun, india

ABSTRACT— Despite considerable effort to date, DNA transcription factor binding sites prediction in whole genome remains a challenge for the researchers. Currently the genome wide transcription factor binding sites prediction tools required either direct pattern sequence or weight matrix. Although there are known transcription factor binding sites pattern databases and tools for genome level prediction but no tool for weight matrix construction. Considering this, we developed a DNA-MATRIX tool for searching putative transcription factor binding sites in genomic sequences. DNA-MATRIX uses the simple heuristic approach for weight matrix construction, which can be transformed into different formats as per the requirement of researcher’s for further genome wide prediction and therefore provides the possibility to identify the conserved known DNA binding sites in the coregulated genes and also to search for a great variety of different regulatory binding patterns. The user may construct and save specific weight or frequency matrices in different formats derived through user selected set of known motif sequences. KEYWORDS; File format, weight matrix and motif prediction. INTRODUCTION It is one of the major challenges in molecular biology is to understand the mechanisms of gene regulation at the level of transcription. An important task in this challenge is to identify DNA regulatory elements for transcription factors. Despite considerable efforts to date, DNA motif finding remains a very complex challenge for computer scientists and biologists. Researchers have taken different approaches in developing motif discovery tools and the progress made in this area of research is very encouraging. Performance comparison of different motif finding tools and identification of the best tools have proven to be a difficult task because tools are designed based on algorithms and motif models that are diverse and complex and our incomplete understanding of the biology of regulatory mechanism does not always provide adequate evaluation of underlying algorithms over motif models. Recent advances in genome sequence availability and in high–

throughput gene expression analysis technologies have allowed for the development of computational methods for motif finding. As a result, a large number of motif finding algorithms have been implemented and applied to various motif models over the past few decades. There are several approaches to identify the regulatory elements but the recent one is through weight matrix based method. So far no such tool is available to construct the optimized weight matrices through aligned known motifs. Earlier algorithms use promoter sequences of co regulated genes from single genome and search for statistically overrepresented motifs. Recent algorithms are designed to use phylogenetic footprinting or orthologous sequences and also an integrated approach where promoter sequences of co regulated genes and phylogenetic footprinting are used. All the algorithms studied have been reported to correctly detect the motifs that have been previously detected by laboratory experimental approaches and some algorithms were able to find novel motifs. However, most of these motif finding algorithms have been shown to work successfully in yeast and other lower organisms, but perform significantly worse in higher organisms. Therefore, computational methods of predicting TF binding sites in DNA are very important for understanding the molecular mechanisms of gene regulation. Over the past few years, numerous tools have become available for the prediction of TF binding sites as in [18,10,1,6,5]. Especially popular are those tools which use information on known TF binding sites that are collected in databases such as TRANSFAC [13,], EpoDB [17], TRANSCompel [10]. Approaches vary between high generalizations provided by weight matrices to high specialization provided by pattern matching approaches. More sophisticated approaches include consideration of nucleotide correlation in different position of the sites, HMMs, taking into account flanking regions and some others as in [15,2,3,16,17,11,8]. But usually, complex approaches require large training sets, which is rather problematic for the most known TFs for which only small sets of binding sites are known (up to 10 sites). So considering this, we have developed a tool called DNA-MATRIX which is a weight matrix construction tool based on aligned known TF binding sites for searching putative transcription factor binding sites in genomic sequences. DNA-MATRIX uses a heuristic approach for position weight matrix construction, which can be used in other tools such as PoSSuMsearch and RSAT–Patser etc. Many web based tools are available such as SIGNAL

90

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

SCAN [15], MATRIX SEARCH [3], MatInsprctor [16], Fuzzy clustering tool [14], FUNSITE [10], Gibbs Sampling tool [12], AliBaba2 [8] and TESS [17] are best known. DNA-MATRIX differs from existing tools by providing liberty to design user defined weight matrix model & signature. ALGORITHM DNA-MATRIX takes aligned DNA Motif sequences ‘S’ and motif width ‘w’ as input, searches for nucleotide frequency at each position ‘F’(ij) and outputs the found consensus patterns/motifs according to conservation priority based on nucleotide frequency ‘F’ (ij)’, constructed frequency matrix, alignment matrix and weight matrix along with motif signature and degenerate consensus sequence according to IUPAC/IUB convention. Scoring of the weight matrix was done through following equation: F (ij) = ∑ (w) [ C (ri) /S] To calculate the nucleotide frequency ‘F(ij) ’ within date set of known motif sequences ‘S’, algorithm uses aligned set of motif sequences for calculating nucleotide repeat conservation ‘C (ri)’ at specific nucleotide position where ‘i’ refers nucleotide (e.g. A, T, G and C) at position ‘j’ in the motif width of ‘w’. This algorithm finds optimized local alignments in related sequences in order to detect short conserved regions of motifs that may not be in the same positions. The matrix length can be set from 4 to 25, which allowed the detection of very short and also longer and more complex conserved sequences. Matrices generated, which represent the nucleotide conservation in each position of the motifs, can be use to search for the motifs positions, copy number and conservation using the Patser and PoSSuMsearch algorithms. The algorithm can be applied to the matrices by transforming it in required format. Moreover, user can transform the matrix or sequences of the most conserved consensus motif as input file for WebLogo program to generate the visual representation of the consensus sequence. IMPLIMENTATION The software is developed in C and the program is wrapped by a Perl script. The top panel is used to paste the input sequences (or aligned known TF binding sites) and to specify the name and width of motif to be search. The results panel contains five major sections: consensus pattern/ motif sequence, frequency matrix, alignment matrix, weight matrix and signature sequence as per IUPAC code. DNA-MATRIX can use both orthologous and co-regulated genes upstream sequences as input data set. For testing, we used the known PurR transcription factor binding sites, to construct the weight matrix similar to earlier reported. Predictions performance showed promising results, as on comparison of weight matrix with known one, we found 92% accuracy with aligned motifs of same width.

CONCLUSION The algorithm is described DNA-MATRIX presented using the appropriate theoretical concepts. TF binding sites prediction in whole genome remains a challenge for the researchers because of the Chemistery of genomic-DNA interaction is poorly understood. DNA-MATRIX uses the simple heuristic approach for weight matrix construction, which can be transformed into different formats as per the requirement of researcher’s for further genome wide prediction and therefore provides the possibility to identify the conserved known DNA binding sites in the co-regulated genes and also to search for a great variety of different regulatory binding patterns. The user may construct and save specific weight or frequency matrices in different formats derived through user selected set of known motif sequences. For testing, we used the known PurR transcription factor binding sites, to construct the weight matrix similar to earlier reported. We found 92% accuracy with aligned motifs of same width. Most of the available web tools are showing their performance upto 100% for a particular case only, So that many research work is required in this direction for the better performance in all the sequences. ACKNOWLEDGEMENT Author is grateful to the technical reviewers for the comments, which improved the clarity and presentation of the paper. Author wishes to thank Mr. Naresh Sen for all the discussions and contributions during the Software development. REFERENCES [1] Bhcher P.(1999).Regulatory elements and expression profiles. Curr.Opin Struct. Biol., 9,4000-407. [2] Chandra Prakash Singh, Feroz Khan, Bhartendu Nath Mishra, Durg Singh Chauhan. “Performance evaluation of DNA motif discovery programs”. Bioinformation 3(5): 205212 (2008) . [3] Chen Q.k., Hertz G.Z. and Stormo G.D.(1995). MATRIXSEARCH 1.0: a computer program that scans DNA sequences for transcriptional elements using a database of weight matrices.Comput. Appl. Biosci.,11,563-566. [4] Conkright M.D.Guzman E. Flechner L.Su. A.I. Hogenesch J.B. and Montminy M.( 2003). Genome-wide analysis of CREB target genes revealsa core promoter requirement for cAMP responsiveness.Mol. Cell, 1101 –1108. [5] Fessele S., Maier H., Zischek C., Nelson P.J. and Wernere T. (2002).Regulatory context is a crucial part of gene function Trends Genet., 18,60-63. [6] Fickett J.W.and Wasserman W.W. (2000). Discovery and modeling of transcriptional regulatory regions .Curr. Opin. Biotechnol., 11,19-24.

91

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

[7] Gobling E. Kel_Margulis O.V., Kel A.E. and Wingender E. (2001).MATCH-TM A tool for searching transcription factor binding sites in DNA sequences. Application for the analysis of human chromosomes. proceedings of the German Conference on Bioinformatics GCB’01 Braunschweig, Germany , October 7-10,pp. 158-161. [8] Grabe N. ( 2000). AlibBaba2: context specific identification of transcription factor binding sites . In Silico Biol., 1,0019. [9] Kel A. Kel –Margoulis O. Borlack J., Tchekmenev D.And Wingender E. (2005). Databases and tools for in silico analysis of regulation of gene expression.In Borlak J. (ed.) ,Handrakhin of Toxicogenomics VCH Weinheim, pp. 253290. [10] Kel A.E. kondrakhin Y.V. ,Kolpakov Ph.A. Kel O.V. Romashenko A.G.,Wingender E., Milanesi L. And Kolchanov N.A. ( 1995). Compuater tool FUNSITE for analysis of eukaryotic regulatry genomic sequences. proc. Int. Conf. Intell. Syst. Mol. Bip;. 3,197-205. [11] Kel –Margoulis O.V., Kel A.E. Reuter I ., Deineko I.V.and Wingender E.(2002).TRANS Compel: a database on comoposite regulatory elements in eukaryotic genes. Nucleic Acids Res., 30,332.334. sequence signals : a Gibbs sampling strategy for multiple alignment.science, 262,208-214. transcriptional regulation, from patterns to profiles .Nucleic Acids Res., 31,374-378. [14] Pickert L., Reuter I., Klawonn F. and Wingender E.(1998). Transcription regulatory region analysis using signal detection and fuzzy clustering Bioindormatics,14 244-251. [15] Prestridge D.S(1996). SIGNAL SCAN 4.0: additional databases and sequence formats. Comput. Appl. Biosci., 12,157-160. [16] Quandt K., Frech K., Karas H., Wingender E. and Werner T. (1995). MatInd and MatInstpector : new fast and versatile tools for detection fo consensus matches in nucleoitide sequence dataNucleic Acids Res., 23, 4878-4884. [17] Stoeckert C.J. Jr Salas F., Brunk B. and Overton G.C. ( 1999). EpoDB: a prototype database for the analysis of genes expressed during vurtebrate erythropoiesis . Nucleic Acids Res 27,200-203. [18] Stormo G.D. (2000) .DNA Binding sites: representation and disxovery. Bioinformatics, 16,16-23. [19] Regulatory Sequence .htt:/rwat.ulb.ae.be./rsay/ Analysis Tools

AUTHORS PROFILE 1.Chandra Prakash Singh is currently a research scholar at Uttarakhand Technical University, Dehradun, India, with the Subject of Computer Science. He completed his master degree in computer application (M.C.A.) in 2001 from the M.G.K.V. University; he has published 2 research papers in international journals and 4 research paper in national journal. He is working as a Lecturer in Deptt.of Computer Application, R.S.M.T, Varanasi. He has 9 Years of the experience in the field of research/academics. 2. Dr. Feroz Khan is completed his M.Sc (2000) from the C.S.J.M. Kanpur Univ., M.Tech (2003), Ph.D. (2008) from the U. P. Tech. Univiversity. Now he is working as a Scientist-C in MSB Division, Central Institute of Medicinal & Aromatic Plants (CSIR), Lucknow (India). He has published 17 international research papers and more than 20 national research paper. He is Serving in the editorial board of academic journal i.e. ‘African Journal of Biotechnology’ (AJB) cited in PubMed, Science etc. as a Member of the Review Team, since July 12, 2006. 3. Sanjay Kumar Singh has completed his M. Tech (C.S.) from Uttar Pradesh Technical University, Lucknow. in 2006. He has done B.E. (C.S.) from NMU, Jalgaon, in 1998. He did PGDM with dual specialization in Marketing and IT from ITS, Ghaziabad in 2000. He received the outstanding staff award in IT from U.P.T. Univ. He is also working as a Reader in Deptt.of Computer Application, R.S.M.T, Varanasi. He has 11 Years of the experience in the field of academics/research. 4. Prof. D.S. Chauhan completed his B.Sc Engg.(1972) in electrical engineering at I.T. B.H.U., M.E. (1978) at R.E.C. Tiruchirapalli (Madras University) and PH.D. (1986) at IIT/Delhi. His brilliant career brought him to teaching profession at Banaras Hindu University where he was Lecturer, Reader and then has been Professor till today. He has been director KNIT Sultanpur in 1999-2000 and founder vice Chancellor of U.P.Tech. University (2000-2003-2006). Later on, he has served as Vice-Chancellor of Lovely Profession University (2006-07) and Jaypee University of Information Technology (2007-2009) Currently he has been serving as Vice-Chancellor of Uttarakhand Technical University for (2009-12) Tenure. He has supervised 12 Ph.D., one D.Sc and currently guiding half dozen research scholars. He has authored two books and published and presented 95 research papers in international journals and international conferences and wrote more than 20 articles on various topics in national magazines. He is Fellow of institution of Engineers and Member IEEE.

[20] Hertz G.Z., Stormok G.D. ( 1999). Identifying DNA and protein patterns with statistically signficant aligaments of multiple sequences. Bioinformatics,15:563-577. [21] Web Logo . htt:/weblogo. berle;eu. edu/logo.cgis

92

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

MULTIPROCESSOR SCHEDULING FOR TASKS WITH PRIORITY USING GA
Mrs.S.R.Vijayalakshmi Lecturer, School of Information Technology and Science, Dr.G.R.D College of Science, Coimbatore -14, India. . Abstract - Multiprocessors have emerged as a powerful computing means for running real-time applications, especially where a uni-processor system would not be sufficient enough to execute all the tasks. The high performance and reliability of multiprocessors have made them a powerful computing resource. Such computing environment requires an efficient algorithm to determine when and on which processor a given task should execute. In multiprocessor systems, an efficient scheduling of a parallel program onto the processors that minimizes the entire execution time is vital for achieving a high performance. This scheduling problem is known to be NP- Hard. In multiprocessor scheduling problem, a given program is to be scheduled in a given multiprocessor system such that the program’s execution time is minimized. The last job must be completed as early as possible. Genetic algorithm (GA) is one of the widely used techniques for constrained optimization problems. Genetic algorithms are basically search algorithms based on the mechanics of natural selection and natural genesis. The main goal behind research on genetic algorithms is robustness i.e. balance between efficiency and efficacy. This paper proposes Genetic algorithm to solve scheduling problem of multiprocessors that minimizes the make span. Key-Words: - Task Scheduling, Genetic Algorithm (GA), parallel processing. I INTRODUCTION Real-time systems are software systems in which the time at which the result is produced is as important as the logical correctness of the result. That is, the quality of service provided by the realtime computing system is assessed based on the main constraint ‘time’. Real-time applications span a large range of activities, which include production automation, embedded systems, telecommunication systems, nuclear plant supervision, surgical Dr.G.Padmavathi Professor and Head, Dept.of Computer Science, Avinashilingam University for Women, Coimbatore – 43,India. . operation monitoring, scientific experiments, robotics and banking transactions. Scheduling is an important aspect in realtime systems to ensure soft and hard timing constraints. Scheduling tasks involves the allotment of resources and time to tasks, to satisfy certain performance needs. In a real-time application, realtime tasks are the basic executable entities that are scheduled. The tasks may be periodic or aperiodic and may have soft or hard real-time constraints. Scheduling a task set consists of planning the order of execution of task requests so that the timing constraints are met. Multiprocessors have emerged as a powerful computing means for running real-time applications, especially where a uni-processor system would not be sufficient enough to execute all the tasks by their deadlines. The high performance and reliability of multiprocessors have made them a powerful computing means in time-critical applications. Real-time systems make use of scheduling algorithms to maximize the number of real-time tasks that can be processed without violating timing constraints. A scheduling algorithm provides a schedule for a task set that assigns tasks to processors and provides an ordered list of tasks. The schedule is said to be feasible if the timing constraints of all the tasks are met. All scheduling algorithms face the challenge of creating a feasible schedule. The two main objectives of task scheduling in real-time systems are meeting deadlines and achieving high resource utilization. Section 1 deals with Introduction. Section 2 deals about the task scheduling in parallel systems. Section 3 about the back ground study. Section 4 about the Genetic algorithm. Section 5 about algorithm design. Section 6 about the experimental results and final section gives the conclusion. II TASK SCHEDULING IN PARALLEL SYSTEMS Multiprocessor scheduling problems can be classified into many different classes based on

1

93

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

characteristics of the program and tasks to be scheduled, the multiprocessor system, and the availability of information. A deterministic scheduling problem is one in which all information about the tasks and their relations to each other, such as execution time and precedence relations are known to the scheduling algorithm in advance. Such problems, also known as static scheduling problems. In contrast to nondeterministic scheduling problems in which some information about tasks and their relations may be non-determinable until runtime, i.e., task execution time and precedence relations may be determined by data input. Within the class of deterministic scheduling problems, the following are the constraints: 1. The number of tasks. 2. Execution time of the tasks. 3. Precedence of the tasks. 4. Topology of the representative task graph. 5. Number of processors. 6. Processors uniformity. 7. Inter task communication. 8. Performance criteria. 2.1 Problem description The goal of multiprocessor scheduling is to find an optimization solution to minimize the overall execution time for a collection of subtasks that compete for computation. Given material for problem • A multiprocessor system with ‘m’ machines. • A task represented by a DAG. • The estimated execution duration of every subtask. 2.2 Problem Statement The goal of multiprocessor scheduling is to find an optimization algorithm to minimize the overall execution time for a collection of subtasks that compete for computation and also the maximum utilization of the processor time. A (homogeneous multiprocessor system is composed of a set P= {P1, …Pm} of ‘m’ identical processors. They are connected by a complete communication network where all links are identical. Task preemption is not allowed. While computing, processor can communicate through one or several of its links. A schedule to, n Minimize { Max [ finish time ( Vj ) ]} j=1

where, the schedule determines, for each subtask, both the processor on which execution will take place and the time interval within which it will be executed. The Problem statement can be given as follows: “Schedule ‘n’ jobs to ‘m’ processors such that the maximum span is minimized”. 2.3 Model A parallel program can be represented as a directed acyclic graph (DAG), G = (V,E), where V is the set of nodes each of which represents a component subtask of the program and E is the set of directed edges that specify both precedence constraints and communication paths among nodes. In the DAG model, each node label gives the execution time for the corresponding task. A task cannot start until all of its predecessor tasks are complete. For a task graph TG = (V, E):  Ti is a predecessor of Tj and Tj is a successor of Ti  Ti is an ancestor of Tj and Tj is a child of Ti if there is a sequence of directed edges leading from Ti to Tj.  PRED(Ti) – The set of predecessor of Ti  SUCC(Ti) – The set of successor of Ti  Et(Ti) – The execution time of Ti. A simple task graph TG, with 8 tasks is illustrated in fig 1.

T2 (3,0)

T1 (2,0)

T5 (2,1)

T4(3,1)

T3 (2,1)

T7 (2,2)

T6(3,2)

T8 (1,3)

Fig 1.A task graph TG

2

94

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

0
P1 P2

11
T3 T6 T7
T8

T1

T5 T2 Finish time: 11

Fig 2 Gantt chart for scheduling two tasks The problem of optimal scheduling a task graph onto a multiprocessor system with ‘p’ processors is to assign the computational tasks to the processors so that the precedence relations are maintained and all of the tasks are completed in the shortest possible time. The time that the last task is completed is called the finishing time (FT) of the schedule. Fig 2 shows a schedule for two processors displayers as Gantt chart. Fig 2 illustrates a schedule displayed as a Gantt chart for the example task graph TG using two processors. This schedule has a finishing time of 11 units of time. An important lower bound for the finishing time of any schedule is the critical path length. The critical path length, tcp of a task graph is defined as the minimum time required to complete all of the tasks in the task graph. 2.4 Height of a task graph: The height of a task in a task graph is defined as0, if PRED (Ti) = 0 Height (Ti) = 1 + max height (Tj) This height function indirectly conveys precedence relations between the tasks. If the task Ti is an ancestor of task Tj, then height (Ti) < height (Tj). If there is no path between the two tasks, then there is no precedence relation between them and order of their execution can be arbitrary. III BACKGROUND List scheduling techniques assign a priority to each task to be scheduled and then sort the list of tasks in decreasing priority. As processors become available, the highest priority task in the task list is assigned to be processed and removed from the list. If more than one task has the same priority, selection among the candidate tasks is typically random. In order to allocate parallel applications to maximize throughput, task precedence graph (TPG) and a task interaction graph (TIG) are modeled. The system usually schedules tasks according to their deadlines, with more urgent ones running at

higher priorities. . The Earliest Deadline First (EDF) algorithm is based on the dead line time constraint. The tasks are ordered in the increasing order of their deadlines and assigned to processors considering earliest deadline first. In multiprocessor real time systems static algorithms are used to schedule periodic tasks whose characteristics are known a priori. Scheduling aperiodic tasks whose characteristics are not known a priori requires dynamic scheduling algorithms. Some researchers analyze the task scheduling problems based on the dynamic load balancing. It minimizes the execution time of single applications running in parallel on multi computer systems. It is essential for the efficient use of highly parallel systems with non uniform problems with unpredictable load estimates. In a distributed real time systems, uneven task arrivals temporarily overload some nodes and leave others idle or under loaded. In the proposed work, the GA technique is involved to solve the task scheduling problem. In the proposed GA technique, the tasks are arranged as per their precedence level before applying GA operators. The cross over operator is applied for the tasks having different height and mutation operator is applied to the task having the same height. The fitness function attempts to minimize processing time. IV GENETIC ALGORITHMS Genetic algorithms try to mimic the natural evolution process and generally start with an initial population of individuals, which can either be generated randomly or based on some algorithm. Each individual is an encoding of a set of parameters that uniquely identify a potential solution of the problem. In each generation, the population goes through the processes of crossover, mutation, fitness evaluation and selection. During crossover, parts of two individuals of the population are exchanged in order to create two entirely new individuals which replace the individuals from which they evolved. Each individual is selected for crossover with a probability of crossover rate. Mutation alters one or more genes in a chromosome with a probability of mutation rate. For example, if the individual is an encoding of a schedule, two tasks are picked randomly and their positions are interchanged. A fitness function calculates the fitness of each individual, i.e., it decides how good a particular solution is. In the selection process, each individual of the current population is

3

95

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

selected into the new population with a probability proportional to its fitness. The selection process ensures that individuals with higher fitness values have a higher probability to be carried onto the next generation, and the individuals with lower fitness values are dropped out. The new population created in the above manner constitutes the next generation, and the whole process is terminated either after a fixed number of generations or when a stopping criteria is met. The population after a large number of generations is very likely to have individuals with very high fitness values, which imply that the solution represented by the individual is good; it is very likely to achieve an acceptable solution to the problem. The population size, the number of generations, the probabilities of mutation and crossover are some of the other parameters that can be varied to obtain a different genetic algorithm. V ALGORITHM DESIGN For genetic algorithm, a randomly generated initial population of search nodes is required. It impose the following height ordering condition on schedules generated: “The list of tasks within each processor of schedule is ordered in an ascending order of their height”. 5.1 Initial Population Algorithm to generate initial population {Generates a schedule of task graph TG for multiprocessor system with p processors} 1. [Initialize] Compute height for every task in TG 2. [Separate tasks according to their height] 3. [loop p-1 times] For each of first p-1 processors, do step 4 4. [form the schedule for a processor] 5. [Last processor] Assign remaining tasks in the set to last processor. 5.2 Fitness Function Multiprocessor scheduling problem will also consider factors such as throughput, finishing time and processor utilization. Genetic algorithm is based on finishing time of a schedule. The finishing time of a schedule, S is defined as follows : FT ( S ) = max ftp( Pj ) Pj where, ftp ( Pj ) is the finishing time for the last task in processor Pj.To maximize the fitness function, one

need to convert the finishing time into maximization form. This can be done by defining the fitness value of schedule, S, as follows : Cmax - FT ( S ) Where, Cmax is the maximum finishing time observed so far. Thus the optimal schedule would be the smallest finishing time and a fitness value larger than the other schedules. 5.3 Genetic Operators Function of genetic operators is to create new search nodes based on the current population of search nodes. By combining good structures of two search nodes, it may result in an even better one. For multiprocessing scheduling problem, the genetic operators used must enforce the intra processor precedence relations, as well as completeness and uniqueness of the tasks in the schedule. For multiprocessor scheduling, certain portions of the schedule may belong to the optimal schedule. By combining several of these optimal parts, one can find the optimal schedule efficiently. For multiprocessing scheduling problem, the genetic operators used must enforce the intraprocessor precedence relations, as well as completeness and uniqueness of the tasks in the schedule. 5.3.1 Crossover The new strings can be created by exchanging portions of two strings using following method: 1. Select sites which differ in height where the lists can be cut into two halves 2. Exchange bottom halves of P1 in string A and string B 3. Exchange bottom halves of P2 in string A and string B. For multiprocessor scheduling, one should ensure that the precedence relation is not violated and that the completeness and uniqueness of tasks still holds after crossover 5.3.2 Reproduction Reproduction process forms a new population of strings by selecting string in the old population based on their fitness values. The selection criterion is that the strings with higher fitness value should have higher chance of surviving to next generation. Good strings have high fitness value and hence should be preserved in the next generation. 5.3.3 Mutation For multiprocessor scheduling problem, mutation is applied by randomly exchanging two tasks with same height. 5.4 Algorithm using GA //Algorithm Find-Schedule

4

96

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

1. [Initialize] 2. Repeat steps 3 to 8 until algorithm is convergent 3. Compute fitness values for each string in the initial population 4. Perform Reproduction. Store string with highest fitness values in BEST_STRING. 5. Perform crossover 6. Perform mutation 7. Preserve the best string in BEST_STRING VI EXPERIMENTAL RESULTS The genetic algorithm has been implemented and tested. The following are the assumptions and conditions under which the experiment is conducted. Assumptions about the task number range from 8 to 110. The number of successors that each task node is allowed is a random number between 3 and 6.The execution time for each task random number between 1 and 25. The task graphs are tested on a listscheduling algorithm. The genetic algorithm used the following parameters throughout the simulation.  Population size = 20.  Maximum number of iterations = 500. The simulation is performed using MATLAB. 6.1 Comparison between GA and LSH This section compares the list scheduling heuristic (LSH) with the genetic algorithm (GA). List scheduling is taken because the tasks are arranged as per the precedence relations. The proposed method using GA also takes the tasks in their precedence relation sequence.
Execution time Comparison 250 200 Execution time 150 100 50 0 0 50 100 Number of tasks 150 LSH GA

Table 1: Execution time GA and LSH Number of tasks ISH Execution Time GA Execution Time

8 11 10 17 24 22 23 28 26 28 47 43 39 60 46 44 56 50 49 103 55 54 68 68 59 121 83 69 139 104 79 153 137 89 157 141 100 222 158 One can infer from the table 1 the finish time of all the range tasks are lesser in the case of GA when compared with LSH. When the number of tasks is increased, GA only gives the minimum finish time. Table 2: Execution time of algorithm Number GA time of tasks to schedule task 8 3.261 17 28 33 39 44 49 54 59 69 79 89 99 100 3.249 2.593 3.125 2.688 2.625 2.859 3.537 0.885 0.842 3.186 3.155 3.217 2.311

Fig 3 Task Vs Execution Time From the fig 3 the LSH and GA produces almost same scheduling time when the number tasks in range 8 to 28.When the number of tasks increase GA gives the better solution. When the tasks are more than 100, the GA gives the best solution.

5

97

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

The table 2 indicates the time taken by the GA algorithm to compute the scheduling time for tasks among many number of processors.
Execution time

No.of Processor Vs.Exec.Time
250 200 150 100 50 0 0 1 2 3 4 5 Number of Processors Best minimum time GA LSH minimum time

Time to execute algorithm 4 3.5 3 2.5 2 1.5 1 0.5 0

Time taken to Algorith execute m

Time to execute Algorithm

0

50 100 Number of tasks

150

Fig 5: Number of processor Vs. GA and LSH Table 4: Precedence relation Vs. Execution time Height Best LSH minimum minimum time GA time 3 5 5 5 6 7 10 22 46 50 158 174 11 24 60 56 222 222

Fig 4: Execution time of algorithm From the fig 4 one can infer the time required to execute the algorithm is minimum when the number of tasks range between 55 and 70.Hence the tasks in this range GA finds the maximum fitness within the minimum period of time. Table 3 Number of processor Vs. GA and LSH No of processor Best minimum time GA LSH minimum time

Precedence relation Vs.Execution time
250 Execution time 200 150 100 50 0 0 2 4 6 8 Height of the task Best minimum time GA LSH minimum time

2 22 24 3 26 28 3 48 61 4 46 60 4 55 103 4 158 222 4 174 222 When the number of processors is increased the LSH takes more time to find the schedule and also the processors are not utilized to the maximum limit. GA gives best solution and the processor performance is increased and the processors are utilized to their maximum limit. The time slot of any processor is not wasted.

Fig 6: Precedence relation Vs. Execution time

6

98

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

One can infer from the table 4 and fig 6 the LSH and GA are producing the same result between the height 3 to 5.But when the height are more than 5 only GA produces the best result. Table 5 Comparisons between LSH and GA Parameter LSH GA Computation time Minimum Higher (task) Communication time Minimum Higher Make span Cost Scheduling time Execution time when Problem size increases Higher Minimum Higher Takes longer time to solve Minimum Optimum Minimum Minimum

even when one or more processors are significantly slower than normal. VII CONCLUSIONS The problem of scheduling of tasks to be executed on a multiprocessor system is one of the most challenging problems in parallel computing. Genetic algorithms are well adapted to multiprocessor scheduling problems. As the resources are increased available to the GA, it is able to find better solutions. The trade off for the increased resources used by the GA is a significantly longer execution time than traditional methods. Overall, the GA appears to be the most flexible algorithm for problems using heterogeneous processors. It also indicates that the GA is able to adapt automatically to changes in the problem to be solved. The advantages of the GA approach presented here are that it is simple to use, requires minimal problem specific information, and is able to effectively adapt in dynamically changing environments. The genetic algorithm with the combination of other scheduling technique may be applied to solve the multiprocessor scheduling problem. It may give even better performance and processor utilization. It may overcome the larger execution time of the algorithm. REFERENCES
1.Joel M Crichlow,”An introduction to distributed and parallel computing” PHI, 2001. 2.Harry F Jordon Gita alaghbad , “ Fundamentals of parallel processing”, PHI,2003. 3. M.Sasikumar, Dinesh Shikhare, P.Ravi prakash, “ Introduction to parallel processing”, PHI, 2003. 4. David.E.Goldberg, “ Genetic Algorithms in search, optimization & machine learning”, Pearson education,2004. 5. Pinaki Mazumdu Elizabeth, M.Pudnick, “ Genetic Algorithms”, Pearson education 2004. 6. Ben Kao and Hector Garcia Molina, December 1997, Deadline assignment in a distributed soft real time system, , IEEE transactions on parallel and distributed systems, Vol. 8, No.12, pp 1268-1274. 7. Andrei Radulescu and Arjan J.C, june 2002, Low cost task scheduling for distributed memory machines, , IEEE transactions on parallel and distributed systems, Vol. 13, No.6, pp 648-657 8. Yair Wiseman and Dror G.Feitelson , June 2003,.Paired Gang Scheduling, , IEEE transactions on parallel and distributed systems, Vol. 14, No.6, pp 581-592 9.Dakai Zhu, Rami Melhem and Bruce R.Childers , July 2005 , Scheduling with dynamic voltage/speed adjustment using slack reclamation in multiprocessor real time systems,

The table 5 shows the comparison of the parameters between GA and LSH. Table 6 gives population vs. GA. In only 20 generations, the GA finds an optimal solution for all the tasks. A suitable solution for 54 tasks is found in less number of generations. Table 6 Population Vs. GA Number Population optimal of tasks Schedule

GA

OSGA/GA * 100

8 20 11 10 10.0 17 20 24 22 9.1 23 20 28 26 7.7 44 20 56 50 12.0 49 20 103 55 87.3 54 20 68 68 0.0 69 20 139 104 33.7 79 20 153 137 11.7 89 20 157 141 11.3 99 20 211 158 33.5 All the results show that Genetic algorithm is better than List Scheduling. The advantages of the GA are simple to use, requires minimal problem specific information, and is able to effectively adapt in dynamically changing environments. It also indicates that the GA is able to adapt automatically to changes in the problem to be solved. Although performance decreases after a target change, the GA immediate begins to improve solutions and is ultimately able to find near optimal solutions

7

99

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

IEEE transactions on parallel and distributed systems, Vol. 14, No.7, pp 686-699. 10.Rabi N.Mahapatra and Wei Zhao, July 2005, An energy efficient slack distribution technique for multimode distributed real time embedded systems, IEEE transactions on parallel and distributed systems, Vol. 16, No.7, pp 650662. 11. Theodore P.Baker, August 2005, An analysis of EDF schedulability on a multiprocessor, IEEE transactions on parallel and distributed systems, Vol. 16, No.8, pp 760-768. 12. Eitan Frachtenberg, Fabrizio Petrini, November 2005 , Adaptive parallel job scheduling with flexible coscheduling , IEEE transactions on parallel and distributed systems, Vol. 16, No.11, pp 1066-1077.

Biographical Centre- Cambridge, England, The Da Vinci Diamond for Inspirational Accomplishment by International Biographical Centre, Cambridge .

Authors Profile

The author is a doctorate holder in Computer science with 21 years of experience in the academic side and approximately 2 years of experience in the industrial sector. She is the Professor and Head of the Department of Computer Science in Avinashilingam University for Women, Coimbatore-43. She has 80 publications at national and International level and executing funded projects worth 2 crores from UGC,AICTE and DRDO-NRB, DRDO-ARMREB. She is a life member of many professional organizations like CSI, ISTE, ISCA, WSEAS,AACE. Her areas of interest include network security, real time communication and real time operating systems. Her biography has been profiled at World’s Who’s Who in Science and Engineering Book, International Biographical Centre- Cambridge, England’s - Outstanding Scientist Worldwide for 2007, International Educator of the Year 2007 by IBC, 21st Century award for Achievement by International Biographical Centre- Cambridge, England, The International president’s award for Iconic achievement, by International

S.R.Vijayalakshmi is a Lecturer in School of Information Technology and Science, Dr.G.R.D college of science, Coimbatore. She received her B.Sc M.Sc, M.Phil in Electronics from the Bharathiar University and also received M.Sc in Computer Science from Bharathiar University and M.Phil in computer Science from Avinashilingam University for women. She has 14 years of teaching experience in the computer science and electronics field. Her research interests include embedded systems, parallel and distributed systems, real time systems, real time operating systems and microprocessors.

8

100

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Measurement of Nuchal Translucency Thickness for Detection of Chromosomal Abnormalities using First Trimester Ultrasound Fetal Images
S. Nirmala Center for Advanced Research, Muthayammal Engineering College,Rasipuram . V. Palanisamy Info Institute of Engineering, Kovilpalayam, Coimbatore – 641 107. .
of fluid in the nuchal region ,the nuchal translucency, can be shown in the ultrasound images for all fetuses during the first trimester[4].Even in presence of a normal karyotype, a bigger NT thickness is also associated with structural defects and genetic syndromes[5]. Furthermore ,an increased NT thickness(>2.5mm) between 10 and 14 weeks has also been associated with an increased risk of congenital heart and genetic syndrome[6],[7],[8].

Abstract—The Nuchal Translucency thickness measurement is made to identify the Down Syndrome in screening first trimester fetus and presented in this paper. The mean shift analysis and canny operators are utilized for segmenting the nuchal translucency region and the exact thickness has been estimated using Blob analysis. It is observed from the results that the fetus in the 14th week of Gestation is expected to have a nuchal translucency thickness of 1.87±0.25mm. Keywords- Down syndrome, Nuchal translucency thickness, Mean Shift Analysis and Blob analysis.

I.

INTRODUCTION

In the recent past, the non invasive prenatal diagnoses of chromosomal disorders have been focused by researchers for detecting Down Syndrome (DS) fetuses. Down syndrome or Trisomy 21 is recognized as severe and common chromosomal abnormality occurring approximately once in every 800 to 1000 live births and the risk increase with the maternal age. It is found from the literature that the Down syndrome is a genetic condition most commonly caused by the extra number 21 chromosome. Affected babies are likely to suffer from severe mental disability and have a high chance of physical disabilities, affecting in particular the heart, gastrointestinal tract, eyes and ears. Individuals with DS have a distinct craniofacial phenotype with a prominent forehead, small overall size of the craniofacial complex and underdevelopment of the fronto-naso maxillary region with missing or small nasal bones. It is also reported that Down syndrome causes Alzheimer’s disease and a 15 to 20 times higher risk of leukimea. The literatures reveal that about 20% die by the age of five years due to cardiac problems [1],[2]. Based on the observations the nuchal translucency (NT) features in first trimester fetal images has been considered to be an important parameter for the detection of DS. Measurement of NT thickness has proved to be one of the most discriminating prenatal markers in screening for chromosomal abnormalities such as tirsomies 13,18 and 21[3].Accumulation

Piotr Sieroszewski et al. [9] have evaluated the NT thickness sonographically and concluded that estimation of the same increased the predictive value of DS. Mehmet Tunc Canda and Namik Demi demonstrated with sonographic studies that increased NT thickness alone can detect 75% of fetuses with trisomy 21 [10] .The reported measurements so far have been made with the decision of skilled sonographers. As the NT thickness is of few millimeters, a small variation in the measurement made by the sonographer may lead to wrong diagnosis. The computer aided evaluation is expected to enhance the NT thickness measurement. The computer aided measurement overcomes the problems inherent in manual measurements. Bernardino et al[11] proposed a semi automatic measurement system, which used the sobel operator to detect the border of NT. However well known edge detection techniques determine the location of an edge by local evaluation of a single image feature such as intensity or the intensity gradient. But no single image feature can provide reliable border measurement in fetal US images. Yu-Bu-lee presented a semi-automated detection procedure for measuring NT thickness based on Dynamic Programming (DP) [12] improved by a nonlinear anisotropic diffusion filtering which has involved preprocessing an image, defining a region of interest in order to reduce interference from the image boundary and to adapt to different fetal head positions and sizes of NT layer. The limitation of the proposed method is that it can be only applied when fetal position is a horizontal as possible.

101

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Keeping the above facts, in the paper a semi-automatic computer aided algorithm to measure the Nuchal Translucency thickness for detecting the DS fetuses is described. The results obtained are promising and will support the physicians for better diagnosis in clinical pathology.

calipers is shown in Figure 3a and Figure 3b illustrates correct and incorrect placement of the calipers on the anatomical structures of the fetus. [13]

II. METHODS The Block diagram of the proposed image processing system is shown in Fig1. The various process involved in NT measurement is provided in this section. The fetus image is obtained from the ultrasound system and subjected to preprocessing for the extraction of features.

Figure. 2. Echo zones and NT borders

Figure. 3. Using calipers to measure NT thickness.

Figure. 1. Proposed Image Processing System.

a) Ultrasonographic sagittal view of a fetus with calipers; b) Placement of calipers on anatomical structures of thefetus C. Median Despeckling Filter The Speckle noise commonly affects all coherent imaging systems and hence the images generated using Ultrasound technique appears to be much inferior to other imaging modalities [14]. Hence speckle filtering in diagnostic ultrasound provides the experts the enhanced diagnostic visibility. The speckle detection can be implemented with a median filter having a sliding window Sw x Sw. If Sw x Sw is a window centered about P (i, j), then the filter coefficient of the median filter is given by

A. Image acquisition The ultrasound fetus images are recorded by the ultrasound machine (model EMP 1100). The result is an exceptional precise beam providing enhancements in focus accuracy, spatial and contrast resolution. The probe used is a multifrequency probe of range 5-10 MHz. A perfect midsagittal view of the fetal profile is obtained. The probe is moved from side to side so that the inner edges of the two thin echogenic lines that border the NT layer is obtained. The magnification of the image should be such that the head and thorax region occupy a major portion of the image in the neutral position. The ultrasound images are obtained as the sequence of moving pictures. Still frame which is suitable for the proposed work is chosen. B. Image Characteristics Nuchal translucency refers to the normal subcutaneous fluid-filled space between the back of the neck of a fetus and the overlying skin. Figure.2 shows a representative image of the NT and a schematic illustration of the echo zones (Z1-Z5) and the borders (B1-B2) of the NT layer. The NT thickness is defined as the maximum thickness of the translucent space (Z3) between the skin (Z4) and the soft tissues (Z2) overlying the cervical spine in the sagittal section of the fetus. Manual measurements of the NT layer are made by placing the crossbar of on-screen calipers on inner edges of the two thin echogenic lines that border the NT layer. Measurement of the maximum thickness should be made with the calipers perpendicular to the long axis of the fetus .An ultrasonographic sagittal view of a fetus and the

mi

n −1

= Med Pj

{

n −1

, j ∈ψ i

Sw

}

(1)

Where P is the pixel to be analyzed and the neighbours are

ψ i Sw = Pi j +1 , Pi+1 j +1 , Pi+1 j , Pi j −1 , Pi−1 j , Pi −1 j +1 , Pi−1 j −1 , Pi+1 j −1

{

}

(2)

The speckle detected sub image sequence ‘Bi’ is generated using the expression given below.

Bi

n −1

= Bi

{

n −1

, Γ < Td }
n −1 Sw

(3)

Where

Γ = mi

−ψ i

(4)

And Td is the threshold value. If the value of Bj exceeds the
102 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

threshold value, the co-efficient of second sub-image set or the decision making set is set to 1 or else it is set to 0. The second set is utilized for noise filtering process. Similar to the speckle detecting scheme, the noise filtering procedure also forms two sequences of sub-images. The first set of sub images can be denoted as {{Ni (0)}, {Ni (1)}, {Ni (2)}, {Ni (3)}….. {Ni (n)}}, is produced by low pass filtering at each Iteration levels. The second subset or the de-speckling subset, denoted by {{Si (0)}, {Si (1)}, {Si (2)}, {Si (3)}….. {Si(n)}}, is a binary set similar to that of the speckle detection scheme, where the good pixels are denoted by ‘1’ and speckle corrupted pixels are marked as’0’. The procedure continues till all flag values in the binary subimage sequence, {Bi (n)} is set to ‘1’. When all gray levels in {Bi (n)} set to ‘1’, after the nth iteration, {Ii (n)} is the enhanced filter output. D. Region of Interest Fetal ultrasound images are expected to have wide variation because of the fetal movement during the scanning process. Therefore it is necessary to define a ROI which can compensate for changes in the fetal head position and NT region. As the main objective is to measure the thickness of NT, a suitable shape must be chosen for better diagnosis. E. Segmentation of NT Region The segmentation of US images is an essential component of computer assisted diagnosis system. The purpose of such systems is always to detect the boundaries of different organs from the diagnostic US images. Many segmentation methods have been proposed for medical imaging. The Mean shift algorithm has been found to be efficient in the segmentation of medical images due to its inherent advantages such as less noise, simple and provides better feature analysis [15, 16]. In the proposed work, the Mean shift analysis has been utilized for segmenting the nasal bone and the frontal bone. Any segment with pixels that share similar features will group together and may form clusters with a densely populated center in the feature space. The shift procedure detects clusters by determining modes in a parametric density function iteratively. The feature vector consists of the pixel position and gray value. Let {Xi} i=1,2,.. n be an arbitrary set of n points in the d dimensional Euclidian space Rd. The multivariate kernel density function for a set of points of {Xi} i=1…n is given by:

estimates computed. The gradient estimate f'(x) is obtained by substituting a differentiable kernel into f'(x).
∆f ' E ( X ) = 1 d +2 ∑(Xi − X ) n( h d cd ) h 2 X i ∈S h ( x )
nx d +2 1 [ n ( h d cd ) h 2 n x

=

∑(X
X i ∈S h ( x )

i

− X )] − −(7)

Sh(x) is a hyper sphere with volume hdcd centered at X. nx is the total number of data points contained inside the hyper sphere. The following equation is the Mean Shift Vector.

1 ∑ ( X i − X ) − − (8 ) nx X i ∈S h ( x) Convergence: The image format is converted into spatial feature space. An arbitrary number of hyper spheres are defined with centers {X1, X2, X3, XC} within the sample set. The locations of the center of the hyper spheres are updated iteratively according to: M
h ,U ( X )

=

X f (n + 1) = X fn + M fh,U ( x )
where the f subscript is the sphere number. Algorithm

(9)

Let {xj} j=1…n be the original data points, {zj} j=1….n the points of convergence, and {Lj} j=1…n a set of labels. i.e. scalars. 1. For each j=1…n run the mean shift procedure for xj and store the convergence point in zj. 2. Identify clusters {Cp} p=1…m of convergence points by linking together all zj which are closer than 0.5 from each other in the joint domain. 3. For each j=1…n assign class labels Lj= {p | zj ε Cp} to clusters. 4. If desired, eliminate regions smaller than P pixels. F. Edge Detection The segmented image is then subjected to canny edge detection for enhanced visibility of edges for further processing of the image. III. RESULTS AND DISCUSSION

f ' ( x) =

1 nh d

∑ k{ h ( x − x )}
i i =1

n

1

(5)

The dimension and h is the radius of the hyper sphere. K is the Epanechikov kernel defined as:
2  1 −1  C (d + 2)(1 − X ), if X < 1 (6) KE ( x) =  2 d 0, otherwise  This yields the lowest Mean Squared Error (MISE) in the

Experiments have been carried out on a set of sonographic images obtained from the Ultrasound machine. A total of 50 images under various weeks of gestation were considered for analysis. The images were preprocessed and the region of interest has been cropped out for further analysis .NT region has been segmented from the cropped image by applying the mean shift cluster analysis. The data at the edges has been enhanced using canny operator to improve the visibility of the data. The NT thickness has

103

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

been estimated using Blob analysis. The implementation steps were as follows. Step1: Image acquisition using Ultrasound system. Step2: Removal of Speckle noise using Median filter. Step3: Identification of NT (ROI). Step4: Applying Mean shift segmentation Algorithm for segmenting the ROI. Step5: Edge Detection of the segmented NT by Canny edge detection. Step6: Binary masking of the entire image Step7: Blob analysis for the measurement of thickness of Nuchal Translucency.

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
Full Segment

Figure. 6 (a). Segmented image of normal fetus

Figure. 6 (b). Edge detected output of normal fetus
Full Segment

Figure. 4. Sample and Filtering Image.

Figure. 6 (c). Segmented output of abnormal fetus

The sample image and the despeckled image after preprocessing is shown in Figure 4.

Figure. 6 (d). Edge detected output of abnormal fetus

Figure. 5. Selection of ROI

The region of interest has been obtained for further segmentation. Figure.5 shows different views of ROI of fetal images. The selection of ROI will lead to precise segmentation.

The segmented image and edge detected image are shown in Figure 6.The Mean shift algorithm is utilized for obtaining the segmentation of NT region. For normal and abnormal fetus images, the visibility of the edges are improved using canny operator. The boundary of the segmented NT region is shown in Figure 6(b) and 6(d) for normal and anueploid fetus respectively.It is obvious from the results that the abnormal fetus has enlarged NT thickness when compared with the normal fetus.

104

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. XXX, No. 3, 2009 7
6

NT Thickness in mm

5

4 Normal 3 Abnormal

2

1

0 11-11.6 12-12.6 13-13.6 14-14.6 Gestational Age in Weeks

Figure. 9. Plot of mean values of NT for various weeks of gestation

Figure. 7. Image of various fetus for different weeks of gestation

Figure 7(a) and 7(c) depict the normal and abnormal fetus image of different weeks of gestation. In Figure 7(b) and 7(d) the extracted NT region is superimposed on the original image and marked as red. It is made as a verification methodology for ensuring the segmentation process.

The estimated NT thickness for the total population including the normal and abnormal fetus has been graphically shown in Figure 9. It can be observed from the graphical results that the normal fetuses have an average Nuchal translucency thickness which is considerably less compared to the abnormal fetuses. It is also observed that the computer aided measurement enhances the estimation. The deviation of the Nuchal Translucency thickness from the average thickness measured for the set of samples indicates that the measurement is more reliable. For example the fetus in 14 weeks of gestation is expected to have the NT thickness of 1.87 ± 0.25 mm. IV CONCLUSION The measurements of NT for normal and abnormal fetus have been carried out. It is concluded that the computed aided measurement will provide valuable information to the physicians to take accurate decision. The results reveal that the normal fetus with gestation week of 14 must not have greater than 2.12 mm of NT thickness ACKNOWLEDGMENT The authors would like to thank Dr.M.Madheswaran, Principal, Muthayammal Engineering College for his valuable suggestions. They would like to express their gratitude to Dr. S.Suresh, Mediscan Systems, Chennai for providing the necessary images for this study.

Figure. 8. Images of fetus for different weeks of gestation with NT region Superimposed TABLE 1. NT THICKNESS IN FIRST TRIMESTER MEASURED FOR NORMAL FETUS WITH GESTATIONAL PERIOD VARYING FROM 11 TO 14 WEEKS

S.no

Gestation Weeks

Number of subjects 11 10 12 10

Average (mm)

x

Standard Deviation

Variance

σ
0.114 0.223 0.215 0.117

σ2
0.0145 0.0595 [2] 0.0337 0.0183 [3] [1]

REFERENCES
KH. Nicolaides, G. Azar, D.Byrne,C. Mansur,K. Marks . “Fetal nuchal translucency: ultrasound screening for chromosomal defects in first trimester of pregnancy”. BMJ 1992; 304: pp 867–889 KH.Nicolaides,“Nuchal translucency and other first-trimester sonographic markers of chromosomal abnormalities”, Am Journal of Obstet Gynecol 2004; 191: pp45–67. RJ.Snijders,P.Nobel,N.Sebire,A.Souka,KH.Nicolaides(1998) UK “Multi center project on assessement of risk of trisomy 21 by maternal age and fetal nuchal translucency thickness at 10-14 weeks of gestation”.lancet 352(9125):343-346. N.Zosmer,VL.Souter,CS. Chan ,IC. Huggon ,KH. Nicolaides (1999) “Early diagnosis of major cardiac defects in chromosomally normal

1 2 3 4

11/11.6 12/12.6 13/13.6 14/14.6

1.35±0.41 1.45±0.43 1.11±0.62 1.87±0.25

[4]

105

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
fetuses with increased nuchal translucency” .Br J Obstet Gynaecol 106(8):829-8233. [5] A.P.Souka,E.Krampl,S.Bakalis,V.Heath and K.H. Nicolaides” Outcome of pregnancy in chromosomally normal fetuses with increased nuchal translucency in first trimester “, Ultrasound Obstet. Gynecol.Vol.18,PP.9-17 2001. [6] J.Hyett, G.Moscosco,KH. Nicolaides (1995) “Cardiac defects in1st trimester fetus with trisomy 18”. Fetal Diag Ther 10(6):381–386 [7] AP.Souka,E.Krampl,S.Bakalis,V.Heath,KH.Nicolaides (2001)”Outcome of pregnancy in chromosomally normal fetuses withincreased nuchal translucency in the first trimester”. Ultrasound Obstet Gynecol 18(1):9– 17 [8] N.Zosmer, VL.Souter,CS. Chan , IC.Huggon ,KH.Nicolaides (1999) “Early diagnosis of major cardiac defects in chromosomally normal fetuses with increased nuchal translucency”. Br J Obstet Gynaecol 106(8):829–833 [9] Piotr Sieroszewski, Malgorzata Perenc. Elzbieta Bas-Budecka, Jacek Suzin, “Ultrasound diagnostic schema for determination of increased risk for chromosomal fetal aneuploidies in the first half of pregnancy”, Journal of Appl Genet, pp 177-185, 2006. [10] Mehmet Tunc, Namik, “Contemporary Screening in Pregnancy”, Journal of Turkish-German Gynecol Association, vol 8(3), 331338,200 [11] F.Bernadino,R. Cardoso,N. Montenegro,J. Bernardes ,J. Marques de Sa (1998) “Semiautomated ultrasonographic measurement of fetal nuchal translucency using a computer software tool”. Ultrasound Med Biol 24(1):51–54 [12] Yu-Bu Lee Æ Min-Jeong Kim Æ Myoung-Hee Kim “Robust border enhancement and detection for measurement of fetal nuchal translucency in ultrasound images”. Med Bio Eng Comput (2007) 45:1143–1152. [13] KH.Nicolaides, N.Sebire, R.Snijders (1999) “The 11–14 week scan: the diagnosis of fetal abnormalities”. Parthenon Publishing, New York [14] Edund Hui Ng, “Speckle Noise Reduction Via Homomorphic Elliptical Threshold Rotations in complex Wavelet Domain”, A MS Thesis Presented to University of Waterloo, Canada, 2005 [15] Mark Fashing and Carlo Tomasi,”Mean Shift is a Bound Optimization “,IEEE Transaction of Pattern Analysis And Machine Intelligence,2004. AUTHORS PROFILE S.Nirmala received her B.E in Electrical and Electronics Engineering from Government College of Engineering, Salem, Madras University and Master Degree in Applied Electronics from Anna University, Chennai. She is doing her Ph.D in Biomedical Imaging at Anna University, Chennai. She is at present working as an Assistant Professor in the Department of Electronics and Communication Engineering, Muthayammal Engineering College, Rasipuram, Namakkal District, Tamilnadu. Her field of interest is Biomedical Engineering, Image Processing, Digital Signal Processing, VLSI Design and Embedded systems.She is a member of IEEE and vice chairman of IEEE India CAS chapter. Dr.V.Palanisamy received his B.E degree in Electronics and Communication Engineering from PSG College of Technology, Coimbatore and Master Degree in Communication systems from University of Madras. He also received his Ph.D in Antennas Theory from Indian Institute of Technology, Karagpur. Since 1974 he has been working in various capacities in the Department of Technical Education in Tamilnadu. He is at present working as a Principal of Info Institute of Engineering, Coimbatore. His field of interest is Electronics, Antennas, Image Processing, Communication Systems and VLSI Technologies.

106

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

An Improved Image Mining Technique For Brain Tumour Classification Using Efficient classifier
P.Rajendran
Department of Computer science and Engineering K. S. Rangasamy College of Technology, Tiruchengode-637215, Tamilnadu, India. Phone: +91 4288 274741, Fax: +91 4288 274757 .
Abstract— An improved image mining technique for brain tumor classification using pruned association rule with MARI algorithm is presented in this paper. The method proposed makes use of association rule mining technique to classify the CT scan brain images into three categories namely normal, benign and malign. It combines the low-level features extracted from images and high level knowledge from specialists. The developed algorithm can assist the physicians for efficient classification with multiple keywords per image to improve the accuracy. The experimental result on pre-diagnosed database of brain images showed 96% and 93% sensitivity and accuracy respectively. Keywords-Data mining; Image ming; Association rule mining; Medical Imaging; Medical image diagnosis;. Classification.

M.Madheswaran
Center for Advanced Research, Department of Electronics and Communication Engineering, Muthayammal Engineering College, Rasipuram – 637 408, Tamilnadu, India. Phone: +91 4287 220837, Fax: +91 4287 226537 . The first reason is that scanner images contain anatomical information which offers the possibility to plan the direction and the entry points of the radiotherapy rays which have to target the tumor and to avoid some risk organs. The second reason is that CT scan images are obtained using rays, which is the same physical principle as radiotherapy. This is very important because the radiotherapy rays intensity can be computed from the scanner image intensities. Due to the high volume of CT [3] images to be used by the physicians, the accuracy of decision making tends to decrease. This has further increased the demand to improve the automatic digital reading for decision making [4]. It also significantly improves in the field of conservative treatment of CAD diagnosis. It is an interdisciplinary field that combines techniques like data mining, digital image processing, radiology and usability among others. In this paper image mining concepts have been used. It deals with the implicit knowledge extraction, image data relationship and other patterns which are not explicitly stored in the images. This technique is an extension of data mining to image domain. It is an inter disciplinary field that combines techniques like computer vision, image processing, data mining, machine learning, data base and artificial intelligence [5]. The objective of the mining is to generate all significant patterns without prior knowledge of the patterns [6]. Rule mining has been applied to large image data bases [7]. Mining has been done based on the combined collections of images and it is associated data. The essential component in image mining is the identification of similar objects in different images [8]. The method proposed in this paper classifies the brain CT scan images into three categories: normal, benign and malignant. Normal ones are those characterizing a healthy patient, benign cases represents CT scan brain images showing a tumor that are not formed by cancerous cells, and Malign cases are those brain images that are taken from patients with cancerous tumors. CT scan brain images are among the most difficult medical images to be read due to their low contrast and differences in the type of tissues. This paper illustrates the importance of data cleaning phase in building an accurate data mining architecture for image classification [9, 10, and 11].

I.

INTRODUCTION

In health care centers and hospitals, millions of medical images have been generated daily. Analyses have been done manually with an increasing number of images. Even after analyzing a minimal number of images, radiologist becomes more tiresome. Nowadays, physicians are providing with computational techniques in assisting the diagnosis process. In the recent past, the development of Computer Aided Diagnosis (CAD) systems for assisting the physicians for making better decisions have been the area of interest [1]. This has motivated the research in creating vast amount of image database. In CAD method, computer output has been used as a second opinion for radiologist to diagnose the information more confident and quicker mechanism as compared to manual diagnosis. Pathologies are clearly identified using automated CAD system [2]. It also helps the radiologist in analyzing the digital images to bring out the possible outcomes of the diseases. In the last few years, inexpensive and available means of database containing rich medical data have been provided through the internet for health services globally. It has been reported that the brain tumor is one of the major cause leading to higher incidence of death in human. Physicians have faced a challenging task in extracting the features and decision making. The Computerized Tomography (CT) has been found to be the most reliable method for early detection of tumors because this modality is the most used in radiotherapy planning for two main reasons.

107

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Histogram Equalization

Feature Extraction

Feature Vector

Transactional Database

Association Rule Mining

TRAINING PHASE
Association Rules

---------------------------------------------------------------------------------------------------------------------

Histogram Equalization

Feature Extraction

Feature Vector

Suggestion of diagnosis

Classification

Figure 1. Overview of the proposed system

The method presented here is based on the associative classification scheme. This approach has an advantage of selecting only the most relevant features during mining process and obtaining multiple keywords when processing a test image [12]. II. SYSTEM DESCRIPTION

[16, 17]. In the test phase, the feature vector obtained from the test images are submitted to the classifier which makes use of the association rules to generate keywords to compose the diagnosis of the test image. These keywords have been used to classify the three categories of CT scan brain images as normal image, benign (tumor without cancerous tissues) image and malignant (tumor with cancerous tissues) image. A. Pre-Processing Since most of the real life data is noisy, inconsistent and incomplete, preprocessing becomes necessary [18]. The cropping operation can be performed to remove the background, and image enhancement can be done to increase the dynamic range of chosen features so that they can be detected easily.In general, most of the soft tissues have overlapping gray-levels and the condition of illumination at that time of CT scan taken is also different. The histogram equalization can be used to enhance the contrast within the soft tissue of the brain images and also hybrid median filtering

Overview of the proposed system is shown in Fig 1. The proposed system is mainly divided into two phases: the training phase and the test phase. Data cleaning and feature extraction are common for both the training set of brain images and the test set [13, 14]. In the training phase, features are extracted from the images, represented in the form of feature vectors. Next, the features are discretized into intervals and the processed feature vector is merged with the keywords related with the training images [15]. This transaction representation is submitted to the MARI (Mining Association Rule in Image database) algorithm for association rule mining, which finally produces a pruned set of rules representing the actual classifier

108

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

technique can be used to improve the image quality. Good texture feature extraction can be done by increasing the dynamic range of gray-levels using the above mentioned technique [19]. B. Texture Feature Extraction As the tissues present in brain are difficult to classify using shape or intensity level of information, the texture feature extraction is found to be very important for further classification [20,21, and 22]. The analysis and characterization of texture present in the medical images can be done using several approaches like run length encoding, fractal dimension, and discrete wavelet transform and co-occurrence matrices [23, 24]. Though many texture features have been used in the medical image classification, Spatial Gray Level Dependent Features (SGLDF) can be used to calculate the intersample distance for better diagnosis [25, 26]. In order to detect the abnormalities in medical images association rule mining is built using texture information [27, 28]. This information can be categorized by the spatial arrangement of pixel intensities. In order to capture the spatial distribution of the gray levels within the neighborhood, two dimensional co-occurrence matrices can be applied to calculate the global level features and pixel level features. The following ten descriptors can be used for extracting texture features. Entropy = − Energy = Contrast =

Cluster_Tendency = Correlation =

∑ ∑ (i + j − 2μ ) P[i, j]
k

M i

N j

(9) (10)

∑∑
i j

M

N

(i − μ )( j − μ )P[i, j]
σ2

where P is the normalized co-occurrence matrix, (i, j) is the pair of gray level intensities and M by N is the size of the co-occurrence matrix. The intersample distance is estimated based on estimation of the second-order joint conditional probability density function for the pixel (i, j), P[i, j | d, θ]

for θ = 0 o ,45 o ,90 o and135 o . The function is the probability that
two pixels which are located with an inter sample distance d and a direction θ. The estimated joint conditional probability density functions are defined as

⎧((k, l), (m, n)) ∈[Lx × Ly ]×[Lx × Ly ] :⎫ P i, j d,0o =#⎨ / T(d,0o ) k = ml − n = d,S(k, l) = i,S(m, n) = j⎬ ⎩ ⎭

[

]

⎧((k, l), (m, n)) ∈[Lx × Ly ]×[Lx × Ly ] :⎫ ⎪ ⎪ o P i, j d,45 =#⎨(k − m = d, l − n = −d)or ⎬/ T(d,45 ) (k − m = −d, l − n = d) ⎪ ⎪S(k, l) = i,S(m, n) = j ⎭ ⎩

[

o

]

∑ ∑ P[i, j]log P[i, j]
i j M i N j 2
N j

M

N

(1) (2) (3) (4)

∑∑ P [i, j]
2 ∑∑ (i − j) P[i, j] i M

, ⎧((k,l),(m n)) ∈[Lx ×Ly]×[Lx ×Ly]:⎫ ⎪ ⎪ o Pi, jd,90 =#⎨(k−m=d,l =n),S(k,l) =i, ⎬/T(d,90) ⎪S(m n) = j ⎪ ⎩ , ⎭

[

o

]

Homogeneity =

∑∑ 1 + i − j
i j

M

N

P[i, j]

⎧((k, l), (m, n)) ∈[Lx ×Ly ]×[Lx ×Ly ] :⎫ ⎪ ⎪ o o P i, jd,135 =#⎨( k − m = d, l − n = −d), ⎬/ T(d,135) ⎪ ⎪S(k, l) = i,S(m, n) = j ⎭ ⎩

[

]

SumMean=

1 M N ∑∑ (i ∗ P[i, j] + j ∗ P[i, j]) 2 i j

(5)

where # denotes the number of elements in the set, S(x, y) is the image intensity at the point (x, y) and T (d, θ) stands for the total number of pixel pairs within the image which has the intersample distance d and direction θ. Co-occurrence matrices can be calculated for the directions 0 ,45 ,90 ,135 and their respective pixels are denoted as 1, 2, 3 and 4. Once the co-occurrence matrix is calculated around each pixel, the features such as entropy, energy, variance, homogeneity and inverse variance can be obtained for each matrix with respect to the intersample distance. From the co-occurrence matrices the feature vectors can be calculated and stored in the transaction database. Next, the continuous valued features are discretized into intervals, where each interval represents an item in the process of mining association rules [15].
o o o o

M N Variance= 1 ∑∑ (i − μ)2 P[i, j] + ( j − μ)2 P[i, j] 2 i j

(

)

(6)

Maximum_Probability =

Max P[i, j]
i, j

M, N

(7)

Inverse_Diference_Moment =

∑∑ i − j
i j

M

N

P[i, j]
k

(8)

109

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

III.

BUILDING THE CLASSIFIER

The extracted features can be stored in to transaction database for classification using the trained brain images. The image mining techniques can be applied to match the extracted features with trained sets for proper classification [29]. A. Association Rule Mining Association rule mining aims at discovering the associations between items in a transactional database [8, 27, and 30]. Given a set of transactions D= {t1, t2….tn} and a set of items I= {i1, i2 ,…in} such that the transaction T in D is a set of items in I, an association rule is an implication of the form X ⇒ Y , where X is called body or antecedent of the rule, and Y is called as head or consequent of the rule [31]. The problem of mining association rules involves finding rules that satisfy minimum support and minimum confidence specified by the user. In this approach, modified version of ARC-AC algorithm [32] can be used for mining the association among the features from the transactional database. The proposed algorithm named MARI has been explained as follows. Algorithm: MARI Find association on the training set of the transactional database. Input:

(8) (9) (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) (21) (22) (23) (24) (25) (24) (25) (26) (27) (28)

kw.support ← kw.support. count (kw , p ) } }

F ← kw ∈ C kw.support 〉 σ 2 2 P2 ← Filter Table (P1 , F2 ) For (i ← 3; Fi −1 ≠ φ ; i ← i + 1 ) do { C i ← (Fi −1 F2)

{

}

C ←C − kw(i −1)Item- setof kw∉F i i i −1 Pi ← Filter Table ( Pi −1 , Fi −1 ) For each Patches p in Pi do { for each kw in C i do
{ kw.support ← kw.support + count (kw, p) } } } }

{

}

Fi ← {kw ∈ Ci kw.support> σ}

Pi : {k 1 , k 2 .........k m , f 1 , f 2 ,......f n } where k i is a keyword attached to the patches and f j are the selected features for the patches, a minimum support threshold σ .
A set of association rules of the form

A set of Image patches (P1) of the form

Rule= φ for each item set I in sets do { Rule ← Rule+{f ⇒ kw| f ∪kw∈I ∧ f Is an itemset∧ kw∈C0} }

sets ← U i { kw ∈ Fi i > 1}

Output:

f 1 ∧ f 2 ∧ ................ ∧ f n ⇒ k i where k i is a keyword, f j is a feature and kw is a
class category. Method: C0 (1) (2) (3) (4) (5) (6) (7)

The association rules are constrained such that the antecedent of the rules is composed of conjunction of features from the brain image while the consequent of the rule is always the class label to which the brain image belongs [33]. Pruning Techniques The rules generated in the mining phase are expected to be very large. This could be a problem in applications where fast responses are required. Hence, the pruning techniques become necessary to eliminate the specific rules and which are conflicting with the same characteristics pointing different categories [34, 35]. B. This can be achieved using the following conditions: Condition 1: Given two rules R1 ⇒ C and R 2 ⇒ C , the first rule is a general rule if R1 ⊆ R2. To attain this, ordering the association rules must be done as per condition 2. Condition 2: Given two rules R1 and R2, R1 is higher ranked than R2 if: (1) R1 has higher confidence than R2, (2) If the confidences are equal, support of R1 must exceed support of R2.

← {candidate keywords and their
support} ← {frequent keywords and their

F0

support} C1 ← {candidate keyword 1 item sets and their support} F1 ← {frequent 1 item sets and their support} C 2 ← {candidate pairs (k , f ) , such that For each patches p in P1 do {

(k, f ) ∈ P1 and k ∈ F0 and f ∈ F1 }
For each kw = (k , w ) in C 2 do {

110

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

(3) If both confidences and supports are equal, but R1 has less attributes in left hand side than R2. Condition 3: The rules R1 ⇒ C1 and R1 ⇒ C 2 , represents are conflict in nature. Based on the above conditions, duplicates have been eliminated. The set of rules that are selected after pruning represents the actual classifier. These conditions have been used to predict to which class the new test image belongs. C. Classification Of Test Image After the completion of training phase, an actual classifier with pruned set of association rules can be built for training the brain images [36]. Each training image is associated with a set of keywords. Keywords are representative words chosen by a specialist to use in the diagnosis of a medical image. The knowledge of specialists should also be considered during the processing of mining medical images in order to validate the results. The extracted features of the test image and the feature vector generated can be submitted to the classifier, which uses the association rules and generates set of keywords to compose the diagnosis of a test image. Algorithm: Input: Feature vector F of the test image, threshold Output: set of keywords S Method: (1) for each rule r ∈ R of the form body → head do (2) { (3) for each itemset h ∈ head do (4) { (5) if body matches F then (6) increase the number of matches by 1 (7) Else (8) increase the number of non matches by 1 (9) } (10) } // to generate keywords (11) for each rule r∈ R of the form body → head do (12) { (13) for each item set h ∈ R head do (14) { (15) if (n(Mh )/ n(Mh) + n(Nh))) ≥ T then (16) if h ∉ S then (17) add h in S (18) } (19) } (20) return S This classifier returns the multiple classes when processing a test image. The algorithm developed has been employed to generate suggestions for diagnosis. This algorithm stores all item sets (i.e. Set of keywords) belonging to the head of the rules in a data structure. An item set h is returned in the suggested diagnosis if the condition is satisfied as the given equation

(15) n(M)/(n(M) + n(N)) ≥ T h h h where, n(Mh) is the number of matches of the item set h and n(Nh) is the number of non-matches. Threshold T is employed to limit the minimal number of matches required to return an item set in the suggested diagnosis. A match occurs when the image features satisfy the body part of the rule. D. Performance evaluation criteria The confusion matrix can be used to determine the performance of the proposed method and is shown in Fig 2. This matrix describes all possible outcomes of a prediction results in table structure. The possible outcomes of a two class prediction be represented as True positive (TP), True negative (TN), False Positive (FP) and False Negative (FN). The normal and abnormal images are correctly classified as True Positive and True Negative respectively. A False Positive is when the outcome is incorrectly classified as positive (yes) when it is a negative (no). False Positive is the False alarm in the classification process. A false negative is when the outcome is incorrectly predicted as negative when it should have been in fact positive. From the confusion matrix, the precision and recall values can be measured using the formula. Precision: It is defined as the fraction of the classified image, which is relevant to the predictions. It is represented as

Pr ecision =

TP TP + FP

(16)

Recall: It is defined as the fraction of the classified image for all the relevant predictions. It is given as

Re call =

TP TP + FN

(17)

Identified Yes No

Actual

Yes

TP

FN

No

FP

TN

Figure 2.Confusion matrix

111

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

IV.

RESULT AND DISCUSSIONS 135 90 45

An experiment has been conducted on a CT scan brain image data set based on the proposed flow diagram as shown in Fig 1. The pre-diagnosed databases prepared by physicians are considered for decision making. Fig 3(a) represents the original input image and Fig 3(b) shows the result of histogram equalization and hybrid median filtered original image, which is used to reduce the different illumination conditions and noises at the scanning phase. After preprocessing, feature extraction has been done to remove the irrelevant and redundant content of the information present in the input image [23]. Haralick cooccurrence method has been used to determine the discrimination of the tissue level variations and shown in Fig.4. In Fig 4(a) pixel 1 represents 0° , pixel 2 represents 45° , pixel 3 represents 90° and pixel 4 represents 135° from the centered pixel, at a distance value of one. Fig 4(b) shows the preprocessed CT scan brain image merged with angle representation. Fig 4(c) represents the pixel representation matrix for distance one and degree zero, similarly the pixel representation matrix have to be calculated for the remaining degrees. From the pixel representation matrix, the cooccurrence matrices are also calculated and represented in the Fig 4(d).

3 4 2 1 0

Figure 4(a). Matrix representation for center pixel and all around pixels

135

90

45

0

Fig ure 4(b). Preprocessed CT scan brain image merged with angle representation Figure 3(a). Input CT scan brain image

0 3 2 3 3

0 1 1 2 3

1 1 3 1 2

3 3 0 0 1

1 1 3 3 2

Figure 3(b). Histogram equalized image

Figure 4(c). Pixel representation matrix for Zero Degree

112

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

i /j 0 1 2 3

0 1 1 0 1

1 1 1 3 3

2 0 1 0 2

3 2 3 0 1

Image

Diagnosis Malign Assessment=5 Subtlety=5 Abnormality=1 Benign Assessment=3 Subtlety=4 Abnormality=1 Benign Assessment=3 Subtlety=3 Abnormality=1

Figure 4(d). Co-occurrence matrix for distance one and Degree zero

Fig 5 represents the flow of texture feature extraction and object segregation process. For each object co-occurrence matrix has been calculated and texture features are extracted. Four different directions 0 o , 45 o ,90 o ,135 o generate 16 cooccurrence matrices. The texture features are then calculated for each co-occurrence matrix and stored in the database. The feature vectors have calculated and the obtained vectors are stored in the transaction database from the co-occurrence matrix values.

Figure 6. Images of the dataset and their diagnoses

MARI algorithm has been applied on the transaction database which consists of the feature vectors and the diagnosis information about the training CT scan images. Fig 6 represents the images of the sample dataset and their diagnosis information. In Fig 7 the precision and recall values of the proposed method, Association Rule Mining (ARM) method and Naïve Bayesian method are plotted in the graph. It shows that the performance of proposed method is better compared to the existing methods. The effectiveness of the proposed method has been estimated using the following measures:

Input CT scan brain image

Histogram equalized image

Preprocessed image

Accuracy= (TP+TN)/ (TP+TN+FP+FN) Sensitivity= TP/ (TP+FN) Specificity= TN/ (TN+FP) where, TP, TN, FP, and FN are the number of True Positive cases (abnormal cases correctly classified), the number of True Negatives (normal cases correctly classified), the number of False Positives (normal cases classified as abnormal), and the number of False Negatives (abnormal cases classified as normal) respectively. Accuracy is the proportion of correctly diagnosed cases from the total number of cases. Sensitivity measures the ability of the proposed method to identify abnormal cases. Specificity measures the ability of the method to identify normal cases. The value of minimum confidence is set to 97% and the value of minimum support is set to 10%. The features of the test images and the association rules have been generated using the threshold value=0.001. The results show that the proposed classifier gives higher values of sensitivity, specificity and accuracy such as 96%, 90% and 93% respectively. In order to validate the obtained results, the algorithmic approach has been compared with the well known classifier, a naive bayesian classifier and associative classifier [37, 38 and 39].

Texture feature calculation and object segregation

Co-occurrence matrix are calculated for all the segregated objects
Figure 5. Texture feature extraction and object segregation

113

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
1 0.9 0.8 0.7 Precision 0.6 0.5 0.4 0.3 0.2 0.2 ARM NBC Proposed-MARI

0.3

0.4

0.5

0.6 Recall

0.7

0.8

0.9

1

Figure 7. P & R graph using naive bayesian, association rule mining and MARI association rule mining

Figure 8. Tumor classification by ROC analysis

Table 1 illustrates the sensitivity, accuracy, specificity, area under the curve (Az), standard error (SE) and execution time of naive bayesian classifier, association rule mining and proposed method. The experimental results have shown that the proposed method achieves high sensitivity (up to 96%), accuracy (up to 93%) and less execution time and standard error in the task of support decision making system.
TABLE 1 PERFORMANCE COMPARISION FOR CLASSIFIERS

Table 2 represents the results of the classifiers, here 150 images are taken for training and 95 images are taken for the testing in both benign and malignant categories, which are classified using different classifiers. The results show that the proposed system gives better percentage of correct classification as compared to naive bayesian classifier and association rule based classifier.
TABLE 2 RESULTS OF THE CLASSIFIERS

Sensitivity (%)

Specificity (%)

Accuracy(%)

No. of correctly classified data

No. of data Training/ Testing

Association rule based classifier

Association rule based classifier

Approach Pruned Association rule with MARI Algorithm based classifier

Az

SE

Pruned Association rule with MARI Algorithm based classifier

Naive Bayesian classifier

Naive Bayesian classifier

Classes

Naive Bayesian classifier [22]

75

63

70

0.89

0.08

30.91

Benign

150 / 95

Association rule based classifier [22] 84 91 93 88.4 95.78 97.90

95

84

91

0.91

0.03

9.75

Malignant

150 / 95

85

92

94

89.4

96.84

98.95

Pruned Association rule with MARI Algorithm based classifier

96

90

93

0.98

0.02

2.15

Average

89.9

96.31

98.42

114

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

Time (ms)

Percentage of correct classification

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

The Receiver Operating Characteristic (ROC) curves are plotted with respect to sensitivity and specificity. The area under the ROC plays a vital role since it has been used to determine the overall classification accuracy. Fig 8 shows the comparison of the ROC curve for various classifiers. It clearly shows that the proposed mining based classification with pruned rules has higher value of (A z ) as compared to other methods. V. CONCLUSION

An improved image mining technique for brain tumor classification using pruned association rule with MARI algorithm has been developed and the performance is evaluated. The proposed algorithm has been found to be performing well compared to the existing classifiers. The accuracy of 93% and sensitivity of 96% were found in classification of brain tumors. The developed brain tumor classification system is expected to provide valuable diagnosis techniques for the physicians.

ACKNOWLEDGEMENT The authors would like to express their gratitude to Dr. D. Elangovan, Pandima CT scan centre, Dindigul for providing the necessary images for this study. REFERENCES
X.R. Marcela, H.B. Pedro, T.J. Caetano, M.A.M. Paulo, A.R. Natalia, and J.M Agma , ”Supporting Content-Based Image Retrieval and Computer-Aided Diagnosis Systems with Association Rule-Based Techniques,” Data & Knowledge Engineering, in press. [2] G.D. Tourassi, “Journey toward Computer-Aided diagnosis role of image texture analysis,” Radiology, 1999, pp 317-320. [3] S.R. Daniela, ”Mining Knowledge in Computer Tomography Image Databases,” Multimedia Data Mining and Knowledge Discovery, Springer London, 2007. [4] B. Thomas, ”Analyzing and mining image databases. Reviews Drug Discovery Today,” Biosilico 10(11), 2005, pp.795-802. [5] C. Ordonez, E. Omiecinski ,”Image mining: A new approach for data mining,” Technical Report GITCC-98-12, Georgia Institute of Technology, College of Computing, 1998, pp 1-21. [6] H. Wynne, L.L Mong ,and J. Zhang, “Image mining: trends and developments. Journal of Intelligent Information Systems,” 19 (1): 2002, pp 7–23. [7] P. Stanchev , M. Flint, “Using Image Mining For Image Retrieval,” In. Proc: IASTED conf. Computer Science and Technology, 2003, pp. 214218. [8] C. Ordonez, E.Omiecinski, ”Discovering association rules based on image content,” In Proc: IEEE Forum ADL, 1999, pp. 38–49. [9] V. Megalooikonomou, J. For ,F. Makedon , ”Data Mining in Brain Imaging,” Statistical Methods in Medical Research, 2000, pp. 359-394. [10] H. Pan, L.Jianzhong, W.Zhang, ”Incorporating domain knowledge into medical image clustering,” Applied Mathematics and Computation, 2007, pp. 844–856. [11] J.B.W. Pluim, J.B.A. Maintz, and M.A. Viergever, ”Mutual-informationbased registration of medical images,” A survey. IEEE Transactions on Medical Imaging, 22(8):2003, pp. 986-1004. [12] M.L. Antonie, O.R. Zaiane, and A.Coman, ”Associative classifiers for medical images,” Revised Papers from MDM/KDD and PAKDD/KDMCD, 2002, 68–83. [1]

[13] B.A. Dogu, H. Markus, A.Tuukka, D. Prasun, and H. Jari , ”Texture Based Classification and Segmentation of Tissues Using DT-CWT Feature Extraction Methods,” In Proc: 21st IEEE International Symposium on Computer-Based Medical Systems, 2008, pp.614-619. [14] P. Dollar, T. Zhuowen, T. Hai, and S. Belongie, “ Feature Mining for Image Classification,” In Proc: IEEE Conference on Computer Vision and Pattern Recognition, 2007, pp. 1-6. [15] J. Dougherty, R. Kohavi, and M. sahami, “Supervised and Unsupervised Discretization of continuous features,” In Proc: 12th International Conference Machine Learning, 1995, pp.56-69. [16] R. Agrawal, T. Imielinski, And A.N. Swami, ”Mining association rules between sets of items in large databases,” In Proc: ACMSIGMOD Int. Conf. Manage, Washington, DC, 1993, pp. 207-216. [17] S. Kotsiantis, D.Kanellopoulos, ”Association Rules Mining: A Recent Overview,” GESTS International Transactions on Computer Science and Engineering, 32 (1):2006, pp. 71-82. [18] R. Agrawal, R.Srikant, ”Fast algorithms for mining association rules,” In Proc: Int. Conf. VLDB, Santiago, Chile, 1994, pp. 487-499. [19] J.C. Felipe, A.J.M Traina, and C. Traina, ”Retrieval by content of medical images using texture for tissue identification,” In Proc: 16th IEEE Symp. Computer-Based Med. Systems. CBMS 2003, pp. 175–180. [20] R. Abraham, J.B. Simha,and S.S Iyengar, ”Medical datamining with a new algorithm for Feature Selection and Naive Bayesian classifier.,” In Proc: 10th International Conference on Information Technology (ICIT), 2007, pp. 44-49. [21] C. Christophe, S.G. Jean, L.M. Gael, and K. Michel, ”Efficient Data Structures and Parallel Algorithms for Association Rules Discovery,” In Proc: Fifth Mexican International Conference in Computer Science (ENC), 2004, pp. 399-406. [22] P.G. Foschi, D.Kolippakkam, H. Liu, and A. Mandvikar , ” Feature extraction for image mining,” In Proc: 8th Int. Workshop Multimedia Inf. Syst, Tempe, AZ, 2002, pp. 103-109. [23] R.M. Haralick, K.Shanmugam, and I. Distein, ”Textural features for image classification.,” IEEE Trans. Syst, Man, Cybern, vol. SMC-3, 1973, pp. 610–621. [24] A. Ranjit, B.S. Jay, and S.S. Iyengar, ”Medical Data mining with a New Algorithm for Feature Selection and Naive Bayesian Classifier,” In Proc: 10th International Conference on Information Technology (ICIT), 2007, pp.44-49. [25] L. Hui, W. Hanhu, C. Mei, and W. Ten , ”Clustering Ensemble Technique Applied in the Discovery and Diagnosis of Brain Lesions,” In Proc: Sixth International Conference on Intelligent Systems Design and Applications (ISDA) , vol. 2: 2006, pp. 512-520. [26] C.F. Joaquim, X.R. Marcela, P.M.S. Elaine, J.M.T. Agma, and T.J. Caetano, ”Effective shape-based retrieval and classification of mammograms,” In Proc: ACM symposium on Applied computing, 2006, pp. 250 – 255. [27] N.R. Mudigonda, R.M. Rangayyan , ”Detection of breast masses in mammograms by density slicing and texture flow-field analysis,” IEEE Trans. Med. Imag. 20(12), 2001, pp. 1215–1227. [28] K. Murat, I.M. Cevdet,”An expert system for detection of breast cancer based on association rules and neural network,” An International Journal Expert Systems with Applications 36: 2009, pp. 3465–3469. [29] K. Lukasz, W. Krzysztof, ”Image Classification with Customized Associative Classifiers,” In Proc: International Multiconference on Computer Science and Information Technology, 2006, pp. 85–91. [30] E. Laila, A.A. Walid, ”Mining Medical Databases using Proposed Incremental Association Rules Algorithm (PIA),” In Proc: IEEE Second International Conference on the Digital Society, 2008, pp 88-92. [31] A. Olukunle, S.A. Ehikioya, ”A Fast Algorithm for Mining Association Rules in Medical Image Data,” In Proc: IEEE Canadian Conf. Electr. Comput. Eng. Conf, 2002, pp. 1181–1187. [32] L.A. Maria, R.Z. Osmar, and C. Alexandru, ”Associative Classifiers for Medical Images.Mining,” Lecture Notes in Computer Science, Multimedia and Complex Data, Springer Berlin / Heidelberg, 2003, pp.68-83.

115

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
[33] X.Wang, M.Smith, and R. Rangayyan, “ Mammographic information analysis through association-rule mining,” In Proc: IEEE CCGEI, 2004, pp. 1495-1498. [34] B. Liu, W. Hsu, Y. Ma, ” Pruning and Summarizing the Discovered Associations,” In Proc: ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , 1999, pp. 81-105. [35] R.Z. Osmar, L.A. Maria,” On Pruning and Tuning Rules for Associative Classifiers,” In Proc: 9th International Conference(KES) ,K-BasedIntelligent Information and Engineering Systems Melbourne, part III, 2005. [36] S. T. Vincent, W. Ming-Hsiang, and S.J. Hwung, “A New Method for Image Classification by Using Multilevel Association Rules,” In Proc: 21st International Conference on Data Engineering Workshops (ICDEW), 2005, pp.1180-1188. [37] G.H. John, P. Langley, ”Estimating Continuous Distributions in Bayesian Classifiers,” In Proc: 11th conference on uncertainty in artificial intelligence Morgan Kaufmann, 1995, pp. 338-345. [38] J. Kazmierska, J. Malicki, "Apllication of the Navie Bayeian classifier to optimize treatment decisions,” Journal of Radiotherapy and Oncology, 86(2), 2008, pp. 211-216. [39] X.R. Marcela, J.M.T. Agma, T.Caetano, M.A.M. Paulo, ”An Association Rule-Based Method to Support Medical Image Diagnosis With Efficiency,” IEEE transactions on multimedia, 10( 2):2008, pp. 277-285. AUTHORS PROFILE P.Rajendran obtained his MCA degree from Bharathidhasan University in 2000, ME Degree in Computer science and engineering from Anna University, Chennai, in 2005. He has started his teaching profession in the year 2000 in Vinayakamissions engineering college, salem. At present he is an Assistant Professor in department of computer science and engineering in K.S.Rangasamy college of Technology, Thiruchengode. . He has published 10 research papers in International and National Journals as well as conferences. He is a part time Ph.D research scalar in Anna University Chennai. His areas of interest are Data mining, Image mining and Image processing. He is a life member of ISTE. Dr. M. Madheswaran has obtained his Ph.D. degree in Electronics Engineering from Institute of Technology, Banaras Hindu University, Varanasi in 1999 and M.E degree in Microwave Engineering from Birla Institute of Technology, Ranchi, India. He has started his teaching profession in the year 1991 to serve his parent Institution Mohd. Sathak Engineering College, Kilakarai where he obtained his Bachelor Degree in ECE. He has served KSR college of Technology from 1999 to 2001 and PSNA College of Engineering and Technology, Dindigul from 2001 to 2006. He has been awarded Young Scientist Fellowship by the Tamil Nadu State Council for Science and Technology and Senior Research Fellowship by Council for Scientific and Industrial Research, New Delhi in the year 1994 and 1996 respectively. His research project entitled “Analysis and simulation of OEIC receivers for tera optical networks” has been funded by the SERC Division, Department of Science and Technology, Ministry of Information Technology under the Fast track proposal for Young Scientist in 2004. He has published 120 research papers in International and National Journals as well as conferences. He has been the IEEE student branch counselor at Mohamed Sathak Engineering College, Kilakarai during 1993-1998 and PSNA College of Engineering and Technology, Dindigul during 2003-2006. He has been awarded Best Citizen of India award in the year 2005 and his name is included in the Marquis Who's Who in Science and Engineering, 2006-2007 which distinguishes him as one of the leading professionals in the world. His field of interest includes semiconductor devices, microwave electronics, optoelectronics and signal processing. He is a member of IEEE, SPIE, IETE, ISTE, VLSI Society of India and Institution of Engineers (India).

116

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Mining Spatial Gene Expression Data Using Negative Association Rules
*M.Anandhavalli, M.K.Ghose
Department of Computer Science Engineering SMIT Majitar, India .
Abstract— Over the years, data mining has attracted most of the attention from the research community. The researchers attempt to develop faster, more scalable algorithms to navigate over the ever increasing volumes of spatial gene expression data in search of meaningful patterns. Association rules are a data mining technique that tries to identify intrinsic patterns in spatial gene expression data. It has been widely used in different applications, a lot of algorithms introduced to discover these rules. However Priori-like algorithms has been used to find positive association rules. In contrast to positive rules, negative rules encapsulate relationship between the occurrences of one set of items with absence of the other set of items. In this paper, an algorithm for mining negative association rules from spatial gene expression data is introduced. The algorithm intends to discover the negative association rules which are complementary to the association rules often generated by Priori like algorithm. Our study shows that negative association rules can be discovered efficiently from spatial gene expression data. Keywords- Spatial Gene expression data; Association Rule; Negative Association Rule

K.Gauthaman
Department of Drug Technology Higher Institute of Medical Technology Derna, Libya . themselves but rather based on the co-occurrence of the items within the database. The associations between items are commonly expressed in the form of association rules. In general, an association rule represents a relationship between two sets of items in the same database. It can be written in the form A → C, where A and C are item sets and A∩C=Φ. The left-hand side (LHS) of the rule is called the antecedent, while the right-hand (RHS) is called consequent. Negative association rules are complementary to the sorts of association rules and have the forms A→¬C or ¬A→C. The rule in the form of ¬A →¬C is equivalent to a positive association rule in the form of C →A. In this paper, an attempt has been made to study a novel algorithm for discovering of positive association rules and generating meaningful negative association rules in effective manner from spatial gene expression data. II. → means ―implies‖ III. NOTATIONS U means Union ¬ means negation

I.

INTRODUCTION (HEADING 1)

MATERIALS AND METHODS

The main contribution here has been a great explosion of genomic data in recent years. This is due to the advances in various high-throughput biotechnologies such as spatial gene expression database. These large genomic data sets are information-rich and often contain much more information than the researchers who generated the data might have anticipated. Such an enormous data volume enables new types of analyses, but also makes it difficult to answer research questions using traditional methods. Analysis of these massive genomic data has two important goals: 1) To determine how the expression of any particular gene might affect the expression of other genes 2) To determine what genes are expressed and not expressed as a result of certain cellular conditions, e.g. what genes are expressed in diseased cells that are not expressed in healthy cells? The most popular pattern discovery method in data mining is association rule mining. Association rule mining was introduced by [4]. It aims to extract interesting correlations, frequent patterns, associations or casual structures among sets of items in transaction databases or other data repositories. The relationships are not based on inherent properties of the data
This study has been carried out as part of Research Promotion Scheme (RPS) Project under AICTE, Govt. of India. * Corresponding author
117

A. Spatial Gene Expression Data The Edinburgh Mouse Atlas gene expression database (EMAGE) is being developed as part of the Mouse Gene Expression Information Resource (MGEIR) [1] in collaboration with the Jackson Laboratory, USA. EMAGE (http: //genex.hgu. mrc.ac.uk/Emage/database) is a freely available, curated database of gene expression patterns generated by in situ techniques in the developing mouse embryo [9, 10]. The spatial gene expression data are presented as N×N similarity matrix. Each element in the matrix is a measure of similarity between the corresponding probe pattern and gene-expression region. The similarity is calculated as a fraction of overlap between the two and the total of both areas of the images. This measurement is intuitive, and commonly referred to as the Jaccard index [2, 6]. When a pattern is compared to itself, the Jaccard value is 1 because the two input spatial regions are identical. When it is compared to another pattern, the Jaccard Index will be less than one. If the Jaccard Index is 0, the two patterns do not intersect. If a Jaccard Index value is close to 1, then the two patterns are more similar. However, biologists are more interested in how gene expression changes under different probe patterns. Thus, these

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

similarity values are discretized such that similarity measure greater than some predetermined thresholds and converted into Boolean matrix.

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

frequent itemsets these two subsets are also frequent and their support is already calculated. Now these two subsets may generate a rule A →C, if the confidence of the rule is greater than or equal to the specified minimum confidence.

B. Data Preprocessing Preprocessing is often required before applying any data mining algorithms to improve performance of the results. The preprocessing procedures are used to scale the data value either 0 or 1. The values contained in the spatial gene expression matrix had to be transformed into Boolean values by a socalled discretization phase. In our context, each quantitative value has given rise to the effect of discretization procedure [2]: Max minus x% method. Max minus x% procedure [7] consists of identifying the highest expression value (HV) in the data matrix, and defining a value of 1 for the expression of the gene in the given data when the expression value was above HV – x% of HV where x is an integer value. Otherwise, the expression of the gene was assigned a value of 0 (Figure 1). In the similarity matrix, the items I are genes from the data set, where a transaction T I consists of genes that all have an expression pattern intersecting with the same probe pattern. The sets of transactions are constructed by taking, for each probe pattern r, every gene g from which its associated gene expression pattern ge satisfies the minimum similarity β, i.e., similarity(r, ge) > β, to form the itemsets. a b c d e f g h α (Input) 0.096595 0.123447 0.291310 0.126024 0.155819 0.288394 0.000000 0.215049 α (after discretization) 0 0 1 0 0 1 0 1

D. Algorithm Details 1. Let I={i1, i2, …, in} be a set of items, where each item ij corresponds to a value of an attribute and is a member of some attribute domain Dh={d1, d2, …, ds}, i.e. ij Є Dh. If I is a binary attribute, then the Dom (I)={0,1}. 2. A transaction database is a database containing transactions T in the form of (d, E), where d Є Dom(D) and E Є I. Let D be a transaction database, n be the number of transactions in D, and minsup be the minimum support of D. The new_support is defined as new_support = minsup × n. Proposition 1: According to [8], By Boolean vector with AND operation, if the sum of ‗1‘ in a row vector Bi is smaller than k, it is not necessary for Bi to involve in the calculation of the k- supports. Proposition 2: According to [5], Suppose Itemsets X is a kitemsets; |FK-1(j)| presents the number of items ‗j‘ in the frequent set FK-1. There is an item j in X. If | FK-1(j)| is smaller than k-1, itemset X is not a frequent itemsets . Proposition 3: |FK| presents the number of k-itemsets in the frequent set FK. If |FK| is smaller than k+1, the maximum length frequent itemsets is k. A positive association rule represents a relationship between two sets of items in the form of A →C, where A I, C I and A∩C=Φ. A positive association rule represents a relationship between two sets of items in the form of A → ¬C or ¬A→C, where A I, C I and A∩C=Φ. The rule A → ¬C has support s% in the data sets, if s% of transactions in T contain itemset A while do not contain item set C. The support of negative association rule supp(A → ¬C, is the frequency of occurrence of transactions with item set A in the absence of item set C.

3.

4.

5.

6.

7.

8.

9.

Figure 1. Results of Max minus 25% discretization method

C. Association Rule Mining The Apriori-like algorithms adopt an iterative method to discover frequent itemsets. The process of discovering frequent itemsets need multiple passes over the data. .The algorithm starts from frequent 1-itemsets until all maximum frequent itemsets are discovered. The Apriori-like algorithms consist of two major procedures: the join procedure and the prune procedure. The join procedure combines two frequent kitemsets, which have the same (k-1)-prefix, to generate a (k+1)itemset as a new preliminary candidate. Following the join procedure, the prune procedure is used to remove from the preliminary candidate set all itemsets whose k-subset is not a frequent itemsets [3]. From every frequent itemset of k>=2, two subsets A and C, are constructed in such a way that one subset C, contains exactly one item in it and remaining k-1 items will go to the other subset A. By the downward closure properties of the

10. Let X be the set of transactions that contain all items in A. The rule A →¬C holds in given data set with confidence c%, if c% of transactions in X do not contain item set C. Confidence of negative association rule, conf(A → ¬C), can be calculated with supp(AU¬C)/supp(A). 11. Given supp(A →C) and conf(A →C), the support and confidence of the negative rule A→¬C can be computed as follows: supp(A →¬C) = supp(A) – supp(A →C)……..…(1) conf(A →¬C) = 1 – conf(A →C) ……………........(2) 12. Given supp(A→C) and conf(A→C), the support and confidence of the negative rule ¬A→C can be computed as follows: supp(¬A→C) = supp(C) – supp(C →A)………….(3)

118

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

conf(¬A→C)= (supp(C)/1-supp(A)) (1–conf(C→A))….(4) The introduced algorithm for finding both RHS negative and LHS negative rules in terms of spatial gene expression data in the form of similarity matrix consists of four phases as follows: 1. 2. 3. 4. Transforming the similarity matrix into the Boolean matrix Generating the set of frequent itemsets using Fast Mining algorithm for spatial gene expression data Generating positive rules using Apriori algorithm. Generating negative rules based on existing positive rules.

The part I of the algorithm given in Figure 2 is capable of discovering all possible set of frequent itemsets subject to a user specified minimum confidence. The part II of the algorithm given in Figure 2 is capable of finding all positive and negative association rules from the frequent itemsets subject to a user specified minimum confidence very quickly. The function genRule(FreqSetk) generates all positive subject to a user specified minimum confidence. The function genNegCand(r) generates all negative itemsets for the given positive association rule, using the support value calculated using the formula given in equations (1) and (3). From the generated negative item sets, the negative association rules are generated subject to the confidence calculated using the formula given in equations (2) and (4). IV. RESULTS AND DISCUSSION

A detailed description of the introduced algorithm is described as follows: Part 1: Algorithm for generating frequent itemsets and positive rules.
Input: Spatial Gene Expression data in similarity matrix (M), the minimum support. Output: Set of frequent itemsets F. 1. Normalize the data matrix M and transformed into Boolean Matrix B; // Frequent 1-itemset generation 2. for each column Ci of B 3. If sum(Ci) >= new_support 4. F1 = { Ii}; 5. Else delete Ci from B; // By Proposition 1 6. for each row Rj of B 7. If sum(Rj) < 2 8. Delete Rj from B; // By Proposition 2 and 3 9. for (k=2; | Fk-1| > k-1; k++) 10. { // Join procedure 11. Produce k-vectors combination for all columns of B; 12. for each k-vectors combination { Bi1, Bi2,…Bik} 13. { E= Bi1 ∩ Bi2 ∩.…∩Bik 14. If sum(E) >= new_support 15. Fk = { Ii1, Ii2,…Iik} 16. } // Prune procedure 17. for each item Ii in Fk 18. If |Fk(Ii)| < k 19. Delete the column Bi according to item Ii from B; 20. for each row Rj of B 21. If sum(Bj) < k+1 22. Delete Bj from B; 23. k=k+1 24. } 25. Return F = F1UF2….UFk

The introduced algorithm has been implemented in Java and tested on Linux platform. A comprehensive experiment on spatial gene expression data has been conducted to study the impact of normalization. A few sample records from spatial gene expression data (EMAGE) are listed in Table I. Different support-confidence thresholds were tested. A few positive and negative rules are listed in Table I. They are generated under the support and confidence constraints of 2% to 9% and 30% and 60%, respectively. Note that rules in the right column are negative rules discovered with respect to positive rules in the left column. Table II shows the number of positive rules and negative rules vs. user-specified minimum support and minimum confidence. Thus, the algorithm can successfully generate negative rules and the number of negative rules discovered is reasonable. The number of negative rules tends to be related to the number of positive rules given in Table II. However, it is inversely proportional to the minimum support threshold. The reason is less and less high-support 1-item set survives with increases in support threshold, which reduces the number of candidate negative rules significantly.
TABLE I. Spatial gene expression data from EAMGE database

Uniqid
EMAGE:1024 EMAGE:111 EMAGE:114 EMAGE: 117

Gene name

EMAGE :1024

EMAGE: 111

EMAGE: 114

EMAGE: 117

Cer1 T Mesp1 Pou5f1

1 0 0 0

0 1 1 0

0 1 1 1

0 0 1 1

Part 2: Algorithm for generating positive and negative association rules.
Input: Set of Frequent (F), minimum support and minimum confidence. Output: Set of Positive and Negative Association rules. 26. postiveRule = genRule(FreqSetk); 26. Rule = postiveRule; // Generate Negative Rules 27. for all rules r Є postiveRule. 28. negativeRuleSets = genNegCand(r); 29. for all rules tr Є negaitveRuleSets 30. Rule = {Rule, Neg(tr) | Neg(tr).supp >minsup, Neg(tr).conf > minconf }; 31. endif 32. endfor. Fig. 2 Mining Negative Association Rules

TABLE II. Positive and negative rules generated (min_support=3% and

min_confidence = 30% and 60% ) Confidence 100% 50% 67%

Positive rules T→Mesp1 T→Pou5f1 Mesp1→Pou5f1

Negative rules Cer1→ ¬ T Cer1→ ¬Mesp1 Cer1→ ¬Pou5f1

Confidence 100% 100% 100%

Given the number of positive rules P, the complexity of the algorithm is O(P). In this algorithm the complexity does not depend on the number of transactions since it is assumed that the supports of item sets have been counted and stored for use in this as well as other mining applications. However if we are considering discovering positive rules, which is necessary in

119

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

generating negative rules, the algorithm must browse all combinations of items. The complexity of discovering positive rules depends on not only the number of transactions, but also the sizes of attribute domains as well as the number of attributes. The overall complexity will be proportional to that of discovering positive rules. The performance is also affected by the choice of minimum support. A lower minimum support produces more numerous item sets and, with the same confidence constraint more positive rules will be generated, which adds to computation expense. The trend in the number of negative and the number of positive rules with different minimum support are shown in Figure 3.

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

association rule mining ensures that genomic data mining will continue to be a necessary and highly productive field for the foreseeable future. REFERENCES

Figure 3. Positive and Negative Rules Discovered under Different Supports and Confidences

V.

CONCLUSION

In this paper, a novel method of mining positive and negative association rules from the spatial gene expression data has been introduced to generate frequently occur genes very quickly. The introduced algorithm does not produce candidate itemsets, it spends less time for calculating k-supports of the itemsets with the Boolean matrix pruned, and it scans the database only once and needs less memory space when compared with Apriori algorithm. The introduced algorithm is good enough for generating positive and negative association rules from spatial gene expression data very fast and memory efficient. Finally, the large and rapidly increasing compendium of data demands data mining approaches, particularly

Baldock,R.A., Bard,J.B., Burger,A., Burton,N., Christiansen,J., Feng,G., Hill,B., Houghton,D., Kaufman,M., Rao,J. et al., ―EMAP and EMAGE: a framework for understanding spatially organized data‖, Neuroinformatics, vol. 1, pp. 309–325, 2003. [2] Pang-Ning Tan, Micahel Steinbach, Vipin Kumare, Intoduction to Data Mining Pearson Education, second edition, pp.74, 2008. [3] Agrawal, R. & Srikant, R., ―Fast Algorithms for Mining Association Rules in large databases‖. In Proceedings of the 20th International Conference on Very Large Databases pp. 487-499. Santiago, Chile, 1994. [4] Agrawal, R., Imielinski, T., & Swami, A.,‖Mining association rules between sets of items in large databases‖. Proceedings of the ACM SICMOD conference on management of data‖, Washington, D.C, 1993. [5] Xu, Z. & Zhang, S., ―An Optimization Algorithm Base on Apriori for Association Rules‖. Computer Engineering vol. 29(19), pp. 83-84, 2003. [6] J. van Hemert and R. Baldock, ―Mining spatial gene expression data for association rules‖, In S. Hochreiter and R. Wagner, editors, Proceedings of the 1st International Conference on BioInformatics Research and Development, Lecture Notes in Bioinformatics, pp.66–76. SpringerVerlag, 2007. [7] Céline Becquet et al., ―Strong-association-rule mining for large-scale gene-expression data analysis: a case study on human SAGE data‖, Genome Biology vol. 3(12):research0067.1–0067.16, November 2002. [8] M.Anandhavalli, M.K.Ghose, K.Gauthaman, ―Mining Spatial Gene Expression Data Using Association Rules‖, IJCSS vol. 3(5), 2009. [9] S. Venkataraman, P. Stevenson, Y. Yang, L. Richardson, N. Burton, T. P. Perry, P. Smith, R. A. Baldock, D. R. Davidson, and J. H. Christiansen. Emage—edinburgh mouse atlas of gene expression: 2008 update. Nucleic Acids Research, 36(D):860–865, Jan 2008. [10] J. van Hemert and R. Baldock. Matching spatial regions with combinations of interacting gene expression patterns. In M. Elloumi, J. K¨ung, M. Linial, R. Murphy, K. Schneider, and C. Toma, editors, Proceedings of the 2nd International Conference on BioInformatics Research and Development, volume 13 of Communications in Computer and Information Science, pages 347–361. Springer Verlag, 2008.

[1]

120

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Hierarchical Route Optimization by Using Tree information option in a Mobile Networks
K.K.Gautam
Department of Computer Science & Technology Roorkee Engineering & Management Technology Institute Shamli-247 774 (INDIA) .
Abstract—The Networks Mobility (NEMO) Protocol is a way of managing the mobility of an entire network, and mobile internet protocol is the basic solution for Networks Mobility. A hierarchical route optimization system for mobile network is proposed to solve management of hierarchical route optimization problems. In present paper, we study Hierarchical Route Optimization Scheme using Tree Information Option (HROSTIO). The concept of optimizationfinding the extreme of a function that maps candidate ‘solution’ to scalar values of ‘quality’ – is an extremely general and useful idea. For solving this problem, we use a few salient adaptations and we also extend HROSTIO perform routing between the mobile networks. Keywords-Route Optimization, Tree Information Option,, personal area networks, NEMO, IP.

Menu Chaudhary
Department of Computer Science & Technology Roorkee Engineering & Management Technology Institute Shamli-247 774 (INDIA) Mobile networks can here very complex form of hierarchy e.g. Mobile networks in a mobile network Visiting Mobile Nodes (VMNS) in mobile networks and so on. This situation is repaired as nested mobile network. Many important problems arising in science, industry and commerce, mobile networks fall very neatly into the read-made category of optimization problem. That is to say, these problems are solved if we can simply find a solution that maximizes or minimizes some important and measurable property. For example, in Fig. 1, we might want to find the set of mobile router in a simple illustration of nested mobile network at the beginning MR1, MR2, and Visiting. Mobile Nodes (VMNS) are attached to their own home link. After MR1 moves to a foreign link, MR2 moves this make a simplest form of nested mobile networks. II. NEMO ARCHITECTURE

I.

INTRODUCTION

In the trend of ubiquitous computing, many electric appliances and electronic devices capable of integrating with wireless communications are being added. The mobile internet protocol (IP) working group within the Internet Engineering Task Force (IETF) has proposed the mobile IP protocol [1], [2] to support host mobility in IP based networks. The mobile IP aims at maintaining internet connectivity while a host is moving. The Networks Mobility (NEMO) protocol is a way of managing the mobility of an entire network, viewed as a single unit, which changes its points to attachments in the internet [3]. Such an internet will include one or more Mobile Routers (MRs) that connect it to the global internet. A mobile network can connect it to the global internet. A mobile network can have a hierarchical structure; in this paper we propose a Hierarchical Route Optimization Scheme using Tree Information Option (HROSTIO) for mobile network. In addition to routing inefficiency, other criteria are important in designing a route optimization scheme for mobile networks. The concepts of network mobility have been introduced to reduce the signaling overheads of a number of hosts moving as group. The NEMO basic support protocol uses a bidirectional tunnel between the Home Agent (HA) and the Mobile Networks Needs (MNNS) from sending all there location registration simultaneously when the MR changes its point of attachment. The characteristic is called mobility transparency, which is a very desirable feature for the route optimization scheme.

When a mobile network moves from one place to another, it changes its points of attachment to the internet, which also makes changes to its reach ability and to the Internet topology. NEMO (Network Mobility) working group has come up with NEMO support solution. NEMO support is a mechanism that maintains the continuity of session between Mobile Networks Node (MNN) and their Correspondent Nodes (CN) upon a mobile Router’s change of point attachment. NEMO support is divided into two parts: 1). NEMO Basic Support 2). NEMO Extended Support NEMO Basic Support is a solution for persevering session continuity by means of bidirectional tunneling between Home Agent (HA) and a mobile network. And NEMO extended Support is a solution for providing the necessary optimization between arbitrary Mobile Networks Nodes and correspondent Nodes, including routing optimization [5]. There has not been much research done with the NEMO extended Support Protocol. A mobile Network is composed of one or more IP subnets viewed as a single unit. The Mobile Router is the gateway for the communication between the mobile network and the internet.

121

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

An Access Router (AN) is a router at the edge of an access network which provides wireless link to mobile nodes. A link is simply a physical medium via which data is transformed between multiple nodes. A Home Link is the link attached to the interface at the Home Agent on which the Home Prefix is configured. Any Link other than Home link is foreign link. NEMO link is the link within the mobile network. A Mobile Router has two interfaces:Ingress Interface: The interface of the MR attached to a link inside the mobile network. Egress interface: The interface of the MR attached to the home link if the MR is at home and to foreign link if it is a foreign network. NEMO Basic Support protocol is an extension to the Mobile Ip version 6 (MIPv6) [2]. MIPv6 is a version of Internet Protocol (IP) that supports mobile nodes. III. MOBILE ROUTERS A Mobile Router is a router that can change its point of attachment to the network by moving from one link to another. All the Internet traffic to and from the mobile network passes through the Mobile Router. Therefore, Mobile Router has to perform certain operations to be able to support network mobility. IV. HROSTIO

VII.

EXTENDED TIO

The tree information option for Hierarchical Route Optimization scheme in network mobility is divided in two parts: A. Basic Hierarchical Route Optimization Scheme using Tree Information Option (HROSTIO). This is define hierarchical Route Optimization scheme between the Corresponding Nodes (CNs) and the Local Fixed Nodes (LFNs) B. Extended Hierarchical Route optimization Scheme using Tree Information Option (HROSTIO). VII. HIERARCHICAL ROUTE OPTIMIZATION BY PACKES DISRUPTION TREE (HROPD) For the Hierarchical Route Optimization Scheme using the Edmond’s Theorem. The tree system we study a one routing assignment, for information packets to go from the root of the directed tree, if sender is x and receiver is y. then we present a subgraph Gk = (Vk,Ek,Ck)≤ G=(V,E,C) when Ek = Edge Set Vk = Vertex Ck = Edge Capacity Hence a distribution tree Gk = (Vk,Ek,Ck)can deliver communication to receiver at a rate R(Gk) = min ck(e). e€Ek Hence we present hierarchical edge packs, as described below:
●

For the hierarchical Route Optimization scheme using tree information option (HROSTIO) we use an assistant data structure and call it MNN-CN(mobile network node-corresponding node) list .It is stored at MRs and records the relationship of the MNNCN. V. TREE INFORMATION OPTION

The tree information option Tio [4] avoids routing loops in a nested NEMO Fig. 2 shows the TIO formet in an RA message to prevent a loops MRs performance topology based on various metrics fig.1 is defined information option. A=0
Type Preference Tree depth

B=8016
Length
Tree pref.

C=150031
G H Reserved

Hierarchical edge bundling is a flexible and generic method that can be used in conjunction with existing tree visualization techniques to enable users to choose the tree visualization that they prefer and to facilitate integration into existing tools. Hierarchical edge bundling reduces visual clutter when dealing with large number of adjacency. Hierarchical edge bundling provide an intuitive and continuous way to control the strength of bundling. Low Bundling strength mainly provides low-level, node-to-node connectivity information, whereas high bundling strength provides high-level information as well by implicit visual of adjudges between parent nodes that they are the result of explicit adjacency edges between their respective child nodes Hierarchical edge bundling provide an intuitive and continuous way to control the strength of bundling. Low

BTRHRO
Tree delay

●

Path digest Tree information system Figure 1. BTRHRO(Boots Time Random for Hierarchical route Optimization system) VI. PERSONAL AREA NETWORK

●

A mobile network can have a hierarchical structure e.g. a mobile network within another mobile network. This situation is referred to as mobile network. A Personal Area Network (PAN) may travel a vehicle, which also contains a mobile network of larger scale MR-1, MR-2, MR-3 … are attaché their own home link. A wireless personal area network moves as a single unit with one or more mobile routers that connect it to global internet.
122

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

REFERENCES
Access Router
[1] Hosik, Cho, Taekyoung Kwon and Yanghee Choi, “Route Optimization using Tree Information Option for nested Mobile Networks.” IEEE, Journal of Selected Vol. 24.9, page 1717-1722, Area in Communications, 2006. [2] C. Perkins, “IP mobility support for IPv4, IETFRFC 3344, Aug. 2002. [3] Jhonson, C. Perkins and J. Arkko, “Mobility support in JPVG”, IETF, RFC3775, Jun., 2004. [4] T. Ernst and H. Lach, “Network mobility support terminology”, Interior draft, Feb., 2005 [Online] http://infreport.isoc.org/idref/draft-ietf-nemo-terminology. [5] M.A. Abido, A new multiobjective evolutionary algorithm for environmental /economic power dispatch. In Power Engineering Society Summer Meeting, Vol. 2, page 1263-1268, IEEE, 2001. [6] Pragya Lamsal, “Network Mobility”, Research Seminar on Hot Topics in Internet Protocols, page 1-2. [7] Joshua Knowles and David Corne, “Memetic algorithms for multiobjective optimization: issues, methods and prospect, page 1-40. [8] Yunnan Wu, Philip A. and Kamal Jain, “Acomparison of Network Coding and Tree Packing”, One Microsoft Way, Redmond, WA 98052-6399 USA. [9] Danny Holten, “Hierarchical Edge Bundles: Visualzation of Adjacency Relations in Hierarchical Data”, IEEE Transactions on visualization and Computer Graphics, VOL 12 NO. 5, September/October, 2006.

MR1

MR2

MR3

MR4

MR5

MR1, MR2……………………………MRn

Fig.-2 CONCLUSION Hierarchical Route Optimization scheme optimization scheme in mobile network modifying the process of Tree information option. And hence the NEMO basic support protocol needs to be extended with an appropriative route optimization scheme. the optimization scheme to easily solved by Tree information option. We propose a scheme can achieve the Hierarchical Route Optimization Scheme using Tree Information Option (HROSTIO) for route optimization environment.

AUTHORS PROFILE Authors Profile .. K K Gautam is the Dean in the Roorkee Engineering & Management Technology Institute, Shamli-247774, India. Prof Gautam is basically a mathematician and is working in the area of mobile network and wireless network for the past 3 years. Meenu Chaudhury is a lecturer in the Department of Computer Science at Roorkee Engineering & Management Technology Institute, Shamli-247 774, India. Meenu has B Tech degree in Computer Science to her credit.

123

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Seeing Beyond the Surface: Understanding and Tracking Fraudulent Cyber Activities
1

Longe,O. B. 2Mbarika, V.

3

Kourouma, M

4

Wada, F. & 5Isabalija, R

Int. Centre for IT & Development Southern University Baton Rouge, LA 70813
.

Dept. of Computer Science Southern University Baton Rouge, LA 70813

Nelson Mandela School of Public Policy Southern University Baton Rouge, LA 70813

Abstract The malaise of electronic spam mail that solicit illicit partnership using bogus business proposals (popularly called 419 mails) remained unabated on the internet despite concerted efforts. In addition to these are the emergence and prevalence of phishing scams that use social engineering tactics to obtain online access codes such as credit card number, ATM pin numbers, bank account details, social security number and other personal information[22]. In an age where dependence on electronic transaction is on the increase, the web security community will have to devise more pragmatic measures to make the cyberspace safe from these demeaning ills. Understanding the perpetrators of internet crimes and their mode of operation is a basis for any meaningful effort towards stemming these crimes. This paper discusses the nature of the criminals engaged in fraudulent cyberspace activities with special emphasis on the Nigeria 419 scam mails. Based on a qualitative analysis and experiments to trace the source of electronic spam and phishing e-mails received over a six months period, we provide information about the scammers’ personalities, motivation, methodologies and victims. We posited that popular e-mail clients are deficient in the provision of effective mechanisms that can aid users in identifying fraud mails and protect them against phishing attacks. We demonstrate, using state of the art techniques, how users can detect and avoid fraudulent e-mails and conclude by making appropriate recommendations based on our findings. Keyword: Spammers, Scamming, E-mail, Fraud, Phishing, Nigeria, IPLocator

I.

INTRODUCTION

Cybercrime refer to misconducts in the cyber space as well as wrongful use of the internet for criminal purposes. Various categories of these crimes include cyber stalking, phishing (identity theft), virus attacks, malware attack, the use of anonymous proxies to masquerade and sniff information and the popular electronic spam mail problem. Unfortunately, Cybercrime seems to be yielding much to criminals all over the world so the malaise is not going tom be curbed without some resistance from the criminals. Offline crime rates have reduced because the offline criminals have gone cyber. In fact, it is highly likely that Cybercrime and its perpetrators will continue developing and upgrading knowledge to stay steps ahead of the law. It is generally believed that most fraudulent mails in cyberspace and phishing attacks originate from or are traceable to Nigeria or Nigerians in other nations. These assumptions remains to be validated using empirical research [27]. Since the implication of cybercrime is well beyond its immediate points of perpetration and utilization of online facilities depends strongly on consumer trust, attention must be given to factors that can impede the effective use of information technology platforms for electronic transactions. The negative effects of online criminal activities on the use of the internet for electronic commerce, e-banking and other forms of usage has therefore increased interest in studying the factors that motivates these criminals, their tactics and what can be done to mitigate their activities[13][17]. The remaining part of the paper is organized as follows: In section 2 we discussed cybercrime in Nigeria with particular emphasis on scamming and phishing. Section 3 examined the nature of the cyber criminals, the victims and the tools used by these criminals. The scammers mode of operations are discussed in section 4. In section five we

.
124 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

provide simple guidance on how users can trace the source of suspected scam mails. We concluded the paper in section 6 by suggesting possible measures that can be adopted to address the problem and gave direction for future research.

A. Scamming Scamming or 419 fraud is usually in the form of "Advance Fee Fraud" (named after the relevant section of the Criminal Code of Nigeria that deals with such crimes). It begins when the target receives an unsolicited fax, e-mail, or letter often concerning Nigeria or another African nation containing either a money laundering or other illegal proposal. With the increase in use of internet facilities for electronic commerce, online banking, social networking and other financial transactions phishing attacks are also on the increase. Spamming was said to be one of the most prevalent activities on the Nigerian Internet landscape accounting for 18% of all online activities amongst others [11]. Information released by the United States Internet Crime Complaint Centre between 2006 and 2008 brings the Nigerian Spam situation to the fore as the nation maintained a high position among the first ten nations that serve as the source of Spam all over the world (see tables below). The report showed that confidence fraud, computer fraud, check fraud, and Nigerian letter fraud round out the top seven categories of complaints referred to law enforcement during the year. Of those complaints reporting a dollar loss, the highest median losses were found among check fraud ($3,000), confidence fraud ($2,000), Nigerian (west African, 419, Advance Fee) letter fraud ($1,650). Diplomatic missions around the world warn visitors to various West African countries such as Nigeria, Côte d'Ivoire, Togo, Senegal, Ghana, Burkina Faso and Benin Republic of susceptibility to 419 scams. Countries outside of West Africa with 419 warnings are South Africa, Spain, and The Netherlands. The effect of fraudulent spamming activities can be measured by the pressure the volume of spam messages placed on internet bandwidth, thus slowing it down. This is not helpful in an age when subscribers are clamoring for faster connections. It also increases the dial-up costs by extending the time a person spends reviewing email. When Spammers use false e-mail addresses and users attempt to respond to them, the e-mail bounces around in cyberspace loops creating huge administrative loads for Internet infrastructures. Other hidden costs involve the claims made on two precious human resources: time and energy. Computer users can spend hours attempting to identify the original sender of an e-mail.

II. CYBERCRIME IN NIGERIA Cybercrimes share some similarities with crimes that have existed for centuries before the advent of the cyber space. The major difference is that the internet now provides an electronic platform with the advantages of speed, anonymity and a tool which increases their potential pool of victims [1] Among the numerous crimes committed daily on the Internet, Nigeria and some other nations on the West African coast are reputed to be at the forefront of sending fraudulent and bogus financial proposals all over the world. The damaging implications resultant on the image of the Nigerian nation and the negative impact this trend has had on e-mail infrastructures are clearly evident. The United States Internet Crime Complaint Centre defined the Nigerian scam as “Any scam that involves an unsolicited email message, purportedly from Nigeria or another African nation, in which the sender promises a large sum of money to the recipient. In return the recipient is asked to pay an advance fee or provide identity, credit card or bank account information[25]. While this genre of cyber crimes are generally targeted at individuals, they require a degree of ingenuity to dealt real damage on the victims. The damage done manifests itself in the real world as human weaknesses such as greed and gullibility are generally exploited. No discussion about organized financial crime is complete in Nigeria without mentioning Fred Ajudua a Nigerian fraud kingpin who is currently standing trial in Nigeria for alleged involvement in advance fee fraud "419" charges. These charges relate to collecting money under false pretences from Nelson Allen, a Canadian who allegedly lost $285,000 to Ajudua and is the only foreigner to have given mail-fraud evidence in a Nigerian court of law. Other suspected Ajudua victims include Technex Import and Export Company of Germany, who lost the equivalent of $230,000 USD, and a German woman who lost the equivalent of $350,000 USD trying to collect purported dividends left by her late husband. It is generally believed that there are many more victims yet to come forward [28].

.
125 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Table 1: Amount Lost by Selected Fraud Type for Reported Monetary Loss Complaint Type % of Reported Total Dollar Of those who reported a loss the Loss Average (median) $ Loss per Complaint Nigerian Letter Fraud 1.7% $5,100.00 Check Fraud 11.1% $3,744.00 Investment Fraud 4.0% $2,694.99 Confidence Fraud 4.5% $2400.00 Auction Fraud 33.0% $602.50 Non-delivery 28.1% $585.00 Credit/debit Card Fraud 3.6% $427.50
Source Internet Crime Complaint Centre Report (2006-2008) [25]

Table 2: Top Ten Countries - Perpetrator of Cybercrime Year 2006 Year 2007 United States United Kingdom Nigeria Canada Romania Italy Netherlands Russia Germany South Africa 60.9% 15.9% 5.9% 5.6% 1.6% 1.2% 1.2% 1.1% 0.7% 0.6% United States United Kingdom Nigeria Canada Romania Italy Spain South Africa Russia Ghana 63.2% 15.3% 5.7% 5.6% 1.5% 1.3% 0.9% 0.9% 0.8% 0.7%

Year 2008 United States United Kingdom Nigeria Canada China South Africa Ghana Spain Italy Romania 66.1% 10.5% 7.5% 3.1% 1.6% 0.7% 0.6% 0.6% 0.5% 0.5%

Source Internet Crime Complaint Centre Report (2006-2008) [25]

With volumes such as this, tremendous burdens are shifted to the ISP to process and store unnecessarily large amount of data [29]. Users bear the costs involved with Spam. Payment occurs when a user is “taken” by the Spammer and pays for services or products never to be realized. This happens more frequently than one might expect. In [30] it was opined that there are also the hidden costs masked as “yearly access fee increases” to help the ISP provide better service to the users. This refers to dealing with regular disruptions to the integrity of the systems which results into the ISPs simply passing their costs down to subscribers and other stakeholders. This has obviously contributed to many of the access, speed and reliability problems seen with lots of ISPs today. Spamming and phishing interferes with ordinary e-mail communication, reduces employee productivity and engender lack of trust in the electronic

infrastructure. It has been projected that spamming will cost 1.4 percent of employee productivity, or N131,100 per year per employee in Nigeria, the equivalence of $874 in the US [11]. Filtering Spam is therefore an extremely urgent problem. With the escalation in the volume of spam mails on the webscape, organizations are also being subject to mounting pressure to deal with issues regarding employee productivity, morale, sexual harassment and congestion of the e-mail infrastructure. B. Phishing Phishing is a social engineering scam that involves luring unsuspecting users to take a cyber-bait much the same way conventional fishing involves luring a fish using a bait. Phishing deceives consumers into disclosing their personal and financial data, such as secret access data or credit card or bank account numbers, it is an identity theft. It is an attempt to elicit a specific response to

.
126 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

a social situation the perpetrator has engineered (Teri, 2002). Identity theft schemes take numerous forms and may be conducted by e-mail (phishing), standard mail, telephone or fax. Thieves may also go through trash looking for discarded tax returns, bank records, credit card receipts or other records that contain personal and financial information so as to use someone’s personal data to steal his or her financial accounts(http://www.irs.ustreas.gov) Phishing scams are on the increase in Nigeria. The most recent phishing attacks were on the customers of Interswitch, the banking and financial system backbone organization with the highest customer base in electronic transactions in the country. According to APWG, The number of unique phishing websites detected during the second half of 2008 saw a constant increase from July – October with a high of 27,739 (http://www.antiphishing.org). The Nigeria Deposit Insurance Corporation (NDIC) disclosed in its 2007 annual report and statement of account that under-hand deals by bank staff, among others, resulted in attempted fraud cases totaling N10.01 billion and actual losses of N2.76 billion in 2007 [15]. With the present economic downturn and appropriate technology, fraudulent actions are most likely to increase and phishing remains one of the means of committing “fraudulent crimes without borders”. The case of three people, two of them Nigerians, who were arrested by the police after having conducting a phishing scheme with approximately 30 victims was reported in [2]. They posted fake e-mails to the clients of a local bank in India asking them to visit a link which required them to enter private details such as credit card number, PIN and other information. Once the users entered the information, the phishers received and used it to transfer over $100,000 from the victims accounts. In their paper “why phishing works” Rachna et al [20] came up with the fact that good phishing websites fooled 90% of users and existing antiphishing browsing cues are ineffective. It was also reported that 23% of users do not look at the address bar, status bar, or the security indicators and on the average users made mistakes 40% of the time in identifying phishing websites. Phishing attacks have convinced up to 5% of their recipients to provide sensitive information to spoofed websites. By hijacking the trusted brands of wellknown institutions, phishers are able to convince a small percentage of recipients to respond to them[7]. The question however remains as to whether or not the success of phishing scams are a result of underlying security flaws in web security

or simply the result of laxities in user assessment of phishing offers.

III. THE CYBERCRIMINALS, THEIR VICTIMS AND THE TOOLS. All forms of crime involve two major players are involved. The criminals and the victims. To analyze conventional criminal conducts, sociologists focus on two main issues viz the crimes and the criminal committing the crime [4][6]. Criminologists are primarily concerned with the sociological factors that cause, or are correlated with a person becoming a criminal and engaging in criminal activities. The social learning theory [5], social bonding theory [12 and rational choice theory are some of the sociological theories that address these issues. Rational choice theory argues that people make a basic decision to commit a crime, or to not commit a crime, based on a simple cost-benefit analysis. The rational choice theory focused on non-sociological factors that can influence the decision to commit crime [31]. For example, electronic mechanisms such as user ID, automated access control systems and surveillance camera can serve as deterrents because they increased the perceived risk of being apprehended [6]. To effectively deal with cybercrime, an understanding of the crime, the criminals, the victims and the tools are required. A. The Criminals – The Nigerian 419 Spammers To understand the nature of cyber criminals, it is important to have an idea of the general misconceptions internet users have about the criminals. In the article “Piercing the darkness – Misconceptions about Cyber criminals” Thinkquest [23] identified four major misconceptions about cybercriminals. These are: (1) Misconception 1: All cyber criminals are smart but social misfits (2) Misconception 2: Cyber criminals are not “real” criminals (3) Misconception 3; Teenagers with computers are all cyber criminals (3) Misconception 3: All cyber criminals have the same characteristics Some studies also opined that the 419 cohorts are illiterates or semi-illiterates. These conclusions were based on the lack of fluency in the use of English, grammatical errors, lexis and the general syntax of the sentences in most 419 mails [21].

.
127 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Other studies claimed that the fraud team consists of Nigerians in Diaspora operating (with assistance from some confederates in the West African coasts) from various places in Europe, Asia, America, Australia and the middle east [19]. There are people who also believed there are not enough computers in Nigeria for these criminals to perpetrate scams on a scale this large and that scammers are all from a particular tribe in Nigeria. While some of these assumptions are tenable others are clearly misguided. The Nigerian Spammers are a mixture of criminally minded elites and semi-elites who simply moved from conventional offline 419 schemes to an electronic platform offered by the Internet. It is clearly an extension of the popular fax and postal mails scams that originated in Lagos, Nigeria back in the 80s[10]. Qualified graduates and undergraduates from universities and polytechnics, high school and secondary school students, kingpins of the underworld, semiilliterates and the unemployed are all involved in these heinous crimes. The level of co-ordination and sophistication now involve very high degree of intelligence and software instruments that are difficult to beat by spam filters and human beings alike. The cybercrime trade is fast turning to an internationally co-ordinated and controlled system. There are indications that computing and allied professionals are recruited specifically to provide expertise in the context of hacking and designing tools that will assist in beating security measures. The evolution of fixed wireless facilities providing internet access anywhere and anytime has enabled a migration of these criminals from the public internet access points to the comfort of their homes and offices. There is therefore a very high degree of mobility of criminals and the network is further enhanced by the availability of high speed wireless connections and the Global System for Mobile communication (GSM). Today, there are in existence, very organized cartel-like organizations mushrooming in different location, especially in the south west and south eastern part of the country. A common impetus among these cyber criminals is their desire to make “big money” that will enable them afford buying “big cars” and “powerful camera phones”. These criminals have successfully swindled locals and foreigners of their money and these success are measureable in terms of very fat bank accounts and exotic cars seen among youths involved in scamming. Their success serves as an impetus for others to get involved and wait for their “lucky day”.

B. The Victims Victims of cybercrime are many and varied. They range from individuals, business organizations, religious organizations, philanthropic organizations and educational institutions. Understanding who the criminal is likely to target can assist in taking preemptive actions to forewarn and prepare for all forms of attack. Intelligent criminals always target people whose circumstances are loaded with some forms of vulnerability. The yahoo boys are fond of people who are easy to deceive. Oral interviews of cyber criminals in cyber cafes in Nigeria yielded the response that “yahoo-yahoo business is all about deceit, if you are gullible, then you become a victim” . Users of internet facilities have to be on guard against all forms of solicitation that comes from strangers with very enticing dividends. One nature that the cybercriminal prey upon most is gullibility and greediness. There are users who see the internet as a very easy tool with which to make money. They gullibly provide private information and account details. Usually, they have some dividends from such transactions at the initial stage. This success takes them deeper into the con as they build more relationship with the scammers and become embroiled in legal and financial entanglements out of which only the perpetrator will make profits. Others are enticed by advertisements and offers that invite them to try out new products and means of making money. The Diversity Visa Lottery spam is used to fool local victims who are desperate about travelling abroad. Foreigners are engaged and embroiled in conference invitations and asked to pay registration fees that are cornered by the con men. Fictitious websites are set up for employment purposes using phishing. In some cases, these criminals also set up dating sites requesting for personal information and luring the victim to play along as they buy time to extract credit card numbers and render the customer bankrupt. Unfortunately the sheer embarrassment of being defrauded online has prevented most victim from reporting their ordeal, preferring to silently bear the pain and possibly learn their lessons in a bitter way. It is unbelievable at times to imagine the extent of activities involved between the criminal and the victim before money exchange hands. Some correspond for weeks and months, planning, scheming together and building trust with someone they have never met and whose credentials they cannot prove. Victims are enticed by stories that are usually emotion-laden, financially-promising or religiously-toned. The

.
128 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

most unfortunate thing about it all is that most victims act alone. No consultation or efforts are made to seek advice from experts or security agent before committing their time, money and energy to deals offered on the internet. C. The Scammers Tools A combination of social engineering and programming skills are the most potent tools in the hands of the 419 scammers. In order to reach a large volume of users, the scammers require an equally large number of email addresses. These are usually collected by using programs known as spam-bots to search for email addresses listed on web sites and message boards, by performing a dictionary attack (pairing randomly generated usernames with known domain names to ‘guess’ a correct address) or by purchasing address lists from individuals or organizations. Once they have addresses, spammers can use programs known as “bulk mailers” to automate the sending of spam. These programs can send huge volumes of email messages in a small amount of time. Some bulk mailing programs engaged by the spammers use open-relays to send messages, effectively hiding the true address of the spammer. Bulk mailers can also fabricate the from address in email message headers to further hide the identity of the spammer [8]. Another technique spammers utilize to send emails is with the use of zombie networks, also known as bot networks. Zombie is the term given to a computer that has been infected by a virus, worm, or Trojan Horse [9], which allows remote entities to take control and use it for their own (usually illegal) purposes. A large amount of these computers, usually called a network or army can be co-opted to send spam emails, requiring little of the spammer’s own computing power and network bandwidth. This technique is also popular as it protects the identity of the spammer [18]. Another popular method employed by scammers is the use of dating sites as a powerful tool to get attention and e-mail addresses. A number of victims have fallen victim to dating scams. Religious persuasions and emotional-laden mails are designed to attract attention and sympathy from religious organizations. Just as the web security community developed personalized electronic mail filtering systems, scammers also develop tactics that are personalized. They profile individuals, trace their business history with individuals they have been involved with in the past. The knowledge of old business acquaintances abroad are employed to compose emotion-laden letters with bogus business proposals from

Nigerians purportedly in government looking for opportunities to launder money abroad through a friend or two. Keystroke loggers are also used by these criminals to carefully collect personal information from unsuspecting victims. This trick is employed when unsuspecting users log on to the wrong website during a request for program update. This is particularly targeted at financial organizations. Most phishing attacks from either state that there has been some sort of fraud detected on bank accounts or that for security reasons the company just wants everyone to validate their usernames and passwords. In either event, the attack preys on fear and naiveté to get people to respond by providing sensitive information such as usernames, passwords, account numbers, etc. Cyber criminals can combine phishing with the 'Nigerian Bank Scam' to use greed rather than fear as the driving force to prey on individuals [24].

IV. THE MODUS OPERANDI In order to understand current techniques used by the 419 scammers, we set up electronic mail accounts on popular free e-mail clients such as yahoo, hotmail and excite for a six month period. Additionally, we also operated the e-mail account from locations outside Nigeria, specifically in the United States of America and United Kingdom to ascertain if there are variations in the content and style of electronic mails received from the conmen at these locations. We deliberately did not subscribe to any form of promotion, materials or subject while registering the e-mail account. This is to enable us have a clean slate to operate and avoid other forms of spam. We employed the software tools IP2Location, IPGeolocator and GeoBytes IP Locator to track the source of the electronic mails that were received. We provide samples of these e-mails in the figures below. A breakdown and analysis of the various genre of email is presented in Table 3.

.
129 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Table 3: Genre of Spam Mails and their Statistics
Total No Total of % of false Spamming Technique Employed of Mails False Negative One sentence e-mail consisting only a weblink that 78 21 26.9% take the recipients to another page where the actual scam is to be perpetrated 73 39.0% Spoofing/phishing This type of mail comes with fictitious sender 187 addresses. They usually direct users to respond to an entirely different address(es) from the sender address. Usually, the two addresses looks similar. i.e longeolumide@yahoo.com and olumidelonge@yahoo.com Image With These e-mails contains an image of any company or 152 62 40.7% Superimposed organization on which text are superimposed to Spam communicate to the user. An example is shown in Fig. 2 Subject Title and This e-mail type makes the subject of the e-mail and 107 24 22.4% Sender ID the sender header the same. sender and specify the Similarity same. i.e a message title: From Longe Sender : From Longe Bayesian This type of e-mail usually contain a link to another 261 77 29.5% Poisoning website with an e-mail body loaded with a lot of grammatical error and meaningless sentences to confuse the Bayesian filter. At other times, tokens are manipulated using word toggling, a combination of numerals and letters i.e AgeiNg numbe3rs etc. Attachment Only These e-mails contain just an attachment or two 78 19 24.3% E-mail with no contents at all. Note: False negatives are spam mails wrongly classified as good mails (found in the Inbox) Type One line Subject Link

As at the time of writing, a total of 1113 unsolicited electronic mails have been received from the e-mail accounts set up for the research between April – October, 2009. Out of this number, purely pornographic e-mail numbered 45 with 37 of them received when the client moved between Europe and America. Marketing e-mail for medication, commercial products and webinars were 76 in number. The remaining 955 mails collated from all the various locations are purely 419 mails. Out of this number, 187 are purely spoofed/phishing e-mails that direct users or recipients to websites set up for collecting personal information. Others in this category direct the reader to subscribe to one product or the other with the clause that “if the e-mail is received in error, the reader should click a link to unsubscribe” in order to stop receiving the mail. A click on such link is a subtle way to validate e-mail addresses as it reflects the fact that the mail has been received by a user. There is also a degree of overlap among 68 of the remaining e-mails making them fall into other categories.

These were discarded. Of the 800 remaining 419 e-mails, 633 were selected because they satisfy stratification for language, type and content.

Fig. 1: Lottery Winning and Financial Scam .
130 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

V. EXPERIMENT WITH E-MAIL SOURCE AND IP ADDRESS LOCATION It is important as a first line of defence for emailers to know how to track e-mail origin. We use yahoo mail account for our demonstration here since it is one of the most popular e-mail clients. The sequence is almost the same for all other email clients. The User need to be able to view the full header of an e-mail before they can track the origin of any mail. For yahoo, the “Full Header” option can be seen at the bottom right side of each e-mail message received. The header section contains a lot of information relating to the mail. Sometimes there are several Received From’s in the header. This is so because the header contains the IP addresses of all servers involved in routing e-mail from one point to another. With this addresses a user can track an e-mail origin and possibly the identity of the sender. We selected some mails randomly from our corpus of spam mails. To find the actual location from which the e-mail originates we pick the “Received From” IP that is at the bottom of the list on the header view. The results below is among others for an e-mail purportedly originating from Interswitch Nigeria limited claiming that “an account has been suspended and personal details should be provided for reactivation”. For the sake of brevity, a portion of the header details is given below (Fig. 3).

Fig. 2: E-mail Scam using Images

Fig 3. Images Spam With Text Superimposed.

Fig 4: Pure 419 Scam for Phishing

Fig. 5: Letter Purportedly from Interswitch Nigeria Limited

.
131 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

The next step is to look up the IP address. There are freeware open source software that are designed to do just that. In this work, we use the IP2Location and GeoBytes IP Locator and TraceE-mail.

We checked for the source of this e-mail using the IP address on the e-mail as well as the e-mail address itself. IP2Location produced the result below when fed with the IP address.

Now, you can run IP address query in your desktop even without network. Please download and install the IP2Location Application today.
IP Address 160.36.0.84 160.36.0.84 160.36.0.84 Country (Short) US US US Country (Full) UNITED STATES UNITED STATES UNITED STATES Flag Region City ISP Map

TENNESSEE KNOXVILLE UNIVERSITY OF TENNESSEE TENNESSEE KNOXVILLE UNIVERSITY OF TENNESSEE TENNESSEE KNOXVILLE UNIVERSITY OF TENNESSEE

Fig 6: Result from IP2Location on the validity of the IP address

To further substantiate this result, we used another locator IP Location Finder and IP Locator and obtain similar results shown below.

Fig.7: Result from IP Location Finder and IP Locator on the validity of the IP address

Next we use the e-mail address “onlineservices-email@interswitchng.com” to check the validity of the email on IP Locator and E-mail trace. We obtained the result showing the letter emanates from Tennessee in the United States.

nvalid Domain Search
The email address onlineservices-email@interswitchng.com appears to be invalid. Click here to run another search. If you received a message from this address, the email was likely "spoofed" and actually from someone else. This is quite common.

Fig. 8; Result from IP Locator on the validity of the e-mail address

.
132 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

For a letter purportedly originating in Dubai (68.142.236.201) and inviting the authors for a job interview in Dubai. The test showed that the letter emanated from Sunnyvale, United States

Fig. 9: E-Mail trace validity of the e-mail address

For a letter purportedly posted inviting us to a conference from “Miss (Regina Pedro) presently working with (GLOBAL YOUTH ORGANIZATION FOR HUMAN RIGHT) California, USA” the e-mail address finder produced the result below. Fig. 11: Result from IP2Location With Map To be sure that the locators are reporting correctly, we search for the location of the computer being used at the time of writing this paper using IP2Locator and IP tracer. The results obtained are shown in Fig. 12. The authors were actually wordprocessing the paper at the Southern University, Baton Rouge, Louisiana, USA

Fig. 10: I P Goe Locator Validation

Fig. 12: Authors Location Authoring Paper

.
133 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

VI. DISCUSSION OF FINDINGS For the purpose of our research, we used medium spam filtering for the e-mail accounts by setting the spam filtering option to medium. This is intended to balance the degree of indirection between expert users and novice when using email accounts. Our findings from the analysis of the mail corpus showed that the greatest challenge the scammer face is how to beat existing filtering systems. As a result of this, they craft their e-mails in such a way that the victim get baited to phishing websites where they can divulge information from which the scammer can make financial gains. They therefore stick to tactics yielding dividends until the security system evolve to tackle it. Interestingly, we received mails from mailboxes set up abroad informing us about access to ATM CARD "POWERED BY INTERSWITCH" being temporarily limited” we were asked to click a link to restore the account in 48 hours. Interswitch, is the backbone organization that links all financial institution on ATM and smartcards in Nigeria. The link takes us to http://ussonline.free.fr/uss/administrator/compon ents/com_remository/login.html. The letter itself purportedly originated from onlineservicesemail@interswitchng.com>. The image–based spam with super-imposed text/content is a scheme that is yielding very positive results for the scammer. Over 40% of such mails get past current filters. This is followed by spoofing and Bayesian poisoning. A lot of efforts in literature has focused on dealing with Bayesian poisoning, unfortunately, it seems the spammers are always one step ahead as there a number of ways a mail can be composed to defeat these content-based spamming method. Lately, most Bayesian-poisoned e-mails are coined to guide users to phishing websites. The percentage of such e-mails that beat current spam filters are outstanding. The scammers tactics to also use attachment only, one subject link and sender and subject similarity are producing commensurate results (over 24% on the average). The spammer needs a turnaround of less that 5% to get good returns for their efforts. This research however cast some shadows on the general assumption that most fraudulent mails originate from Nigeria. While not excusing this view, empirical research is warranted to study trends as far as spam mail sources or origins are concerned. Most of the e-mails passed through the locators or IP trackers yield results that showed that they emanated from places outside Nigeria.

The interesting thing is that some of them lay claim to wanting to solve problems for people living in Nigeria. The validation of previous claims about the sources of fraudulent e-mails will have to be subjected to scientific experiments such as we have done in order to direct filtering efforts and legislation to reach out to all instances of location of the spam problem.

VII. CONCLUSION For spamming and identity theft to be successful weaknesses such as ignorance, oversight and lack of awareness are exploited by the spammers. Most users are devoid of the knowledge of the workings of the mailing system as well as online vulnerabilities. For users generally, disposing off bank statement carelessly close to ATM machines calls for great concern. Customers are unconscious of the fact that scammers can fiddle with trash bins to look for account details and use them to withdraw funds from their account. It is worth noting that phishers are getting smarter and it will not be out of place to say that in the future their methodologies could advance to using some elements of frustration and disruptions to financial network systems to compel users to provide information. A qualitative interview of regular emailers in the UK, the US and Nigeria on the problem of spam mails revealed that some users are not aware that there are provisions within the e-mail client to mark messages as spam in order to prevent them from receiving spam from the same source in the future.

The awareness of blacklisting spam mails and whitelisting wrongly classified mails is also low. Most users are not conscious of or have never use reporting facilities that help authorities track phishing websites. An average user cannot differentiate between a legitimate Universal Resource locator (URL) and a fake. In some cases, users are not knowledgeable about how to identify and distinguish between good and fake security indicators on visited websites. This lack of awareness and information on internet scam and what to look out for in order to prevent phishing attacks is one of the major reasons why phishing and other electronic fraud are successful. A. Recommendations Considering the far reaching effects of scamming, it is appropriate at this juncture to recommend that service providers and organizations orient users

.
134 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

about their vulnerabilities online and be equipped with standard practice measures to forestall and prevent them from falling victim to cybercrime. The present web design security architecture and common e-mail clients are deficient in the provision of effective mechanisms that can aid users (as the last line of defense) to identify and avoid fraud mails and phishing attacks. It is important to address the issue of security indicators in web design. These indicators have to be made more glaring and obvious enough for users to identify. E-mail clients can also make it easy for users to avoid and fight spamming by providing full lexical information in the headers rather than the codification of IP addresses and [17] Mailing addresses that we currently have. B. Direction For Future Work In the future, we intend to perform a comprehensive experiment relative to the source/origin and location on the corpus of spam e-mails and phishing websites that we are building. This will provide a scientific basis regarding the type of e-mails that come from particular origins and show the skewness of these mails by location. ACKNOWLEDGEMENT
This work is supported by the (2009) University of Ibadan MacArthur Foundation Grant for Manpower Training & Development.

REFERENCES
[1] Aghatise, E.J. 2006. Cybercrime definition”Computer Crime Research Center Publications. [2] Bogdan, P (2008): Nigerian Phishers Arrested Three phishers forced to go on a trip behind bars. Security and Search Engines Editor. 10th of April 2008, 20:31 MThttp://news.softpedia.com/news/Nigerian-PhishersArrested-83024.shtml [3] Loftesness, S. Responding to "Phishing" Attacks. Glenbrook Partners (2004) [4] Akers, R. L. (1994) Criminological Theories: Introduction and Evaluation, Roxbury Publishing Company: Los Angeles, 1-236. [5] Burgess, R. L. and Akers, R. L. (1966) A Differential Association-Reinforcement Theory of Criminal Behaviour, Social Problems, 14, 128-147. [6] Clarke, R. V. (1997) Situational Crime Prevention: Successful Case Studies, Harrow and Heston Publishers: Guilderland, 1-357. [7] Dahamija, R, Tygar, J and Hearst, M. (2006); Why Phishing Works. Why Phishing Works. UC Berkeley: Experimental Social Science Laboratory (Xlab). Retrieved from: http://escholarship.org/uc/item/9dd9v9 [8] Garcia F, Hoepman J, Van J, and Nieuwenhuizen J.(2004): Spam filter analysis. Proceedings of 19th IFIP international information security conference, WCC2004SEC, Toulouse, France. Kluwer Academic Publishers. [9] Levy, Steven (1984). Hackers: Heroes of the Computer Revolution. Doubleday. ISBN 0385191952

[10] Longe, O., Chiemeke, S., Onifade, O. And Longe, F. (2008): Camouflages and Token Manipulations-The Changing Faces of the Nigerian Fraudulent 419 Spammers. African Journal of Information & Communication Technology, Vol 4, No 3 [11] Longe, O.B. and Chiemeke, S.C. 2006. The design and implementation of an e-mail Encryptor for combating internet spam” Proceedings of the Ist International Conference of the International Institute of Mathematics and Computer Sciences. Ota, Nigeria. June. pp 1 - 7 [12] Hirschi, T. (1969) Causes of Delinquency, University of California Press: Berkeley. [13] http://www.irs.ustreas.gov/newsroom/article/0,,id.htm. [14] http://www.antiphishing.org [15] http://www.dailytrust.com, 2008 [16] Schwartz, A. and S. Garfinkel (1998): Stopping Spam: Stamping Out Unwanted E-mail and News Postings . O’Reilly & Associates, Inc., Sebastopol, CA [17] Patrick, F. (1999) Spam: Not just for breakfast anymore unsolicited e-mail in the business environment. Available online at http://www.bc.edu/bc.html#fna [18] Paulson, LD 2004, 'Spam hits instant messaging', Computer, vol. 37, no. 4, p. 18. [19] Peter, G. and Grace, D. 2001. Red flags of fraud. Trends and Issues in Crime and Criminal Justice, No. 200, Australian Institute of Criminology, Canberra. pp. 1-6. [20] Rachna D., Tygar, J. and Marti Hearst, (2006): "Why Phishing Works" in the Proceedings of the Conference on Human Factors in Computing Systems (CHI2006). [21] Smith, R. G., Holmes, M. N. And Kaufmann, P. (1999): Nigerian Advance Fee Fraud., Trends and Issues in Crime and Criminal Justice, No. 121, Australian Institute of Criminology, Canberra. http://www.aic.gov.au [22] Teri, B. 2002. Hack proofing your identity in the information age. First edition. Pg 3, 4. Syngress Publishing, Inc, 800 Hingham Street, Rockland, MA 02370, retrieved Oct. 15, 2008 [23] Thinkquest (2004): http://library.thinkquest.org [24] Tony, B (2005); 'Nigerian Bank Scam' Meets Phishing Attackhttp://netsecurity.about.com/b/2005/02/20/nigerianbank-scam-meets-phishing-attack.htm [25] USIC3 - Internet Crime Complaint Centre Report (20062008). www.ic3.gov/media/annualreports.aspx [26]Tony, B. (2009): Gone Phishing http://netsecurity.about.com/od/secureyouremail/a/.htm [27] Wendy L. Cukier, Eva J. Nesselroth, Susan Cody: Genre, Narrative and the "Nigerian Letter" in Electronic Mail. Proceedings of the 40th Annual Hawaii International Conference on System Sciences HICSS 2007: 70 [28] The Wikimedia Project (2008): Electronics SPAM – New Trends. Available online at http://www.en.wikipedia.org/ELECTRONICSPAM [29] Deborah, F. (2005) “Spam: How it is hurting e-mail and degrading life on the Internet,” Deborah Fallows, Pew Internet & American Life Project Available online at <http://www.pewinternet.org/report_display.asp?r=102>; [30] Schwartz, A. and Garfinkel, S. (1998): Stopping Spam: Stamping Out Unwanted E-mail and News Postings . O’Reilly & Associates, Inc., Sebastopol, CA [31] Beebe, N and Rao, V. (2005): Using Situational Crime Prevention Theory to Explain the Effectiveness of Information Systems Security. Proceedings of the 2005 SoftWars Conference, Las Vegas, NV, Dec 2005.

.
135 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

On the Efficiency of Fast RSA Variants in Modern Mobile Phones
Klaus Hansen, Troels Larsen and Kim Olsen
Department of Computer Science University of Copenhagen Copenhagen, Denmark

Abstract—Modern mobile phones are increasingly being used for more services that require modern security mechanisms such as the public-key cryptosystem RSA. It is, however, well-known that public-key cryptography demands considerable computing resources and that RSA encryption is much faster than RSA decryption. It is consequently an interesting question if RSA as a whole can be executed efficiently on modern mobile phones. In this paper, we explore the efficiency on modern mobile phones of variants of the RSA cryptosystem, covering CRT, Multi-Prime RSA, Multi-Power RSA, Rebalanced RSA and R-Prime RSA by comparing the encryption and decryption time using a simple Java implementation and a typical RSA setup. Keywords—Public-key cryptography; RSA; software; mobile phones.

do this by looking into various variants of the original RSA cryptosystem, thus taking an algorithmic approach rather than discussing hardware or software optimization schemes. The paper is organized as follows. In section II, we introduce the original RSA cryptosystem and its variants (except Batch RSA, which has low relevance for our investigation), and in section III discuss our implementation of these cryptosystems in a specific mobile environment. In section IV, we show our experimental results of these cryptosystems on modern mobile phones and comment on the degree to which the variants improve the decryption time of the original RSA. Finally, in section V, we conclude the paper and take stock on the state of RSA as a whole on modern mobile phones, i.e. whether both encryption and decryption can be executed efficiently on modern mobile phones. This paper is based on the results of a project carried out by the authors at the University of Copenhagen in the spring of 2008 [9]. II. RSA AND ITS FAST VARIANTS The original RSA cryptosystem was proposed in 1978 by Rivest, Shamir and Adelman [1] and consists of three parts: • Key generation: Given an integer n, generate two different primes p and q of (n/2)-bit each and compute N = pq and φ(N) = (p-1)(q-1). Choose a random integer 1 < e < φ(N) such that gcd(e,φ(N)) = 1. Next, compute the uniquely defined integer 1 < d < φ(N) satisfying ed ≡ 1 (mod φ(N)). The public key is <N,e> and the private key <N,d>. Encryption: To encrypt a message X with the public key <N,e>, transform the message X to an integer M in {0,…,N-1} and compute the ciphertext C = Me mod N. Decryption: To decrypt the ciphertext C with the private key <N,d>, compute M = Cd mod N and employ the reverse transformation to obtain the message X from M.1

I.

INTRODUCTION

Today, the virtually ubiquitous mobile phone is used for more complex services than just traditional voice calls and text messages. Many of these services require modern security mechanisms such as the Secure Sockets Layer (SSL) protocol. A number of these mechanisms, including SSL, make use of public-key cryptography. It is, however, well-known that public-key cryptography demands considerable computing resources. On limited platforms such as the mobile phone, this problem is exacerbated and one may ask if it is possible at all to implement public-key cryptography efficiently on mobile phones, i.e. where the time needed to perform encryption or decryption is sufficiently small to avoid having a negative impact on the user experience while retaining the security of the cryptosystem. In 2004, Groβshadl and Tillich showed that both encryption and decryption using public-key cryptography based on elliptic curves could be executed efficiently on some mobile phones [7]. The authors also showed that encryption using the publickey cryptosystem RSA [1] was feasible but that decryption was not. To the best of our knowledge, no basic research results have been published showing that (the decryption of) RSA can be executed efficiently on mobile phones. This paper investigates methods to optimize the execution of the decryption of RSA on modern mobile phones. We will

•

•

1

Note, we use the notation a mod b to mean the remainder when a is divided by b. The notation a ≡ c (mod b) means that a and c result in the same remainder when divided by b, i.e. a mod b = c mod b.

136

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Key generation is only performed occasionally so the efficiency of that part is less important than the two other parts, encryption and decryption. Their efficiency is determined by 1) the transformation of the message X to the integer M, and back, and 2) the modular exponentiations Me mod N and Cd mod N. The transformations can be performed with standard algorithms, e.g. from the Public Key Cryptography Standards (PKCS) published by RSA Security [13]. This leaves the modular exponentiation as the single most important component of RSA in regards to its efficiency. There exist a number of different methods for fast modular exponentiation [6]. In general, a modular exponentiation ab mod c can be performed using t-1 modular squarings and h-1 modular multiplications where t is the length of the binary representation of b and h is the number of 1's in the binary representation (i.e. its Hamming weight). In practice, an often used method to perform modular exponentiations is the so-called repeated square-and-multiply algorithm [10]. In its binary version, the algorithm processes a single bit of the exponent b at a time and on every iteration squares its intermediate result and multiplies it with a if the current bit is set. Thus, the algorithm always performs t-1 modular squarings and at most t-1 modular multiplications. Since an upper bound on a single modular multiplication – and therefore also squaring – is O(v2), the repeated square-andmultiply algorithm has a running time of O(tv2) where t is the bitlength of the exponent b and v is the bitlength of the modulus c. Given this bound, the encryption exponent e in the original RSA cryptosystem is typically chosen to be a small number, often 216 + 1. There are two reasons for this: the relative short bitlength of 216 + 1 will result in a small amount of modular squarings, and 216 + 1 has only two 1's in its binary representation leading to the fewest possible modular multiplications for valid RSA encryption exponents, namely one [14]. Expressed informally, choosing e this small almost effectively yields an encryption running time dependent only on the bitlength n of the modulus N, i.e. O(n2). The structure of the decryption exponent d cannot be tailored to fit the repeated square-and-multiply algorithm in the same way, but will often be long and consist of a random number of 1's in its binary representation. The worst case scenario is that |d| ≈ |N| yielding a running time of O(n3). This means that encryption is much faster than decryption in the original RSA. A number of variants of the original RSA cryptosystem have been published over the years all seeking to improve the time-consuming decryption operation. We take a closer look at these variants in the following subsections. A. CRT RSA CRT RSA is the most commonly known RSA variant for speeding up decryption. It was first described by Couvreur and Quisquater in 1982 [5]. The idea behind CRT RSA is to split the costly decryption into two smaller and faster modular exponentiations using the Chinese Remainder Theorem, hence the acronym CRT RSA.

The Chinese Remainder Theorem tells us that a system of r congruences x ≡ a1 (mod n1), …, x ≡ ar (mod nr), where n1,…,nr are relatively prime integers and a1,…,ar ordinary integers has a unique solution modulo N = n1n2···nr. This solution can be written as x = (a1N1y1 + … + arNryr) mod N, where Ni = N/ni and yi = Ni-1 mod ni for 1 ≤ i ≤ r [15]. CRT RSA uses the Chinese Remainder Theorem the following way: • Key generation: Generate e and d the same way as in the original RSA. Next, compute dp = d mod p-1 and dq = d mod q-1. The public key is <N,e> and the private key <p,q,dp,dq>. Encryption: Encryption is the same as for the original RSA, C = Me mod N. Decryption: Decryption is split into the following computations. First, compute Mp = Cdp mod p and Mq = Cdq mod q. Then, using the Chinese Remainder Theorem, find M as M = (Mpq(q-1 mod p) + Mqp(p-1 mod q)) mod N.

• •

If we ignore the contribution from the sum function of the Chinese Remainder Theorem, decryption using CRT RSA requires two times O((n/2)3) since the bitlength of both the exponents and the moduli are n/2 and a single modular exponentiation has an upper bound of O(tv2) where t is the bitlength of the exponent and v is the bitlength of the modulus. Compared to the O(n3) decryption of the original RSA, CRT RSA improves decryption time with a factor n3 / (2 · (n/2)3) = 4.2 B. Multi-Prime RSA Multi-Prime RSA represents a natural generalisation of CRT RSA: By adding more primes to the generation of N, decryption can be split into an arbitrary number of smaller exponentiations instead of just two. This variant of the original RSA was first described by Collins et al. i 1997 [4]: • Key generation: Given two integers n and r ≥ 3, generate r different primes p1,…, pr each (n/r)-bit long. Set N = ∏ri=1 pi and φ(N) = ∏ri=1 (pi-1). Compute e and d as in the original RSA. Next, compute di = d mod pi1 for 1 ≤ i ≤ r. The public key is <N,e> and the private key <p1,…,pr,d1,…,dr>. Encryption: Encryption is the same as for the original RSA, C = Me mod N. Decryption: Decryption is split into r exponentiations, Mi = Cdi mod pi, for 1 ≤ i ≤ r. Using the Chinese Remainder Theorem, M is found as M = (M1N1y1 + … + MrNryr) mod N where Ni = N/pi and yi = Ni-1 mod pi for 1 ≤ i ≤ r.

• •

This sort of big-o manipulation is formally not sound as big-o implies an arbitrary constant factor, but it does give a rough notion of the speed-up to expect in practice.

2

137

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

If we again ignore the contribution of the Chinese Remainder Theorem, decryption in Multi-Prime RSA requires r times O((n/r)3). Compared to the O(n3) decryption of the original RSA, Multi-Prime RSA improves decryption time with a factor n3 / (r · (n/r)3) = r2. Obviously, the individual primes need to have a certain size to guard against factorisation attacks. This puts a natural limit to the size of r and thus the actual improvement in decryption. Hinek suggest the following guidelines [8]:

as it cannot be efficiently applied for usual sizes of N in RSA [3]. D. Rebalanced RSA Basically, in the original RSA, encryption is more efficient than decryption because e is small and d is large. A straightforward way to optimise decryption would be to “switch” the exponents, i.e. make e large and d small. However, this is not to be recommended. Small values of d open up for Wiener's Low Decryption Exponent Attack [2]. Instead Wiener proposed in 1990 the variant Rebalanced RSA [17] that retains the size of d but makes dp = d mod p-1 and dq = d mod q-1 small (at the expense of a larger e): • Key generation: Given integers n and w, generate two different primes p and q each (n/2)-bit long such that gcd(p-1,q-1) = 2. Set N = pq and φ(N) = (p-1)(q-1). Compute two w-bit integers dp and dq satisfying gcd(dp,p-1) = gcd(dq,q-1) = 1 and dp ≡ dq (mod 2). Find a d such that d = dp mod p-1 and d = dq mod q-1. Compute e = d-1 mod φ(N). The public key is <N,e> and the private key <p,q,dp,dq>. Encryption: Encryption is the same as for the original RSA, C = Me mod N, but with a much larger e (on the order of N). Decryption: Decryption is the same as for CRT RSA but with smaller dp and dq – each w-bit long in Rebalanced RSA versus (n/2)-bit in CRT RSA. Usually, w ≥ 160 and n/2 ≥ 512.

n max r

1024 3

2048 3

4096 4

8192 4

C. Multi-Power RSA Multi-Power RSA is a variant of Multi-Prime RSA. In Multi-Prime RSA, the modulus N consists of r different primes whereas the modulus in Multi-Power RSA has the structure N = pr-1q for r ≥ 3. As shown below, this different structured modulus gives rise to a more efficient decryption than MultiPrime RSA. Multi-Power RSA was first described by Takagi in 1998 [16]: • Key generation: Given two integers n and r ≥ 3, generate two different primes p and q each (n/r)-bit long. Set N = pr-1q. Compute e as in the original RSA and d satisfying ed ≡ 1 (mod (p-1)(q-1)). Finally, compute dp = d mod p-1 og dq = d mod q-1. The public key is <N,e> and the private key <p,q,dp,dq>. Encryption: Encryption is the same as for the original RSA, C = Me mod N. Decryption: Conceptually, decryption is the same as CRT RSA. First, compute Mq = Cqdq mod q where Cq = C mod q and Mp = Cpdp mod pr-1 where Cp = C mod pr1 . Next, apply the Chinese Remainder Theorem to find M.

•

•

• •

If we again ignore the contribution of the Chinese Remainder Theorem, decryption in Rebalanced RSA requires two times O(w(n/2)2). Compared to the O(n3) decryption of the original RSA, Rebalanced RSA improves decryption time with a factor n3 / (2w · (n/2)2) = 2n/w. With respect to the security of Rebalanced RSA, it is recommended to set w ≥ 160 thereby limiting the actual improvement of decryption in practice. Note also that the speed-up in decryption comes at the cost of a much slower encryption. In the original RSA, e is small – typically 16 bit or less – whereas in Rebalanced RSA, e is on the order of N. This means that encryption in Rebalanced RSA is as slow as decryption in the original RSA. E. R-Prime RSA In Rebalanced RSA, decryption is the same as in CRT RSA. Since Multi-Prime RSA is a generalisation of CRT RSA, why not generalise Rebalanced RSA to use Multi-Prime RSA in its decryption as well. This is the idea behind R-Prime RSA, first described by Paixão in 2003 [12]: • Key generation: Given n and w, generate r ≥ 3 different primes p1,…,pr each (n/r)-bit long such that gcd(p1-1,…,pr-1) = 2. Set N = ∏ri=1 pi and φ(N) = ∏ri=1 (pi-1). Compute r w-bit integers dp1,…, dpr satisfying gcd(dp1,p1-1) = … = gcd(dpr,pr-1) = 1 and dp1 ≡ … ≡ dpr

Using Hensel Lifting, Mp can be computed using only one modular exponentiation modulu p instead of modulu pr-1, and some extra arithmetic. Ignoring the contribution from the Chinese Remainder Theorem and the extra arithmetic from the Hensel Lifting, this means that decryption in Multi-Power RSA requires two times O((n/r)3). Compared to the O(n3) decryption of the original RSA, Multi-Power RSA improves decryption time with a factor n3 / (2 · (n/r)3) = r3/2. With respect to the security of Multi-Power RSA, the same guidelines as for Multi-Prime RSA apply limiting the size of r and thus the actual improvement in practice. The special Lattice Factoring Method designed to factor integers with the structure N = prq is not a security issue for Multi-Power RSA

138

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

(mod 2). Find a d such that d = dp1 mod p1-1,…, d = dpr mod pr-1. Compute e = d-1 mod φ(N). The public key is <N,e> and the private key <p1,…,pr,dp1,…,dpr>. • Encryption: Encryption is the same as for the original RSA, C = Me mod N, but with a much larger e (as was the case with Rebalanced RSA). Decryption: Decryption is the same as for Multi-Prime RSA. That is, decryption is split into r modular exponentiations Mi = Cdpi mod pi for 1 ≤ i ≤ r after which the Chinese Remainder Theorem is applied. The difference lies in the length of dpi (denoted di in MultiPrime RSA). In R-Prime RSA, these values are w-bit each. In Multi-Prime RSA, they are n/r each.

No single variant outperforms all the others. Which variant is best is a question of usage scenario. If both encryption and decryption is needed, then Multi-Prime or Multi-Power RSA should be best. If only encryption is needed, there is no need to upgrade from the original RSA. And if only decryption is needed, then R-Prime RSA promises the most improvement. III. IMPLEMENTATION

•

We have implemented the original RSA and all its variants (except Batch RSA) to demonstrate their actual improvement in practice on modern mobile phones. As most mobile phones support Java's runtime environment, our implementation is written in Java's mobile platform Java 2 Micro Edition (J2ME). J2ME contains a subset of the Java Standard Edition (JSE). The Connected Limited Device Configuration (CLDC) and the Mobile Information Device Profile (MIDP) define the available APIs. We use the CLDC 1.1 and MIDP 2.1. These contain no general cryptographic API or a support for multi-precision computations like the BigInteger class. Therefore, we base the mathematical operations such as modular exponentiations on the BigInteger class from the bouncycastle.org [11]. The lack of built-in cryptographic APIs has the effect that each java application that require the use of cryptographic funcitons has to be bundled with code that provides such features. This leads to larger code sizes (in our case around 100 KB) and prevents code sharing. Finally, note that we do not implement any textual transformations but only the core encryption and decryption of M and C, respectively. IV. EXPERIMENTAL RESULTS

This means that the decryption in R-Prime RSA requires r times O(w(n/r)2) (ignoring the Chinese Remainder Theorem). Compared to the O(n3) decryption of the original RSA, RPrime RSA improves decryption time with a factor n3 / (rw · (n/r)2) = nr/w. R-Prime RSA inherits the same security considerations as Rebalanced RSA and Multi-Prime RSA, i.e. it is recommended to set w ≥ 160 and limit the value of r with the respect to n as listed earlier. As in Rebalanced RSA, the speed-up in decryption means a much slower encryption. Since Multi-Power RSA is faster than Multi-Prime RSA, why not use Multi-Power RSA in R-Prime RSA? The reason lies in the technique Hensel Lifting that is used in decryption in Multi-Power RSA. Hensel Lifting makes use of a number of exponentiations modulo e which is not a problem in MultiPower RSA since e is small. But in R-Prime RSA, e is on the order of N making Hensel Lifting (and consequently, MultiPower RSA) disproportionately expensive. F. Comparison The following table compares the decryption time of the original RSA to all its variants (excluding Batch RSA). The table shows the complexity of decryption for each variant, their general theoretical improvement and their approximated improvement in practice when n = 1024, and r = 3 and w = 160 where applicable: Variant Original CRT Multi-Prime Multi-Power Rebalanced R-Prime Complexity O(n ) 2 · O((n/2)3) r · O((n/r)3) 2 · O((n/r)3) 2 · O(w(n/2)2) r · O(w(n/r)2)
3

Our implementation of the original RSA and its variants is tested on a HTC Touch Dual from 2007 with a Qualcomm MSM 7200 400MHz processor and 128MB SDRAM. We test encryption and decryption of the original RSA and decryption for all the variants. We set n = 1024, and r = 3 and w = 160 where applicable. M is the same for all tests. The following table summarizes the results of our experiment by listing the average of twenty decryptions for each cryptosystem (and encryptions for the original RSA), the actual improvement in decryption and the approximated improvement in decryption (all figures are in ms): Variant Original CRT Multi-Prime Multi-Power Rebalanced R-Prime Enc. 29 Dec. 2098 558 283 210 187 151 Actual 1.0 3.8 7.4 10.0 11.2 13.9 Appr. 1.0 4.0 9.0 13.5 12.8 19.2

Theo. 1.0 4.0 r2 r2/2 2n/w nr/w

Appr. 1.0 4.0 9.0 13.5 12.8 19.2

139

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

As expected, the actual improvement for each variant is less than the approximated improvement due to overhead from the Chinese Remainder Theorem and other arithmetic. This is most outspoken for Multi-Power and R-Prime making Rebalanced faster than Multi-Power while R-Prime retains its position as the fastest decryptor. As evident, the approximated improvements are useful as rough guidelines. It is interesting to note the relative improvement from CRT to Multi-Prime is bigger than the relative improvement from Rebalanced to R-Prime. An closer inspection of the variants reveals that this is due to the fact that the generalization from CRT to Multi-Prime results in both smaller exponents and moduli while the generalization from Rebalanced to R-Prime only leads to smaller moduli as the exponents are fixed at w = 160. All in all, our experimental results show that the actual decryption time of each variant is well below one second – and verify that the actual encryption time of the original RSA is well below that of all the decryption times. V. CONCLUSION

REFERENCES
[1] L. M. Adelman, R. L. Rivest and A. Shamir. “A Method for Obtaining Digital Signatures and Public-Key Cryptosystems.” Communications of the ACM, Vol.21, No.2, pp. 120-126, 1978. D. Boneh and G. Durfee. “Cryptanalysis of RSA with Private Key d Less than n0.292.” IEEE Transactions on Information Theory, 46(4):13991349, 2000. D. Boneh and H. Shacham. “Fast Variants of RSA.” RSA Laboratories, 2002. T. Collins, D. Hopkins, S. Langford and M. Sabin. “Public Key Cryptographic Apparatus and Method.” US Patent #5,848,159, 1997. C. Couvreur and J. J. Quisquater. “Fast Decipherment Algorithm for RSA Public Key Cryptosystem.” Electronic Letters, Vol. 18, 1982. D. M. Gordon. “A Survey of Fast Exponentiation Methods.” Journal of Algorithms, Vol. 27, No. 1, pp. 129-146, 1998. J. Groβshadl and S. Tillich. “A Survey of Public-Key Cryptography on J2ME-Enabled Mobile Devices.” Computer and Information Sciences, Lecture Notes on Computer Science 3280, pp. 935-944, 2004. M. J. Hinek. “On the Security of Multi-Prime RSA.” University of Waterloo, Canada, 2006. T. Larsen and K. Olsen. “Public Key Cryptography in Mobile Phones (in Danish).” University of Copenhagen, Denmark, June, 2008. A. Menezes, P. van Oorschot and S. Vanstone. “Handbook of Applied Cryptography.” 1st Edition, CRC Press, 1997. Legion of the Bouncy Castle. “The Bouncy Castle Crypto APIs for Java.” www.bouncycastle.org, 2009. C. A. M. Paixão. “An Efficient Variant of the RSA Cryptosystem.” University of São Paulo, Brasil, 2003. RSA Laboratories. “PKCS #1 v2.1: RSA Cryptography Standard.” RSA laboratories, 2002. N. Smart. “Cryptography: An Introduction.” McGraw-Hill, 2003. D. R. Stinson. “Cryptography – Theory and Practice.” 3rd Edition, Chapman and Hall/CRC, 2006. T. Takagi. “Fast RSA-type Cryptosystem Modulo pkq.” Proceedings of Crypto 98, Vol. 1462, pp. 318-326, 1998. M. Wiener. “Cryptanalysis of Short RSA Secret Exponents.” IEEE Transactions on Information Theory, Vol. 36, No. 3, pp. 553-558, 1990.

[2]

[3] [4] [5] [6] [7]

[8] [9] [10] [11] [12] [13] [14] [15] [16] [17]

We have implemented the central part of various variants of the RSA cryptosystem: CRT RSA, Multi-Prime RSA, MultiPower RSA, Rebalanced RSA and R-Prime RSA and run a test on typical data that shows that they improve the decryption time of the original RSA considerably, achieving actual decryption times well below one second. Consequently, we are able to assert that both RSA encryption and decryption can be executed efficiently on a modern mobile phone.

140

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

An Efficient Inter Carrier Interference Cancellation Schemes for OFDM Systems
B.Sathish Kumar K.R.Shankar Kumar R.Radhakrishnan
Department of Electronics and Communication Engineering Sri Ramakrishna Engineering College Coimbatore, India.

Abstract— Orthogonal Frequency Division Multiplexing (OFDM) has recently been used widely in wireless communication systems. OFDM is very effective in combating inter-symbol interference and can achieve high data rate in frequency selective channel. For OFDM communication systems, the frequency offsets in mobile radio channels distort the orthogonality between subcarriers resulting in Inter Carrier Interference (ICI). ICI causes power leakage among subcarriers thus degrading the system performance. A well-known problem of OFDM is its sensitivity to frequency offset between the transmitted and received carrier frequencies. There are two deleterious effects caused by frequency offset one is the reduction of signal amplitude in the output of the filters matched to each of the carriers and the second is introduction of ICI from the other carriers. This research work investigates three effective methods for combating the effects of ICI: ICI Self Cancellation (SC), Maximum Likelihood (ML) estimation, and Extended Kalman Filter (EKF) method. These three methods are compared in terms of bit error rate performance and bandwidth efficiency. Through simulations, it is shown that the three techniques are effective in mitigating the modulation schemes, the ML and EKF methods perform better than the SC method. Keywords- Orthogonal frequency Division Multiplexing (OFDM); Inter Carrier Interference(ICI); Carrier to Interference Power Ratio (CIR);Self Cancellation(SC);Carrier Frequency Offset (CFO); Maximum Likelihood(ML); Extended Kalman Filtering(EKF).

performance in unequalized channels and in multipath channels. In this paper, the effects of ICI have been analyzed and three solutions to combat ICI have been presented. The first method is a self-cancellation scheme[1], in which redundant data is transmitted onto adjacent sub-carriers such that the ICI between adjacent sub-carriers cancels out at the receiver. The other two techniques, maximum likelihood (ML) estimation and the extended Kalman filter (EKF) method, statistically estimate the frequency offset and correct the offset [7],using the estimated value at the receiver. The works presented in this paper concentrate on a quantitative ICI power analysis of the ICI cancellation scheme, which has not been studied previously. The average carrier-tointerference power ratio (CIR) is used as the ICI level indicator, and a theoretical CIR expression is derived for the proposed scheme. II. IDEALIZED SYSTEM MODEL The Fig. 1 describes a simple idealized OFDM system model suitable for a time-invariant AWGN channel.

I.

INTRODUCTION

OFDM is emerging as the preferred modulation scheme in modern high data rate wireless communication systems. OFDM has been adopted in the European digital audio and video broadcast radio system and is being investigated for broadband indoor wireless communications. Standards such as HIPERLAN2 (High Performance Local Area Network) and IEEE 802.11a and IEEE 802.11b have emerged to support IP-based services. Such systems are based on OFDM and are designed to operate in the 5 GHz band. OFDM is a special case of multi-carrier modulation. Multi-carrier modulation is the concept of splitting a signal into a number of signals, modulating each of these new signals to several frequency channels, and combining the data received on the multiple channels at the receiver. In OFDM, the multiple frequency channels, known as sub-carriers, are orthogonal to each other. One of the principal advantages of OFDM is its utility for transmission at very nearly optimum

Fig. 1.Idealized OFDM System Model

In an OFDM system, at the transmitter part, a high data-rate input bit stream b[n] is converted into N parallel bit streams each with symbol period Ts through a serial-toparallel buffer. When the parallel symbol streams are generated, each stream would be modulated and carried over at different center frequencies. The sub-carriers are spaced by 1/NTs in frequency, thus they are orthogonal over the interval

141

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

(0, Ts). Then, the N symbols are mapped to bins of an Inverse Fast Fourier Transform (IFFT). These IFFT [11] bins correspond to the orthogonal sub-carriers in the OFDM symbol. Therefore, the OFDM symbol can be expressed as

1) Cancellation Method In self cancellation scheme the main idea is to modulate the input data symbol on to a group of sub carriers with predefined self coefficients such that the generated ICI

X ( n) =

j 2π nm 1 N −1 Xm N ∑ N m =0

(1)

signals within the group cancel each other. The data pair

(X ,− X )

is

modulated

on

to

two

adjacent

where the X m are the base band symbols on each sub carrier. Then, the X (i ) points are converted into a time domain sequence x (i ) via an IFFT operation and a parallel to serial conversion. The digital-to-analog (D/A) converter then creates an analog time-domain signal which is transmitted through the channel. At the receiver, the signal is converted back to a discrete N point sequence y (n) , corresponding to each subcarrier. This discrete signal is demodulated using an N-point Fast Fourier Transform (FFT) operation at the receiver. The demodulated symbol stream is given by

subcarriers (l , l + 1) . The ICI signals generated by the subcarrier l will be cancelled out significantly by the ICI generated by the subcarrier l + 1 .The signal data redundancy makes it possible to improve the system performance at the receiver side. In considering a further reduction of ICI, the ICI cancellation demodulation scheme is used. In this scheme, signal at the ( k + 1) subcarrier is multiplied by "− 1" and then added to the one at the k subcarrier. Then, the resulting data sequence is used for making symbol decision.

N −1 Y ( m) = ∑ y ( n) e n=0

− j 2π nm N + W (m)k
(2)

2). ICI Cancelling Modulation where w (m) corresponds to the FFT of the samples of w (n), which is the time invariant Additive White Gaussian Noise (AWGN) introduced in the channel Then, the signal is down converted and transformed into a digital sequence after it passes an Analog-to-Digital Converter (ADC). The following step is to pass the remaining TD samples through a parallelto-serial converter and to compute N-point FFT. The resulting Yi complex points are the complex baseband representation of the N modulated sub carriers. As the broadband channel has been decomposed into N parallel sub channels, each sub channel needs an equalizer (usually a 1-tap equalizer) in order to compensate the gain and phase introduced by the channel at the sub channel’s frequency. These blocks are called Frequency Domain Equalizers (FEQ).Therefore the groups of bits that has been placed on the subcarriers at the transmitter are recovered at the receiver as well as the high data-rate sequence. III. ICI SELF CANCELLATION SCHEME A. Self-Cancellation ICI self-cancellation is a scheme that was introduced by Yuping Zhao and Sven-Gustav Häggman[1] in to combat and suppress ICI in OFDM. The main idea is to modulate the input data symbol onto a group of subcarriers with predefined coefficients such that the generated ICI signals within that group cancel each other, hence the name self- cancellation. The ICI self-cancellation scheme requires that the transmitted signals be constrained such that using X (1) = − X (0), X (3) = − X (2),......., X ( N −1) = − X ( N − 2) this assignment of transmitted symbols allows the received signal on subcarriers k and k + 1 to be written as N −2 Y (k ) = ∑ X (l )[S (l − k ) − S (l + 1 − k )] + n (3) k l =0,2,4,6

Y' (k +1) =

N −2 ∑ X(l)[S(l −k −1)−S(l −k)]+nk+1 l =0,2,4,6

(4)

and the ICI coefficient S ’ ( l − k ) reffered as

S’(l-k)=S(l-k)-S(l+1-k)

(5)

Fig 2: Comparison of |S(l-k)|, |S`(l-k)|, and |S``(l-k)| for N = 64 and ε = 0.4

142

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Fig. 2 shows a comparison between |S’(l-k)| and |S(l-k)| on a logarithmic scale. It is seen that |S’(l-k)| << |S(l-k)| for most of the l-k values. Hence, the ICI components are much smaller than they are in |S(l-k)|. Also, the total number of interference signals is halved since only the even subcarriers are involved in the summation. 3). ICI Canceling Demodulation ICI modulation introduces redundancy in the received signal since each pair of subcarriers transmit only one data symbol. This redundancy can be exploited to improve the system power performance, while it surely decreases the bandwidth efficiency. To take advantage of this
th

When compared to the two previous ICI coefficients S (1− k ) for the standard OFDM system and
S (1− k ) for the ICI canceling modulation, S ''(1 − k ) has the smallest ICI coefficients, for the majority of l-k values, followed by S (1− k ) and S (1− k ) .The combined modulation and demodulation method is called the ICI self-cancellation scheme. The reduction of the ICI signal levels in the ICI selfcancellation scheme leads to a higher CIR. The theoretical CIR can be derived as 2 -S(-1)+2S(O)-S(1) (8) CIR= N-1 2 -S(l-1)+2S(l)-S(l+1) ∑ l=2,4,6

redundancy, received signal at the (k + 1) subcarrier, where k
th

is even, is subtracted from the k subcarrier.

As mentioned above, the redundancy in this scheme reduces the bandwidth efficiency by half. This could be compensated by transmitting signals of larger alphabet size. Using the theoretical results for the improvement of the CIR should increase the power efficiency in the system and gives better results for the BER. Hence, there is a tradeoff between bandwidth and power tradeoff in the ICI self-cancellation scheme. The Fig. 4 shows the model of the proposed method.

Fig. 3 An example of S(l - k) for N = 16; l = 0. (a) Amplitude of S(l - k). (b) Real part of S(l - k). (c) Imaginary part of S(l - k). Fig.4. Simulation block diagram of the proposed system

This is expressed mathematically as
Y’’ ( k ) = Y’ ( k ) − Y’ ( k + 1) =

N−2 ∑ X(l)[−S(l −k −1) +2S(l −k) −S(l −k +1)]+nk −nk +1 l=0
Subsequently, the ICI coefficients for this received signal becomes

(6)

S’( l − k) = −S( l − k −1) + 2S( l − k) − S( l − k +1)

(7)

ICI self-cancellation scheme can be combined with error correction coding. Such a system is robust to both AWGN and ICI, however, the bandwidth efficiency is reduced. The proposed scheme provides significant CIR improvement, which has been studied theoretically and by simulations. The scheme also works well in a multipath radio channel with Doppler frequency spread. Under the condition of the same bandwidth efficiency and larger frequency offsets, the proposed OFDM system using the ICI selfcancellation scheme performs much better than standard OFDM systems. In addition, since no channel equalization is needed for reducing ICI, the element without increasing system complexity. Fig. 5 shows the comparison of the theoretical CIR curve of the ICI self-cancellation scheme, calculated by, and

143

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

the CIR of a standard OFDM system is calculated. As expected, the CIR is greatly improved using the ICI selfcancellation scheme. The improvement can be greater than 15 dB for 0 < ε < 0.5.

adding the AWGN yields (13) Y1( k ) = R1( k ) + W1( k ) j 2π ∈ (14) + W2 ( k ) Y2 (k ) = R1 ( k )e k = 0,1,...., N − 1 This maximum likelihood estimate is a conditionally unbiased estimate of the frequency offset and was computed using the received data. The maximum likelihood estimate of the normalized frequency offset is given by:  K Im Y ( k )Y * ( k )  ∑ Λ 2 1 1  −1  (15) tan  k =− K ε = K *  2π ∑ Re Y2 ( k )Y1 ( k )    k =− K 
Once the frequency offset is known, the ICI distortion in the data symbols is reduced by multiplying the received symbols with a complex conjugate of the frequency shift and applying the FFT,

Fig. 5. CIR versus ε for a standard OFDM system

IV. MAXIMUM LIKELIHOOD ESTIMATION The second method for frequency offset correction in OFDM systems was suggested by Moose. In this approach, the frequency offset is first statistically estimated using a maximum likelihood algorithm and then cancelled at the receiver. This technique involves the replication of an OFDM symbol before transmission and comparison of the phases of each of the subcarriers between the successive symbols. When an OFDM symbol of sequence length N is replicated, the receiver receives, in the absence of noise, the 2N point sequence {r (n)} is given by
S ''(1 − k )

X ( n) = FFT {Y (n)e

− j 2π n∈ N }

(16)

V.EXTENDED KALMAN FILTERING A. Problem Formulation A state-space model of the discrete Kalman filter is defined as

z (n) = a (n) d (n ) + v (n )

(17)

(11)

In this model, the observation z(n) has a linear relationship with the desired value d(n). By using the discrete Kalman filter, d(n) can be recursively estimated based on the observation of z(n) and the updated estimation in each recursion is optimum in the minimum mean square sense. The received symbols are

where { x( k )} are the 2k + 1 complex modulation values used to modulate 2 k + 1 subcarriers, H ( K ) is the channel th transfer function for k carrier and ε is the normalized frequency offset of the channel.

Y ( n) = X (n)e

j 2π n '∈( n) N + W (n )

(18)

It is obvious that the observation y(n) is in a nonlinear relationship with the desired value εn. At the receiver (19) Y ( n) = f (∈ (n) + W ( n)

A. Offset Estimation The first set of N symbols is demodulated using an N -point FFT to yield the sequence R1(k), and the second set is demodulated using another N -point FFT to yield the sequence R2(k). The frequency offset is the phase difference between R1(k) and R2(k), that is
Where

j 2π n '∈(n) N f (∈ ( n) = X (n )e

(20)

R 2 ( k ) = R1 ( k ) e

j 2πε

In order to estimate ε efficiently in computation, we build an approximate linear relationship using the first-order Taylor’s expansion:

(12)

144

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

y ( n ) ≈ f (ε ^ ( n −1) ) + f ′ (ε ^ ( n −1) ) ε ( n ) − ε ^ ( n −1)  + w ( n )  

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

(21)

where ε^(n-1) is the estimation of ε(n-1)

f ' (∈ ^ ( n − 1) =

∂f (∈ ( n)) ∂ ∈ (n)
(22)

= j
and z ( n ) = y ( n ) − f

2π n ' N

j 2π n '∈( n−1) N e

2. Compute the H(n), the derivative of y(n) with respect to ε (n) at , the estimate obtained in the previous iteration. 3. Compute the time-varying Kalman gain K(n) using the error variance P(n- 1), H(n), and σ2 4. Compute the estimate y^(n)using x(n) and )ε^(n-1)., i.e. based on the observations up to time n-1, compute the error between the true observation y(n) and y^(n). 5. Update the estimate ε^(n) by adding the K(n)weighted error between the observation y(n) and y^(n) to the previous estimation ε^(n-1). 6. Compute the state error P(n) with the Kalman gain K(n), H(n), and the previous error P(n-1). 7. If n is less than Np, increment n by 1 and go to step 2; otherwise stop. It is observed that the actual errors of the estimation ε^(n) from the ideal value ε(n) are computed in each step and are used for adjustment of estimation in the next step. The pseudo code of computation is summarized as Initialize P(n), ε^(0). For n=1,2….Np compute,

(23) (24) (25) (26)

d ( n ) = ε ( n ) − ε ^ ( n − 1)
and the following relationship:

(ε ^ ( n − 1) )

z ( n ) = f ′ ( ε ( n − 1) ) d ( n ) + w ( n )

which has the same form as, i.e., z(n) is linearly related to d(n). Hence the normalized frequency offset ε can be estimated in a recursive procedure similar to the discrete Kalman filter. As linear approximation is involved in the derivation, the filter is called the extended Kalman filter (EKF). The derivation of the EKF is omitted in this report for the sake of brevity. B. ICI Cancellation There are two stages in the EKF scheme to mitigate the ICI effect: the offset estimation scheme and the offset correction scheme. 1). Offset Estimation Scheme To estimate the quantity ε(n) using an EKF in each OFDM frame, the state equation is built as

H (n) =

∂y ( x )
∂x

(29)

=

j 2π n '
N

j 2π n '∈(n−1) N e X (n)

(30)

K (n) = P(n − 1) H *(n)[ P( n − 1) + σ 2 ]−1
ε ^ ( n ) = ε ^ ( n − 1) + Re{K ( n ) [y ( n ) − x ( n ) e

(31)

ε(n)=ε(n-1)
(27)
i.e., in this case we are estimating an unknown constant ε. This constant is distorted by a non-stationary process x(n), an observation of which is the preamble symbols preceding the data symbols in the frame. The observation equation is

j 2π n '∈( n −1) N } (32)

P ( n ) = 1 − K ( n ) H ( n )  P ( n − 1)  
2). Offset Correction Scheme

(33)

j 2π n '∈(n ) N + W ( n) Y ( n ) = X ( n )e

(28)

where y(n) denotes the received preamble symbols distorted in the channel, w(n) the AWGN, and x(n) the IFFT of the preambles X(k) that are transmitted, which are known at the receiver. Assume there are Np preambles preceding the data symbols in each frame are used as a training sequence and the variance σ of the AWGN w(n) is stationary. The computation procedure is described as follows. 1. Initialize the estimate and corresponding state error P(0).

The ICI distortion in the data symbols x(n) that follow the training sequence can then be mitigated by multiplying the received data symbols y(n) with a complex conjugate of the estimated frequency offset and applying FFT, i.e.

j 2π n '∈ X ^ (n ) = FFT {Y ( n)e N }

(34)

As the estimation of the frequency offset by the EKF scheme is pretty efficient and accurate, it is expected that the performance will be mainly influenced by the variation of the AWGN.

145

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

VI. SIMULATED RESULT ANALYSIS A. Performance In order to compare the three different cancellation schemes, BER curves were used to evaluate the performance of each scheme. For the simulations in this paper, MATLAB was employed with its Communications Toolbox for all data runs. The OFDM transceiver system was implemented as specified by Fig. 1. Frequency offset was introduced as the phase rotation. Modulation schemes of binary phase shift keying (BPSK) and Quadrature amplitude modulation (QAM) were chosen as they are used in many standards such as 802.11a.Simulations for cases of normalized frequency offsets equal to 0.05, 0.15, and 0.30.
Table 6.1: Simulation Parameters
PARAMETERS Number of carriers Modulation Frequency offset No. of OFDM symbols Bits per OFDM symbols Eb-No IFFT size VALUES 768 BPSK,QAM [0,0.15,0.30] 100 N*log2(M) 1:15 1024 Fig. 7: BER Performance with ICI Cancellation, ε=0.30 for 16-BPSK

B. BER Performance Fig. 6 to Fig.11 provides comparisons of the performance of the SC, ML and EKF schemes for different alphabet sizes and different values of the frequency offset.
Fig. 8: BER Performance with ICI Cancellation, ε=0.05 for 64-BPSK

Fig. 6. BER Performance with ICI Cancellation, ε=0.15 for 4-BPSK
Fig. 9: BER Performance with ICI, Cancellation ε=0.15 for 4-QAM

146

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

deteriorate the performance too greatly. However, for QAM with an alphabet of size 2, performance degrades more quickly. When frequency offset is small, the 4-QAM system has a lower BER than the BPSK system. But the BER of 4QAM varies more dramatically with the increase the frequency offset than that of BPSK. Therefore it is concluded that larger alphabet sizes are more sensitive to ICI. Tables 6.2 and 6.3 summarize required values of SNR for BER specified
-6

at 10
Table 6.2 Required SNR and improvement for BER of 10^-6 for BPSK Sl No. Fig. 10: BER Performance with ICI Cancellation, ε=0.30 for 16-QAM 1 SC 2 ML 3 EKF 13 dB 12dB 13.5 dB 12.5 dB 12 dB 12 dB 12 dB 11dB 13dB

Method

ε= 0.05

ε= 0.15

ε= 0.30

Table 6.3 Required SNR and improvement for BER of 10^-6 for QAM Sl. No. 1 SC 2 ML 3 EKF 12dB 13 dB 14 dB 11 dB 12 dB 13dB 13 dB 12 dB 11 dB

Method

ε= 0.05

ε= 0.15

ε= 0.30

Fig. 11. BER Performance with ICI Cancellation, ε=0.05 for 64-QAM

It is observed in the figures that each method has its own advantages. In the presence of small frequency offset and binary alphabet size, self cancellation gives the best results. However, for larger alphabet sizes and larger frequency offset such as 4-BPSK and frequency offset of 0.30, self cancellation does not offer much increase in performance. The maximum likelihood method gives the best overall results. The Kalman filter method indicates that for very small frequency offset, it does not perform very well, as it hardly improves BER. However, for high frequency offset the Kalman filter does perform extremely well. It gives a significant boost to performance. Significant gains in performance can be achieved using the ML and EKF methods for a large frequency offset. These results also show that degradation of performance increases with frequency offset. For the case of BPSK, even severe frequency offset of 0.30 does not

For small alphabet sizes and for low frequency offset values, the SC and ML techniques have good performance in terms of BER. However, for higher order modulation schemes, the EKF and ML techniques perform better. This is attributed to the fact that the ML and EKF methods estimate the frequency offset very accurately and cancel the offset using this estimated value. However, the self-cancellation technique does not completely cancel the ICI from adjacent sub-carriers, and the effect of this residual ICI increases for larger alphabet sizes and offset values. VII.CONCLUSION In this paper, the performance of OFDM systems in the presence of frequency offset between the transmitter and the receiver has been studied in terms of the Carrier-toInterference ratio (CIR) and the bit error rate (BER) performance. Inter-carrier interference (ICI) which results from the frequency offset degrades the performance of the OFDM system.

147

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Three methods like ICI self-cancellation (SC), maximum likelihood (ML) estimation and Kalman filtering (EKF) methods were explored in this paper for mitigation of the ICI. The cancellation of the frequency offset has been investigated in this paper and compared with these three techniques. The choice of which method to employ depends on the specific application. For example, self cancellation does not require very complex hardware or software for implementation. However, it is not bandwidth efficient as there is a redundancy of bits for each carrier. The ML method also introduces the same level of redundancy but provides better BER performance, since it accurately estimates the frequency offset. Its implementation is more complex than the SC method. On the other hand, the EKF method does not reduce bandwidth efficiency as the frequency offset can be estimated from the preamble of the data sequence in each OFDM frame. However, it has the most complex implementation of the three methods. In addition, this method requires a training sequence to be sent before the data symbols for estimation of the frequency offset. It can be adopted for the receiver design for IEEE 802.11a because this standard specifies preambles for every OFDM frame. The preambles are used as the training sequence for estimation of the frequency offset. In this paper, the simulations were performed in an AWGN channel. This model can be easily adapted to a flatfading channel with perfect channel estimation. Further work can be done by performing simulations to investigate the performance of these ICI cancellation schemes in multipath fading channels without perfect channel information at the receiver. In this case, the multipath fading may encumber the performance of these ICI cancellation schemes.
REFERENCES [1] Y. Zhao and S. Häggman, “Inter carrier interference selfcancellation scheme for OFDM mobile communication systems,” IEEE Transactions on Communications, vol. 49, no. 7, 2001 P. H. Moose, “A Technique for Orthogonal Frequency Division Multiplexing Frequency Offset Correction,” IEEE Transactions on Communications, vol. 42, no. 10, 1994 R. E. Ziemer, R. L. Peterson, ”Introduction to Digital nd Communications”, 2 edition, Prentice Hall, 2002. J. Armstrong, “Analysis of new and existing methods of reducing inter carrier interference due to carrier frequency offset in OFDM,” IEEE Transactions on Communications, vol. 47, no. 3, pp. 365 – 369., 1999 N. Al-Dhahir and J. M. Cioffi, “Optimum finite-length equalization for multicarrier transceivers,” IEEE Transactions on Communications, vol. 44, no. 1, pp. 56 – 64, 1996 W. G. Jeon, et al, “An equalization technique for orthogonal frequency-division multiplexing systems in time-variant multipath channels,” IEEE Transactions on communications, vol. 47, no. 1, pp. 27 – 32, 2001. Y. Fu, S. G. Kang, and C. C. KO, “A new scheme for PAPR reduction in OFDM systems with ICI selfcancellation,” in Proc. VTC 2002- Fall, 2002 IEEE 56th Vehicular Technology Conf., vol. 3, pp 1418– 1421, 2002.

J.-J. van de Beek, M. Sandell, and P.O Borjesson, “ML estimation of time and frequency offset in OFDM systems,” IEEE Trans. Signal Process., 45, pp.1800–1805, 1997. [8] Tiejun (Ronald) Wang, John G. Proakis, and James R.Zeidler “Techniques for suppression of intercarrier interference in ofdm systems”. Wireless Communications and Networking Conference, IEEE Volume 1,Issue, 13-17 pp: 39 - 44 Vol.1, 2005. [9] X.Cai, G.B.Giannakis,”Bounding performance and suppressing intercarrier interference in wireless mobileOFDM”, IEEE Transaction on communications, vol.51, pp.2047-2056, no.12, Dec.2003. [10] J. G. Proakis, Digital Communications, 2nd ed. New York: McGraw-Hill, 1989. [11] William H.Tranter, K.Sam Shanmugam, Theodore S.Rappaport, Kurt L.Kosbar, “Principles of Communication system simulation with wireless application”, Pearson Education, 2004. AUTHORS PROFILE B.Sathish Kumar currently is a Senior Lecturer, Sri Ramakrishna Engineering College, Coimbatore, Tamil Nadu, India. He received the Masters Degree from Anna University Chennai, in the year 2005 and currently pursuing PhD in Anna University Coimbatore. His research interest includes Wireless Communication, Networking, Signal Processing, Mobile Communication and Multicarrier Communication.

[7]

K.R.Shankar Kumar currently is a Professor, Sri Ramakrishna Engineering College, Coimbatore, Tamil Nadu, India. He received the Masters Degree from Madras University, in the year 2000 and the PhD from Indian Institute of Science, Bangalore, in the year 2004. His research interest includes future broad band wireless communication, Multicarrier Communication systems, Digital Communication, Advanced Signal Processing for communication. He has published more than 20 Journal papers in the field of CDMA systems. His research work was supported by Swarnajayanti Fellowship, Department of Science and Technology (DST), Government of India. Radhakrishnan Rathinavel is currently Professor, Sri Ramakrishna Engineering College, in Electronics and Communication Engineering Department, Coimbatore, Tamil Nadu, India. He received his Masters Degree from P.S.G.College of Technology, Coimbatore, in the year 1997 and the PhD from Anna University Chennai in the year 2008. His research interest includes Wireless Communication, Signal Processing, Networking and Mobile Communication. He has published more than 11 Journal papers in the field of CDMA systems, Mobile communication, Wireless Networking and Signal Processing.

[2]

[3] [4]

[5]

[6]

148

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

HIGH-PRECISION HALF-WAVE RECTIFIER CIRCUIT IN DUAL PHASE OUTPUT MODE
Theerayut Jamjaem
Department of Electrical Engineering Faculty of Engineering, Kasem Bundit University Bangkok, Thailand 10250 .

Bancha Burapattanasiri
Dept. of Electronic and Telecommunication Engineering Faculty of Engineering, Kasem Bundit University Bangkok, Thailand 10250 . signal and negative part signal. Common source circuit, it has functional to sending positive part signal to current mirror circuit, and current mirror circuit, it has functional to establish half-wave positive and negative phase signal. For easy to understand, see diagram block in figure 1. The functional of circuit started when sending input positive and negative part signal to CMOS inverter circuit for comparison between negative ground part signal and positive to common source circuit and part to sending current mirror circuit 1, it has functional to establish half-wave positive phase signal. At that time positive part signal from common source circuit, it will send to current mirror circuit 2, it has functional to establish half-wave negative phase signal.

Abstract—This paper present high-precision half-wave rectifier circuit in dual phase output mode by 0.5 µm CMOS technology, + 1.5 V low voltage, it has received input signal and sent output current signal, respond in high frequency. The main structure compound with CMOS inverter circuit, common source circuit, and current mirror circuit. Simulation and confirmation quality of working by PSpice program, then it able to operating at maximum frequency about 100 MHz, maximum input current range about 400 µAp-p, high precision output signal, low power dissipation, and uses a little transistor. Keywords-component; precession, dual phase. half-wave, rectifier circuit, high-

I.

INTRODUCTION

Rectifier circuit is important circuit in analog working such as AC meter, detection signal circuit, and analog adaptation working circuit etc. And then it always has development and designing about rectifier circuit voltage mode. At the beginning used vacuum tube, diode [1-3] and next time used bipolar transistor [4]. At the beginning of rectifier circuit by diode and bipolar transistor, it has zero-crossing signal error and low precision. Because of diode and bipolar transistor have to use voltage driver working power around 0.3 V for Ge and 0.7 V for Si. So, the lower signal circuit unable to working. Next time, the limited of lower signal circuit has development such as rectifier circuit by Op-amp connected to diode, rectifier circuit by Op-amp connected to diode and bipolar transistor, and rectifier circuit AB class mode [9]. The result from circuit development is able to rectify the limited of lower signal circuit [1-5], responsiveness at narrowness frequency [2-7] and high dissipation current source to transistor. So, this paper is present a new choice of easy to understand and noncomplex of rectifier circuit, but it still have high quality in working function, it able working at high frequency, responding at high input current, high precision signal, low power dissipation, and uses a little transistor. II. DESIGNATION AND FUNCTIONAL Three part component of the high-precision half-wave rectifier circuit in dual phrase output mode that is CMOS inverter circuit, it has functional to comparison positive part
Identify applicable sponsor/s here. (sponsors)

Figure 1. Diagram is show high-precision half-wave rectifier circuit in dual phase output mode.

From figure 1 is able to bring three part of circuit put together after that there are complete high-precision half-wave rectifier circuit in dual phase output mode as figure 2 and when you see from the structure of circuit by sending input current signal, it has equivalent to little than zero as an algebraic equation (1) after that is reflect to M3 and M4 transistor stop working, output voltage return to input of M1 transistor common source circuit current the result is IDM1 equivalent to

149

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

Iin, but in directly opposite input current (Iin) equivalent to more zero is reflect to M1,M3 transistor stop working but M2, M4 are working, so the result is IDM1 equivalent to zero as an algebraic equation (2).

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

I DM 2 = 0 And I DM 1 = I in When I in < 0
I DM 1 = 0 When I in > 0

(1) (2)

400 µAp-p, frequency 100 kHz, then the result is output signal as figure 5, at input current signal at 400 µAp-p, frequency 1 MHz, then the result is output signal as figure 6, at input current signal at 400 µAp-p, frequency 10 MHz, then the result is output signal as figure 7, at input current signal at 400 µAp-p, frequency 100 MHz, then the result is output signal as figure 8, Output signal at input current signal at 400 µAp-p, frequency 10 MHz, temperature 25°, 50°,75°and 100O as figure 9 and characteristic DC current at 400 µAp-p input current, temperature 25°, 50°,75°and 100O as figure 10.

Figure 2. Completely high-precision half-wave rectifier circuit in dual phase output mode.

While M1 transistor current is working, the signal current it has equivalent to little than zero is equivalent current it has M5 transistor at drain pin. It has reflected current pass to transistor M6, so the characteristic of output signal is half-wave negative phase as an algebraic equation (3) at the same time signal current at drain pin of M7 transistor it has passed to drain pin of M8 transistor, so M8 and M9 transistor are bonding in mirror current. So the characteristic of output signal is half-wave positive phase as an algebraic equation (4).

Figure 3. Output signal at input current 400 µAp-p and frequency 1 kHz

I DM 1 = I DM 5 = I DM 6 = − I out
I DM 7 = I DM 8 = I DM 9 = + I out
III. SIMULATION AND MEASUREMENT RESULT

(3) (4)

For confirmation of circuit working function uses PSpice program in testing, so we have to fix MIETEC parameter 0.5 µm for PMOS and NMOS transistor, power supply VDD = 1.5 V, VSS = -1.5 V, by sending input current signal at 400 µAp-p, frequency start from 1 kHz – 100 MHz and to setting W/L = 1.5/0.15 µm, at input current signal at 400 µAp-p, frequency 1 kHz, then the result is output signal as figure 3, at input current signal at 400 µAp-p, frequency 10 kHz, then the result is output signal as figure 4, at input current signal at

Figure 4. Output signal at input current 400 µAp-p and frequency 10 kHz

150

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Figure 5. Output signal at input current 400 µAp-p and frequency 100 kHz

Figure 8. Output signal at input current 400 µAp-p and frequency 100MHz

Figure 6. Output signal at input current 400 µAp-p and frequency 1 MHz

Figure 9. Output signal at input current 400 µAp-p and frequency 10 MHz at temperature 25O, 50O,75O and 100O

Figure 7. Output signal at input current 400 µAp-p and frequency 10 MHz

Figure 10. Output characteristic DC current at 400µAp-p p input current, temperature 25O,50O, 75Oand 100O

151

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

IV.

CONCLUSION

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
[5]

This present is show the circuit, it has component with a little of transistor, noncomplex in working function, dissipation of current source, working at input current mode, output signal is half-wave rectifier circuit in dual phase, without the reflection of temperature with + 1.5 V low voltage and it not chance the structure of circuit. The result of testing it able to guarantee a quality of working function at maximum frequency 100 MHz, maximum output current 400 µAp-p, losses power 198 pW. So, it suitable to uses in development VLSI compound current technology and apply in analog signal processing. APPENDIX The parameters used in simulation are 0.5 µm CMOS Model obtained through MIETEC [10] as listed in Table I. For aspect ratio (W/L) of MOS transistors used are as follows: 1.5 µm / 0.15 µm for all NMOS transistors; 1.5 µm / 0.15 µm for all PMOS transistors.
TABLE I. CMOS MODEL USED IN THE SIMULATION

Toumazou, C., Lidgey, F.J.and Chattong, S., “High frequency current conveyor precision full-wave rectifier,” Electron. Letts, 10: Vol. 30, pp. 745-746, 1994. [6] Wilson, B. and Mannama, V., “Current-mode rectifier with improved precision,” Electron. Letts, 4: Vol. 31, pp. 247-248, 1995. [7] Surakampontorn, W. and Riewruja, V., “Integrable CMOS sinusoidal frequency doubler and full-wave rectifier,” Int.J. Electronics, Letts, 3: Vol. 73, pp. 627-632, 1992. [8] Traff, H., “Novel Approach to High Speed CMOS Current Comparators,” Electron. Letts, 3: Vol. 28, pp. 310-312, 1992. [9] Monpapassorn, A., “Improved Class AB Full-Wave rectifier,” Thammasat Int. J. Sc. Tech., No. 3, November, Vol. 4, 1999. [10] A. Monpapassorn, K. Dejhan, and F.Cheevasuvit, “CMOS dual output current mode half-wave rectifier,” International Journal of Electronics, Vol. 88, 2001, pp. 1073-1084.

AUTHORS PROFILE
Mr.Theerayut Janjaem received the master degree in Telecommunication Engineering, from King Mongkut’s Institute of Technology Ladkrabang in 2005. He is a lecture of Electrical Engineering Faculty of Engineering, Kasem Bundit University, Bangkok, Thailand. His research interests are energy, analog circuit design, low voltage, high frequency and high-speed CMOS technology.

---------------------------------------------------------------------------------------------.MODEL CMOSN NMOS LEVEL = 3 TOX = 1.4E-8 NSUB = 1E17 GAMMA = 0.5483559 PHI = 0.7 VTO = 0.7640855 DELTA = 3.0541177 UO = 662.6984452 ETA = 3.162045E-6 THETA = 0.1013999 KP = 1.259355E-4 VMAX = 1.442228E5 KAPPA = 0.3 RSH = 7.513418E-3 NFS = 1E12 TPG = 1 XJ = 3E-7 LD = 1E-13 WD = 2.334779E-7 CGDO = 2.15E-10 CGSO = 2.15E-10 CGBO = 1E-10 CJ = 4.258447E-4 PB = 0.9140376 MJ = 0.435903 CJSW = 3.147465E-10 MJSW = 0.1977689 .MODEL CMOSP PMOS LEVEL = 3 TOX = 1.4E-8 NSUB = 1E17 GAMMA = 0.6243261 PHI = 0.7 VTO = -0.9444911 DELTA = 0.1118368 UO = 250 ETA = 0 THETA = 0.1633973 KP = 3.924644E-5 VMAX = 1E6 KAPPA = 30.1015109 RSH = 33.9672594 NFS = 1E12 TPG = -1 XJ = 2E-7 LD = 5E-13 WD = 4.11531E-7 CGDO = 2.34E-10 CGSO = 2.34E-10 CGBO = 1E-10 CJ = 7.285722E-4 PB = 0.96443 MJ = 0.5 CJSW = 2.955161E-10 MJSW = 0.3184873 ----------------------------------------------------------------------------------------------

Mr.Bancha Burapattanasiri received the bleacher degree in electronic engineering from Kasem Bundit University in 2002 and master degree in Telecommunication Engineering, from King Mongkut’s Institute of Technology Ladkrabang in 2008. He is a lecture of Electronic and Telecommunication Engineering, Faculty of Engineering, Kasem Bundit University, Bangkok, Thailand. His research interests analog circuit design, low voltage, high frequency and high-speed CMOS technology.

ACKNOWLEDGMENT The researchers, we are thank you very much to our parents, who has supporting everything to us. Thankfully to all of professor for knowledge and a consultant, thank you to Miss Suphansa Kansa-Ard for her time and supporting to this research. The last one we couldn’t forget that is Kasem Bundit University, Engineering Faculty for supporting and give opportunity to our to development in knowledge and research, so we are special thanks for everything. REFERENCES
[1] [2] [3] [4] Barker, R.W.J., “Versatile precision full wave Rectifier,” Electron Letts, 5: Vol.13, pp. 143-144, 1977. Barker, R.W.J. and Hart, B.L., “Precision absolute-value circuit technique,” Int. J. Electronics Letts, 3: Vol.66, pp. 445-448, 1989. Toumazou, C. and Lidgey, F.J., “Wide-Band precision rectification,” IEE Proc. G, 1: Vol.134, pp.7-15, 1987. Wang, Z., “Full-wave precision rectification that is performed in current domain and very suitable for CMOS implementation,” IEEE Trans. Circuits and Syst, 6: Part I, Vol. 39, pp.456-462, 1992.

152

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Internal Location Based System For Mobile Devices Using Passive RFID And Wireless Technology
A.D.Potgantwar
Lecturer in Computer Engg SITRC Nashik India .

Vijay M.Wadhai
Prof & Dean Research MITSOT, MAE, Pune India .

Abstract— We have explored our own innovative work about the design & development of internal location-identification system for mobile devices based on integration of RFID and wireless technology. The function of our system is based on strategically located passive RFID tags placed on objects around building which are identified using an RFID reader attached to a mobile device. The mobile device reads the RFID tag and through the wireless network, sends the request to the server. The server resolves the request and sends the desired location-based information back to the mobile device. We had addressed that we can go through the RFID technology for internal location identification (indoor), which provides us better location accuracy because of no contact between the tag and the reader, and the system requires no line of sight. In this paper we had also focused on the issues of RFID technologies i.e. Non-line–of-sight & High inventory speeds. Keywords- Location Based Services, RFID, J2ME

In addition the short range of the RFID readers used ensures that the system is able to determine position to high degree of accuracy. The architecture proposed is flexible and could be deployed in a range of application domains including tourism, inventory tracking and security access. Section 2 reviews some of the related work carried out on location identification and positioning systems. Section 3 provides background information on the RFID technology. Section 4 describes the design, architecture and implementation of the location-based RFID system. Section 5 concludes the paper and draws direction to future work. II. RELATED WORK

I.

INTRODUCTION

Rapid advances in wireless technologies and the mobile devices provided an opportunity to develop and deliver new types of location-based application and services to users. Different approaches in location identification can be divided into two main sub categories, internal and external location determination. Cell-ID (Cell Identification), TOA (Time Of Arrival), OTD(Observed Time Difference), GPS(Global Positioning System) are the wireless technologies that are being used for external positioning systems, but GPS is the most used preference for external positioning system, whereas, WLAN(Wireless Local Area Networks), Bluetooth and RFID(Radio Frequency Identification) that are being used for internal positioning system [5]. However the majority of the current mobile location systems (MLS) lack sufficient accuracy and there are still open challenges with respect to design, usability, functionality and implementation aspects. This research attempts to address some of these issues and present an indoor positioning system architecture based on a combination of wireless, RFID and J2ME (Java 2 Micro Edition) technology. The approach has been developed using J2ME.This prototype used a java application with a RFID reader attached to a mobile device in order to locate the user’s position in a building with fixed passive RFID tags attached.

Existing approaches to location identification can be divided into two main sub categories, external and internal location determination. External positioning systems are usually GPS based (Global positioning systems) or operationally dependant on the augmentation and utilization of existing infrastructure e.g. location of mobile phone masts. GPS uses satellites and works by calculating the time it takes a signal to travel from a satellite to a receiver on a handheld device. Accuracy to within a few meters is achievable using differential GPS. However this approach can be time consuming and unreliable as the GPS receiver needs to be able to communicate with at least four satellites before location can be found. In addition the receiver must maintain a line-of-site transmission with the satellites. As a direct consequence GPS does not work well in built-up areas such as large cities and is not accessible indoors [6]. Different approaches related to internal positioning systems have been used, some of them being briefly discussed in the following. Active Badge developed at Olivetti research laboratories used infrared technology for indoor location positioning but encountered two major limitations based on line-of-sight requirements and short-range signal transmission [7]. Placelab runs an application on the user’s local device to infer current location by using the known longitude and latitude and unique identifier of existing fixed Wi-Fi hotspots, Bluetooth nodes or GSM stations. This observational approach preserves user privacy as the majority of network access points periodically

153

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

broadca their presence and monitoring the appropriate frequencies allows these signals to be intercepted and utilized for this purpose. The functionality of Placelab is limited in that it can identify the user’s location but then relies on the development of a complementary application to utilize this information effectively. Placelab's accuracy varies widely and is typically within 150 meters using GSM coverage and in the range of 20-50m using Wi-Fi [10]. Due to the limitations of these technologies, RFID has emerged as a more attractive alternative. Radio frequency identification (RFID) is a generic term that is used to describe a system that transmits identity in the form of a unique serial number of an object or person wirelessly using radio waves. This wireless system allows for non-contact reading of RFenabled tags. RFID can be applied to the development of applications for tourism, libraries, health centres, security access, student tracking, etc. The significant advantage of RFID systems is the non-contact and the non-line-of-sight nature of the technology. RFID systems typically consist of a number of components including RFID tags, RFID readers, antennas and system software. There are several approaches in location-sensing systems that use RFID technology. SpotON [8] uses RFID technology for three dimensional location sensing based on radio signal strength analysis. Another approach is mTag [3] which is a distributed eventdriven architecture for determining location specific mobile web services. The mTag architecture uses fixed RFID readers placed around the building and they are touched with a passive RFID tag attached to a mobile phone or PDA. Our proposal uses a different approach from the mTag architecture which is using fixed RFID tags with RFID readers attached to mobile devices. One of the limitations of this approach is that since it needs specific requirements for execution, it can not be run on any mobile device. LANDMARK [6] approach uses active RFID tags for indoor location sensing. The major advantage of LANDMARK is that it employs the idea of having extra fixed location reference tags to improve the overall accuracy of locating objects. Our proposal uses fixed passive RFID tags which are strategically located around buildings and are identified by an RFID reader attached to a mobile device.and Wal-Mart Stores. RFID describes any system of identification wherein an electronic device that uses radio frequency or magnetic field variations to communicate is attached to an item. Due to the nature of the RFID technology which is very applicable in locating and tracking objects, recently it has been applied to many location identification systems. III. RFID TECHNOLOGY

presence and identity, consider the simple scenario depicted in Figure2. In this figure2 the RFID reader transmits radio signals at a preset frequency and interval (usually hundreds of times every second). Any radio frequency tags that are in the range of this reader will pick up its transmission because each

Figure 1:- Components of RFID System [2]

has a built-in antenna that is capable of listening to radio signals at a preset frequency (the size and shape of the antenna determine what frequencies it will pick up). The tags use energy from the reader's signal to reflect this signal back. Tags may modulate the signal to send information, such as an ID number, back to the reader. In-short, in a typical RFID System, tags are attached to objects to be technologies that make use of radio waves to automatically identify individual objects. “Radio frequency identification (RFID) is a generic term that is used to describe a system that transmits identity in the form of a unique serial number of an object wirelessly using radio waves.” A.Components of RFID System An RFID system is composed of three main elements: • RFID Tag (Transponder) : RFID tag or transponder, which usually holds an identification number and is located on the objects to be identified. • RFID Reader (Interrogator): RFID reader or interrogator, which detects tags and reads from and writes to the tags. RFID Antenna: A coil of wound copper wire, which emits radio frequency signals. It also acts as a receiver.

•

RFID (Radio Frequency Identification) technology is not a new technology, RFID technology has been around for at least 50 years. But the advancements necessary for it to enter the corporate mainstream have only been made recently. The cost has finally sufficiently dropped, and read/transmit range has increased. RFID technology has been used in many organizations and agencies such as the U.S. Department of Defense (DoD), the Food and Drug Administration (FDA), To understand how an RFID tag notifies a reader about its

B. Working The RFID tag is located on the objects to be identified and the RFID reader detects tags and reads from and writes to the tags. The reader can then inform another system about the presence of the tagged items. The system with which the reader communicates usually runs software that stands between readers and applications. This software is called as RFID middleware. identified. Each tag has a certain amount of memory to store information about the object, such as its unique tag ID (serial number), or in some cases, more details, e.g. manufacturing date, expiry date etc. When these tags pass through an electromagnetic field generated by the reader, they transmit this information back to the reader, thereby enabling object identification.

154

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

When these tags pass through an electromagnetic field generated by the reader, they transmit this information back to the reader, thereby enabling object identification. C. RFID Tags a. Types of Tag There are two types of tags: Passive Tags: Passive tags do not have internal battery and powered by the signal strength emitted by the reader. That means passive tags obtain their operating power from the field generated by the reader. Active Tags: Active tags are larger, more expensive and powered by an internal battery.

Many basic operations can be performed with an RFID tag, but only two of them are universal. Attaching the tag: Any RFID tag must be attachable to an item in some way. Reading the tag: Any RFID tag must be able to communicate information over some radio frequency in some way.
Many tags also offer one or more of the following features and capabilities: Following table1 shows the RFID Tag comparison.
Table 1:- RFID Tag Comparison

Attributes Tag Power Source Tag Battery Availability of Power Required Signal Strength to tag Range Data Storage

Active RFID Internal to tag Yes

Passive RFID Energy Transferred Using No Only In Field Of Reader

Continuous

Very Low Up to 100m Up to 128Kb

Very High Up to 3-5m, Usually Less 128bytes

Figure 2:-Communication between RFID tags and reader.

Kill/disable: Some tags allow a reader to command them to cease functioning permanently. After a tag receives the correct "kill code," it will never respond to a reader again. Write once: Many tags are manufactured with their data permanently set at the factory, but a write-once tag may be set to a particular value by an end user one time. After that, the tag cannot be changed except, possibly, to be disabled. Write many: Some tags can be written and rewritten with new data over and over. Anti-collision: When many tags are in close proximity, a reader may have difficulty telling where one tag's response ends and another's begins. Anti-collision tags know how to wait their turn when responding to a reader. Security and encryption: Some tags are able to participate in encrypted communications, and some will respond only to readers that can provide a secret password. Standards compliance: A tag may comply with one or more standards, enabling it to talk to readers that also comply with those standards; or, in the case of standards for physical
characteristics, a tag may fit in a particular standard receptacle.

b. Classes of Tags There are four classes of tags: CLASS 0 (Read Only Tags) these are the simplest type of tags, where the data, which is usually a simple Tag ID, is stored only once into the tag during manufacture. CLASS 1 (Write Once Read Only (WORM)) these can be factory or user programmed. In this case data can then either be written by the tag manufacturer or by the user only once. CLASS 2 (Read Write) Data can be read as well as written into the tag’s memory. They contain more memory space than what is needed for just a simple ID number. CLASS 3 (Read Write with on board sensors) These are active tags which may contain sensors for recording parameters like temperature, pressure etc. and can record the readings in tag memory. CLASS 4 (Read Write with integrated transmitters) these tags can communicate with each other without any help from reader. Basic Tag Capabilities

One of its most important attributes is that these tags do not require line of sight to be read. Passive tags are quite smaller and less expensive than active tags. They do not have any

155

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

batteries and hence obtain their operating power from the reader. However, passive tags require a higher-powered reader and support shorter read ranges than active tags. RFID tags are programmable and can store a variety of information including location, destination and product information number. Tags can also be read through a variety of substances such as snow, fog, ice, paint, crusted grime, and other visually and environmentally challenging conditions, where barcodes or other optically read technologies would be useless.

reader may be acceptable for an application in one region of the globe but not in another. Once identified, a reader may read data from or write to tag memory, depending on the permissions granted by the tag. RFID readers generally fall into two categories - high frequency (HF) and ultra-high frequency (UHF).Table2 shows
a comparison between HF and UHF RFID technology.
Table 2:- Comparison of HF and UHF RFID Technology

Frequency

HF RFID 13.56 MHz

Read Range Read Rate Memory Size Power Source Advntage

10-20 cm 50 tags/sec 64-256 bit read/write Inductive magnetic Field Low Cost Standard Frequency

UHF RFID 902-928 MHz N.America 860-868 MHz Europe 950-956 MHz Japan 3-6 meters 400 tags/sec 64-2048 bits read/write Capacitive/Electric Field

High Speed Longer read range

Figure 3.Typical RFID Tags [2]

D. RFID Reader Since the reader communicates with tags using RF, any RFID reader must have one or more antennas. Because a reader must communicate with some other device or server, the reader must also have a network interface of some sort. Examples of common network interfaces are the serial Universal Asynchronous Receiver/Transmitters (UARTs) or RS 232 or RS 485 communications and the RJ45 jack for 10BaseT or 100BaseT Ethernet cables; some readers even have Bluetooth or wireless Ethernet communications built in. Finally, to implement the communications protocols and control the transmitter, each reader must have either a microcontroller or a microcomputer. Following figure4 shows the physical components of an RFID reader.

The most noteworthy advantage of RFID is that there is no contact between the tag and the reader, and the system requires no line of sight. The tags can be read through a variety of substances and surfaces. They can also be read at extremely high speeds. RFID technology presents many advantages, over other technologies, that signify its suitability for developing internal location-based systems: Non-line-of-sight: RFID tags have the advantage that they can be read without line of sight through nonconducting materials. This can save time in processing that would otherwise be spent lining up items. High inventory speeds: Multiple items can be scanned at the same time. As a result, the time taken to count the items drops substantially. RFID tags can be read in less than 100 milliseconds. Variety of form factors: RFID tags come in a wide variety of shapes and sizes. These different form factors allow RFID technologies to be used in a broad range of environments. Rewritable tags: The read/write capability of an active RFID system is also a significant advantage in interactive applications such as work-in-process or maintenance tracking. In the case of reusable container, this can be a big advantage.

Figure 4: Physical Components of Reader [2]

a. Types of Reader Readers, like tags, differ in many ways, and no one reader is a perfect fit for all occasions. Readers come in many shapes and sizes, support different protocols, and often must conform to regulatory requirements, which means that a particular

If we enter unknown premises it will be very difficult for us to identify a particular shop or a particular area or place in that building. In order to avoid the confusion and difficulties in finding the exact place, we have to get the proper information regarding the different places in that building. To overcome this, passive RFID tags are strategically located around building, for example each classroom/cabin has one passive RFID tag that hold unique identification number. The classroom/cabin’s information is stored in the server with the

156

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

corresponding tag number. When user along with PDA or mobile device (with attached RFID reader) comes nearer to the corresponding tags the area’s information will be displayed in the mobile. The user then can send the request to the server, regarding further more information so that some more information from the server will be popped out to the mobile phone as per the user’s request. IV. ARCHITECTURE & IMPLEMENTATION

Then PDA/Mobile device (J2ME emulator) sends the reuest to the server through WiFi or other wireless network. Server resolves the request and sends answer (i.e. location) Back to the PDA/mobile device through WiFi or other wireless network. The client side contents RFID reader, RFID middleware and client application. The server side contents web server and database.

The system architecture is illustrated in figure 5. This figure presents an internal location identification system based on the combination of the wireless and RFID technology. This figure shows the infrastructure and components of the system to be developed. It also depicts the functionality of the application: the active RFID reader with CF (Compact Flash) interface is attached to the mobile device or PDA will detect the RFID tags and will send request through the wireless network to the server. The server resolves the request and sends the answer back to PDA.

Figure 6: The communication model of the system.

The client application is developed using Java 2 Micro Edition (J2ME). Active RFID READER-B SL # 702 that can detect Class 0 tags is connected to the desktop machine through serial TTL interface. The RFID Middleware which is developed using Java takes tag_ID from the reader and gives it to the Client Application running on the The Sun Java Wireless Toolkit which is a software tool that emulates a physical mobile device on a desktop computer. Then client application send request to the server through the wireless network (I used D-Link DIR300 wireless G router for the server and D-Link DWL- G510 wireless desktop adapters for the clients). Web server is built using Apache Tomcat and Java Servlets. Before starting to use this application, the user should connect to the client application. When RFID tag come in the range of the RFID reader then RFID reader reads the RFID tag and client application sends the location ID i.e. tag-ID to the server. The server application then maps the tag_ID with the corresponding entry in the database and retrieves the current location of the user and other related information. Finally, this information is sent back to the user through a wireless network. As a result, the user is presented with information about his/her current location and interest specific info concerning the sight. B. Experimental Result 1. First, RFID Reader detects the RFID Tag. • The active RFID (RFID READER-B SL # 702) reader is connected to the PC through serial TTL interface. • This operates in 135 KHz frequency.

Fig 5: System Architecture

A. Implementation Methodology Adapted Mobile RFID readers are expensive so that we go for implementation using active stationary RFID Reader (RFID READER-B SL # 702). The communication model is as given below. Let us see how ILBSFMDURFID System works. RFID Reader detects the RFID Tag. RFID Middleware takes RFID Tag-Id from RFID Reader and sends it to the Client Application which is running on the PDA/mobile device (J2ME emulator).

157

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

• •

Its read range is 3-4” (10-15 cm). It can read Class-0 tag i.e. passive Read Only tags.
Figure 8. Sun Java WTK and Client Application 5. Following Figure9 shows the login form of the client application, user need to enter username, password and server’s IP Address.

2. RFID Middleware takes RFID Tag-Id from RFID Reader and sends it to the Client Application which is running on the PDA/mobile device (J2ME emulator). Following figure shows middleware screen showing tag_id (110055B53A) detected by the RFID reader.

Figure 7 RFID Middleware

3. Then PDA/Mobile device sends the request to the server through WiFi or other wireless network, server resolves the request and send answer back to the user. 4. Figure 8 shows the screen of the mobile emulator: user first need to launch the application.

Figure 9. Client Application login page

6 See Figure10:- client application is showing current location of the user. This application checks for update after every 2 seconds that means when RFID Reader detects new RFID Tag (enters into the zone of new RFID tag) then client application updates its form. 7. See figure18 below- Client Application showing sight image. In this way our experimental result shows that the system is able to accurately locate user inside the building. This system is based on the integration of J2ME, RFID and wireless technology, using these technologies we were able to provide users with information in different formats including text and images.

158

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

V.

CONCLUSION

This paper has presented an architecture and prototype application for delivering internal location-based services for mobile devices using passive RFID technology. The system presented was based on a combination of wireless and RFID technology and was able to accurately locate user and send information based on their location. Using the J2ME for the client application, we were able to provide users with information in different formats including text, image. The most noteworthy advantage of this system is that there is no contact between the tag and the reader, and the system requires no line of sight. The tags can be read through a variety of substances and surfaces. They can also be read at extremely high speeds. Zero power consumption of the passive tags is the key strength of this system. The reduction of price in RFID tags and readers could lead to extensive development of accurate systems and encourage businesses to use it more and more. The problem with RFID systems is that a tag might not be read, in spite of being in the reader’s range, due to collisions, this problems need to be resolved to provide efficient solution for tag identification. REFERENCES
[1] V. Zeimpekis, G.M. Giaglis, and G. Lekakos, “A taxonomy of indoor and outdoor positioning techniques for mobile location services”, ACM SIGecom Exchanges, ACM, New York, USA, 2003, Vol.3 No.4, pp. 1927. B. Glover, H. Bhatt, RFID Essentials, O’Reilly, January 2006. J. Korhonen, T. Ojala, M. Klemola, and P. Väänänen, “mTag – Architecture for Discovering Location Specific Mobile Web Services Using RFID and Its Evaluation with Two Case Studies”, Proceedings of the Advanced International Conference on Telecommunications and International Conference on Internet and Web Applications and Services (AICT/ICIW 2006), IEEE Computer Society, Washington DC, USA, February 19-25, 2006, pp. 191. L.M. Ni, Y. Liu, Y.C. Lau, and A.P. Patil, “LANDMARC: indoor location sensing using active RFID”, Proceedings of the First IEEE Conference on Pervasive Computing and Communications (PerCom'03), IEEE Computer Society, March 23-26, 2003, pp. 407-415. G.M. Giaglis, P. Kourouthanassis, and A. Tsamakos, “Towards a classification framework for Mobile Location Services”, In Mobile Commerce: Technology, Theory, and Application, IGI Publishing, Hershey, PA, USA, 2003, pp. 67-85. I.A. Junglas, “An Experimental Investigation of Location–Based Services”, Proceedings of the 38th Annual Hawaii International Conference on System Sciences (HICSS'05), IEEE Computer Society, Washington DC, USA, January 3-6, 2005, Vol.3, pp. 85a-85a. R. Want, A. Hopper, V. Falcao and J. Gibbons, “The active badge location system”, ACM Transactions on Information Systems (TOIS), ACM, New York, USA, Jan.1992, Vol.10, Issue 1, pp. 91-102. J. Hightower, C. Vakili, C. Borriello, and R. Want, “Design and Calibration of the SpotON AD-Hoc Location Sensing System”, University of Washington, Department of Computer Science and Engineering, Seattle, WA, August 2001. L. Ho, M. Moh, Z. Walker, T. Hamada, and C.-F. Su, “A Prototype on RFID and Sensor Networks for Elder Healthcare: Progress Report”, Proceeding of the 2005 ACM SIGCOMM workshop on Experimental approaches to wireless network design and analysis, ACM, Philadelphia, Pennsylvania, USA, 2005, pp. 70-75. A. LaMarca, Y. Chawathe, S. Consolvo, J. Hightower, I. Smith, J. Scott, T. Sohn, J. Howard, J. ughes, F. Potter, J. Tabert, P. Powledge, G. Borriello, and B. Schilit, “Place Lab: Device Positioning Using Radio Beacons in the Wild”, Proceeding of the 3rd International Conference on Pervasive Computing: Pervasive2005”, Springer Berlin / Heidelberg, Munich, Germany, May 8-13, 2005, Vol. 3468, pp. 116-133.

Figure 10. Client Application showing current location of the user

[2] [3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

Figure 11. Client Application showing sight image.

159

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

High-Precision Multi-Wave Rectifier Circuit Operating in Low Voltage + 1.5 Volt Current Mode
Bancha Burapattanasiri Department of Electronic and Telecommunication Engineering, Engineering Collaborative Research Center Faculty of Engineering, Kasem Bundit University Bangkok, Thailand 10250 .
Abstract—This article is present high-precision multi-wave rectifier circuit operating in low voltage + 1.5 Volt current modes by CMOS technology 0.5 µm, receive input and give output in current mode, respond at high frequency period. The structure compound with high-speed current comparator circuit, current mirror circuit, and CMOS inverter circuit. PSpice program used for confirmation the performance of testing. The PSpice program shows operating of circuit is able to working at maximum input current 400 µAp-p, maximum frequency responding 200 MHz, high precision and low power losses, and non-precision zero crossing output signal. Keywords-component; rectifier circuit, high-precision, low voltage, current mode.

II.

BASIC PRINCIPLE AND DESIGNATION

A. Current Mirror Circuit The current mirror circuit is compound input and output current circuit, it is the current source to amplifier circuit, because of low input resistance but high output resistance that is the characteristic of current source.

I.

INTRODUCTION

The rectifier circuit is very significance in analog signal processing for example, AC voltmeter, detector signal circuit, demodulate circuit etc. [1,2] then, always development all the time for example, full wave rectifier circuit operating in voltage mode [3], full wave rectifier circuit operating in diode and bipolar transistor [4], there are used voltage operating around 0.3 volt for Ge, and 0.6 volt for Si, then it has signal error in crossing zero, and in low input signal case the circuit doesn’t work, because of the characteristic of diode and transistor has limited. From the limitation has development to use another active device. In the past of century has a lot of the presentation in rectifier circuit in current mode [5], but the circuit still have complicated, dissipation of current source, it has little precision, responding in low frequency and operating in narrow range current, So, In this article would like to show a new choice of easier rectifier circuit, non complicate in active devices connected, but still have the high performance of working by 0.5 µm CMOS technology, and the compound with high-speed current comparator circuit, current mirror circuit, and CMOS inverter circuit. Thus, it able to responding at high operating precision, working at high frequency, responding at high input current, high precision signal, high speed, low disputation power, and non precision zero-crossing output signal.

Figure 1. Basic current mirror circuit

From figure 1 M1 pin drain connected to M1 pin gate, which is the same of diode. When current source is stable I1 pass to M1 then voltage drop between pin gate and pin source of M1, thus voltage drop at M2 too, when M1 and M2 has same characteristic then I1=Iin as equation (1) and (2)
ID = 1 W 2 k ' n (Vgs − Vt ) 2 L

(1) (2)

Io = Ii

(W / L )2 (W / L )1

B. Completely High-Precision Multi-Wave Rectifier Circuit Operating in Low Voltage + 1.5 Volt Current Mode. From the principle of current mirror circuit, when designing to rectifier circuit it able to show the structure of completely high-precision multi-wave rectifier circuit operating in low voltage + 1.5 Volt current mode as figure 2. The structure compound with high-speed current comparator

Identify applicable sponsor/s here. (sponsors)

160

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

circuit, current mirror circuit, and CMOS inverter circuit. The functional working is when high-speed current comparator circuit received positive and negative input signal, then it is comparisons the current by sender negative signal to CM1, and sender positive signal to CM2, from the CM characteristics as equation (1) and (2), then will be has half-wave positive and negative phase output signal from CM1 and CM2

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

From circuit in figure 3 when sender input current signal as equation (3) after that M3 currenting M4 non currently, output voltage sending back to common source circuit, then M1 currenting and Iin < 0 is equation to M2 pin drain current, and it will mirrored current passing to M6. So, the new signal is half-wave negative phase output signal at M6 pin drain as equation (5).

Figure 3. Completely high-precision multi-wave rectifier circuit operating in low voltage + 1.5 volt current mode.

In the opposite things, when input current signal as equation (4) will make M4 currenting and M3 non currenting, Vss voltage sender to input common source circuit, then M4 currenting Iin > 0 it equation to M2 pin drain current and it passed to M8 pin drain, reflected current to M9, after that out put signal will be positive phase at M9 pin drain as equation (6).
Figure 2. Diagram is show high-precision multi-wave rectifier circuit operating in low voltage + 1.5 volt current mode.

I DM 5 = I DM 6 = − I OH

(5) (6)

While, if send CM1 output signal to CM3 and add it to CM2 output signal, then sending to CM4 input after that it will be full wave negative phase signal. In the similarly things, when CM4 output signal send to CM5 input by the characteristic and the principle of CM as equation (1) and (2), then it will be full-wave negative phase output signal at the same time positive and negative input signal passing to high-speed current comparator circuit, then some part of signal passing to input CMOS inverter circuit for establish square wave signal operation in voltage mode. C. Circuit Description From figure 2 rectifier circuit, if high-speed current comparator circuit, current mirror circuit, and CMOS inverter circuit put together and sitting MOS transistor operation at saturation region after that it will be completely high-precision multi-wave rectifier circuit operating in low voltage + 1.5 volt current mode as figure 3.

I DM 8 = I DM 9 = + I OH

From the characteristic and principle of current mirror circuit, if send M7 pin drain signal (it has equation to M6 pin drain signal) pass to M11 pin drain, then reflected to M12 for combine with full-wave positive phase current signal at M10 pin drain and sent it to M13 pin drain then reflected current past to M14 after that output signal is full-wave negative phase at M14 pin drain as equation (7)

− I OF = I DM 14 = I DM 10 + I DM 12

(7) (8)

+ I OF = I DM 15 = I DM 17

I DM 2 = 0 And I DM 1 = I in When I in > 0 I DM 1 = 0 And I DM 2 = I in When I in < 0

(3) (4)

In the similarly thing, if sending signal at M15 pin drain to M16 pin drain, then reflected current to M17 that output signal is full-wave positive phase at M17 pin drain as equation (8), after that output signal is full-wave positive and negative phase input signal send to M1, M2, M3 and M4 then some part of signal at M3 and M4 pin drain has passed to M18 and M19 for establish square signal at pin drain as equation (9).

Vout(square =Vss then Iin > 0 and Vout(square) = VDD then Iin < 0 (9) )

161

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

III.

SIMULATION AND MEASUREMENT RESULT

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

The PSpice programs used for confirmation the performance of testing , so setting parameter 0.5 µm of MIETEC for PMOS transistor and NMOS VDD=1.5 V VSS=-1.5 V input current operating at range 0-400 µAp-p. Figure 4 shows half-wave output signal is sending input signal 400 µAp-p at frequency 10, 100, 200 MHz. Figure 5 shows full-wave negative phase output signal is sending input signal 400 µAp-p at frequency 10, 100, 200 MHz. Figure 6 shows full-wave positive phase output signal is sending input signal 400 µAp-p at frequency 10, 100, 200 MHz. Figure 7 shows square-wave output signal is sending input signal 400 µAp-p at frequency 10, 100, 200 MHz. Figure 8 shows the characteristic DC current at input signal 400 µAp-p and temperature 25O,50O, 75O and 100O
(a)

(a)

(b)

(b)

(c) Figure 4. Half-wave output signal at input current 400 µAp-p at (a) frequency = 10 MHz, (b) frequency = 100 MHz and (c) frequency = 200 MHz

(c) Figure 5. Full-wave negative phase output signal at input current 400 µAp-p at (a) frequency = 10 MHz, (b) frequency = 100 MHz and (c) frequency = 200 MHz

162

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

(a)

(a)

(b)

(b)

(c) Figure 6. Full-wave positive phase output signal at input current 400 µAp-p at (a) frequency = 10 MHz, (b) frequency = 100 MHz and (c) frequency = 200 MHz

(c) Figure 7. Square-wave output signal at input current 400 µAp-p at (a) frequency = 10 MHz, (b) frequency = 100 MHz and (c) frequency = 200 MHz

163

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

ACKNOWLEDGMENT

The researchers, we are thank you very much to our parents, who has supporting everything to us. Thankfully to all of professor for knowledge and a consultant, thank you to Miss Suphansa Kansa-Ard for her time and supporting to this research. The last one we couldn’t forget that is Kasem Bundit University, Engineering Faculty for supporting and give opportunity to our to development in knowledge and research, so we are special thanks for everything. REFERENCES
Barker, R.W.J., “Versatile precision full wave Rectifier,” Electron Letts, 5: Vol.13, pp. 143-144, 1977. [2] Barker, R.W.J. and Hart, B.L., “Precision absolute-value circuit technique,” Int. J. Electronics Letts, 3: Vol.66, pp. 445-448, 1989. [3] Toumazou, C. and Lidgey, F.J., “Wide-Band precision rectification,” IEE Proc. G, 1: Vol.134, pp.7-15, 1987. [4] Wang, Z., “Full-wave precision rectification that is performed in current domain and very suitable for CMOS implementation,” IEEE Trans. Circuits and Syst, 6: Part I, Vol. 39, pp.456-462, 1992. [5] Toumazou, C., Lidgey, F.J.and Chattong, S., “High frequency current conveyor precision full-wave rectifier,” Electron. Letts, 10: Vol. 30, pp. 745-746, 1994. [6] Wilson, B. and Mannama, V., “Current-mode rectifier with improved precision,” Electron. Letts, 4: Vol. 31, pp. 247-248, 1995. [7] Surakampontorn, W. and Riewruja, V., “Integrable CMOS sinusoidal frequency doubler and full-wave rectifier,” Int.J. Electronics, Letts, 3: Vol. 73, pp. 627-632, 1992. [8] Traff, H., “Novel Approach to High Speed CMOS Current Comparators,” Electron. Letts, 3: Vol. 28, pp. 310-312, 1992. [9] Monpapassorn, A., “Improved Class AB Full-Wave rectifier,” Thammasat Int. J. Sc. Tech., No. 3, November, Vol. 4, 1999. [10] A. Monpapassorn, K. Dejhan, and F.Cheevasuvit, “CMOS dual output current mode half-wave rectifier,” International Journal of Electronics, Vol. 88, 2001, pp. 1073-1084. [1]

Figure 8. Output characteristic DC current at 400 µAp-p p input current, temperature 25O,50O, 75Oand 100O

IV.

CONCLUSION

The circuit has designed by high-speed current comparator circuit, current mirror circuit and CMOS Inverter circuit. Setting transistor operating at saturation region, so the circuit is able to high precision, but low voltage at + 1.5 Volt. The simulation result is able to confirm the performance of working at maximum frequency 200 MHz, maximum output 400 µAp-p, high precision, dissipation loss power and non precision zero-crossing output signal, so it suitable to apply in analog signal processing. APPENDIX The parameters used in simulation are 0.5 µm CMOS Model obtained through MIETEC [10] as listed in Table I. For aspect ratio (W/L) of MOS transistors used are as follows: 1.5 µm/0.15 µm for all NMOS transistors; 1.5 µm/0.15 µm for all PMOS transistors.
TABLE I. CMOS MODEL USED IN THE SIMULATION

AUTHORS PROFILE
Mr.Bancha Burapattanasiri received the bleacher degree in electronic engineering from Kasem Bundit University in 2002 and master degree in Telecommunication Engineering, from King Mongkut’s Institute of Technology Ladkrabang in 2008. He is a lecture of Electronic and Telecommunication Engineering, Faculty of Engineering, Kasem Bundit University, Bangkok, Thailand. His research interests analog circuit design, low voltage, high frequency and high-speed CMOS technology.

---------------------------------------------------------------------------------------------.MODEL CMOSN NMOS LEVEL = 3 TOX = 1.4E-8 NSUB = 1E17 GAMMA = 0.5483559 PHI = 0.7 VTO = 0.7640855 DELTA = 3.0541177 UO = 662.6984452 ETA = 3.162045E-6 THETA = 0.1013999 KP = 1.259355E-4 VMAX = 1.442228E5 KAPPA = 0.3 RSH = 7.513418E-3 NFS = 1E12 TPG = 1 XJ = 3E-7 LD = 1E-13 WD = 2.334779E-7 CGDO = 2.15E-10 CGSO = 2.15E-10 CGBO = 1E-10 CJ = 4.258447E-4 PB = 0.9140376 MJ = 0.435903 CJSW = 3.147465E-10 MJSW = 0.1977689 .MODEL CMOSP PMOS LEVEL = 3 TOX = 1.4E-8 NSUB = 1E17 GAMMA = 0.6243261 PHI = 0.7 VTO = -0.9444911 DELTA = 0.1118368 UO = 250 ETA = 0 THETA = 0.1633973 KP = 3.924644E-5 VMAX = 1E6 KAPPA = 30.1015109 RSH = 33.9672594 NFS = 1E12 TPG = -1 XJ = 2E-7 LD = 5E-13 WD = 4.11531E-7 CGDO = 2.34E-10 CGSO = 2.34E-10 CGBO = 1E-10 CJ = 7.285722E-4 PB = 0.96443 MJ = 0.5 CJSW = 2.955161E-10 MJSW = 0.3184873 ----------------------------------------------------------------------------------------------

164

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Classifying Application Phases in Asymmetric Chip Multiprocessors
A. Z. Jooya
Computer Science dept. Iran University of Science and Technology Tehran, Iran .

M. Analoui
Computer Science dept. Iran University of Science and Technology Tehran, Iran .

Abstract— In present study, in order to improve the performance and reduce the amount of power which is dissipated in heterogeneous multicore processors, the ability of detecting the program execution phases is investigated. The program’s execution intervals have been classified in different phases based on their throughput and the utilization of the cores. The results of implementing the phase detection technique are investigated on a single core processor and also on a multi-core processor. To minimize the profiling overhead, an algorithm for the dynamic adjustment of the profiling intervals is presented. It is based on the behavior of the program and reduces the profiling overhead more than three fold. The results are obtained from executing multiprocessor benchmarks on a given processor. In order to show the program phases clearly, throughput and utilization of execution intervals are presented on a scatter plot. The results are presented for both fixed and variable intervals. Keywords- Heterogeneous multi-core processor; multiprocessor benchmarks; program phase, execution intervals; dynamic profiling; throughput; resource utilization

I.

INTRODUCTION

The amount of diversity among applications that a typical computer is expected to run can be considerable. Also there is significant diversity among different phases of the same application. The utilization of processor resources changes among different phases (a program phase is defined as a contiguous interval of program execution in which the program behavior remains relatively unchanged). Therefore, if a processor had the ability of detecting phase changes, it could adapt its resources based on the new phase requirement. This can cause significant reduction in power consumption [1]. In [2] authors proposed positional adaptation which uses the program structure to identify major program phases. They proposed the use of a hardware based call stack to identify program subroutines. In [3] authors found a relationship between phases and instruction working sets, and they found that phase changes occur when the working set changes. In [4, 5], authors used techniques from machine learning to classify the execution of the program into phases (clusters). In [6] authors improved the execution time and power of multicore processors by predicting the optimal number of

threads depending on the amount of data synchronization and the minimum number of threads required to saturate the offchip bus. In [7] accelerated critical sections (ACS) technique is introduced which leverages the high-performance core(s) of an Asymmetric Chip Multiprocessor (ACMP) to accelerate the execution of critical sections. In [8] authors proposed scheduling algorithms based on the Hungarian Algorithm and artificial intelligence (AI) search techniques which effectively match the capabilities of each core with the requirements of the applications. In [9] authors found that the use of ready and in-flight instruction metrics permits effective co-scheduling of compatible phases among the four contexts. In [10] authors proposed a scheme for assigning applications to appropriate cores based on the information which the job itself presents as an architectural signature of the application, and is composed of certain microarchitectureindependent characteristics. In [11] authors made a case that thread schedulers for heterogeneous multicore systems should balance between three objectives: optimal performance, fair CPU sharing, and balanced core assignment. They argued that thread to core assignment may conflict with the enforcement of fair CPU sharing. In [12] they introduced a cache-fair algorithm which ensures hat the application runs as quickly as it would under fair cache allocation, regardless of how the cache is actually allocated. Comparison between our proposed scheme and previous studies is part of our ongoing research. In heterogeneous multi-core processors, when the phase change is detected, the process will be switched to the core for which the performance is much closer to the new phase requirement. This strategy could best prevent the core resource under utilization or over utilization. In the present study, the phase changes of the program’s processes that are distributed between cores are investigated, focusing on multiprocessor benchmarks’ phase behavior in multicore processors. The rest of the paper has the following organization. In section two, the methodology is described for fix and variable profiling interval length. How the throughput and utilization of intervals are used to detect the phase changes and adjusting the interval lengths dynamically are described. Section three presents simulations results for single and multicore processors

165

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

and section four summarizes the results and contributions of our work. II. METHODOLOGY

If (

U i < δ under _ utilize )

When a process starts running on a core, it goes through different phases during its execution and each phase has specific demand on the core. Therefore, the core selected to run a process will not remain the best for all phases. This leads the scheduler to make assignment decision in a dynamic manner. The goal of such a scheduler is assigning each process to a core which best satisfy its requirements and prevents underutilization or over-utilization of core’s resources. Here presented phase detection technique uses dynamic profiling with variable profiling interval length. The profiling starts with minimum interval length and increases based on program’s behavior. This causes significant reduction in profiling overhead. To detect a phase change, the throughput (the number of instructions retiring in an interval) average of all intervals which belong to the same phase is calculated and the difference percentage of current interval throughput from throughput average of previous intervals of the phase, D i , is considered as phase detection criterion, and defined as follow:
Di = (Th −Th i −1 )×100 i Th i −1

under_utilization_flag = 1; else if ( U i > δ over _ utilize ) over_utilization_flag = 1; if ( Di > δ Th || under_utilization_flag || over_utilization_flag) intervals are not similar; else intervals are similar; Figure 1. The pseudo-code for similarity function.

As previously described, the interval length, τ , is better to be adjusted dynamically in order to reduce the profiling overhead. The related algorithm is illustrated in Fig. 2. In this study, the minimum of 100K cycles is assumed as the default value of τ . A phase is called steady if the difference between throughput average of all previous intervals and throughput average of all intervals doesn't meet the predefined threshold. In this case the interval length is doubled. Otherwise, the interval length is divided by two. It is worth noting that all thresholds, used in the algorithm, have been obtained after numerous simulations.

(1)

Phase_chang_flag=0; if ( Thi − Thi −1 < throughput_average_chang_threshold ) steady_phase_count++; if ( steady_phase_count == steady_phase_upper_bound ) τ *=2; else

in which Thi is the throughput of ith interval and Thi − 1 is The throughput average of previous intervals that belong to the same phase and, for a given i, defined as follow:
Thi + (Thi −1 × N i −1 ) Ni

Thi =

(2)

τ

steady_phase_count = 0; /=2;

Figure 2. The pseudo-code for adjusting interval length dynamically.

in which, Ni is the number of phase intervals. The phase change is detected by comparing D i with a predefined threshold, δTh , for throughput changes. The utilization (of integer and floating point functional units) that is obtained for each interval are profiled separately. The higher utilization is chosen for each interval, because in the most of benchmarks, there are significant difference between the number of integer and floating point instructions, and usually one of them overcomes the other. Comparing the utilization of each interval, Ui , with predefined thresholds, δ over _ utilize and The utilization and throughput of program intervals are shown with scatters. This method of representation is more beneficial to distinguish different phases. The ideal points in the throughput-utilization scatter are those with both high utilization and throughput. Approaching the utilization to 100 % indicates that core resources are over-utilized. This condition implies that the application could use more resources and that switching to a stronger core will most likely boost performance. On the other hand, points with low utilization indicate that the core’s resources are under-utilized. In this condition switching the application to a weaker core will most likely reduce power dissipation without compromising performance. A. Microarchitecture The simulation has been performed on four cores of two types with different level of performance. Table 1 summarizes the characteristics of the cores. The A-type cores have the higher performance. The level one cache is private for each core and the instruction and data caches are separated. A large unified level two cache is shared between all the cores. MOSE

δ under _ utilize ,

for detecting core resource over utilization or

under utilization, respectively. Fig. 1 presents the relevant pseudo-code which summarizes the above discussion of using the throughput and utilization of intervals to make a decision about phase changes.

Identify applicable sponsor/s here. (sponsors)

166

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

protocol is used for cache coherency. This architecture of memory hierarchy is the most commonly used one in multicore processors. We consider that all the cores are implemented in 100 nm technology and run at 2.1 GHz.
TABLE I. Core Issue-width IL1-cache DL1-cache L2-cache B-predictor Int window size FP window size CHARACTRISTICS OF THE CORES. A-type 4 ( Out-Of-Order ) 64 KB , 4 way 64 KB , 4 way 4 Mb , 8 way hybrid 80 32 B-type 2 ( Out-Of-Order ) 32 KB , 2 way 32 KB , 2 way (shared) hybrid 56 16

Comparing between Fig. 3 and 4, one can deduce that for the most of intervals the B-type processor has around 10% higher utilization, especially for intervals with throughput more than 600K instruction, and they could belong to the same phase. Fig. 5 presents the behavior of the FMM benchmark on the first A-type core when it is run on a multicore processor. The results indicate that intervals with high utilization and throughput, could be executed on A-type cores and other intervals could be run on the B-type cores, and this outcome is in consistent with the considerations on figures 3 and 4 for single cores.

B. Simulator and Benchmarks The simulations have been carried out utilizing SESC simulator developed by Jose Renau et al. [13]. This simulator can probably model a very wide set of architectures such as single processors, multicore processors and thread level speculation. Authors made the required modifications to include the phase detection algorithm. Four scientific/technical parallel workloads from splash2 [14] have been used. These workloads consist of three applications and one computational kernel. The kernel is FFT and the three applications that we have used are Barnes, waterspatial and FMM.
Figure 3. a)The throughput-utilization and b) throughput graphs for FMM benchmark with 500K cycle interval length on single A-type core processor.

III.

EXPERIMENTAL RESULTS

In this section, the results of simulation of the benchmarks execution on the multi-core processor are presented. In the first step, the benchmark execution on a single core processor is simulated. The core utilization and the processor throughput have been captured and shown on two graphs. Then, we simulate the execution of the benchmarks on a four-core processor and put into the graph the utilization and the throughput. In the second step, these graphs are reproduced for each core when benchmarks are run on the multi-core processor. The profiling interval kept fixed or variable in both experiments. It is easy to show the results when the interval is fixed whereas a dynamic interval will reduce the overhead. A. Fixed Interval Length For each benchmark, three figures are depicted. The first figure shows the results for the benchmark which is run on a processor with a single A-type core. Next two figures are for the first A-type and the first B-type cores in multicore processor respectively. Figs. 3 to 5 present the simulation results for FMM benchmark. In Fig. 3 that belongs to a single A-type core processor, the throughput-utilization scatter shows a wide distribution of intervals. It is evident from this figure that the intervals could be divided in two or three phases. The throughput scatter indicates repeating in every 100 intervals. Fig. 4 is the similar results for a single B-type processor.

Figure 4. a) The throughput-utilization and b) throughput graphs for FMM benchmark with 500K cycle interval length on single B-type core processor.

The next considered benchmark is FFT (Figs. 6 to 8). As can be found in the presented results, in comparison with the other investigated benchmark, the distribution of this benchmark is dispersed.

167

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Figure 5. a) The throughput-utilization b) throughput graphs for FMM benchmark with 500K cycle interval length on the first B-type core of multicore processor. Figure 7. a) The throughput-utilization and b) throughput graphs for FFT benchmark with 100K cycle interval length on the first A-type core of multicore processor.

The main difference that is clear from the scatter plots is that the FFT benchmark has lower throughput with three distinct phases for both type of processors and cores on multicore processor. The throughput of the first 30 intervals are for integer instructions that consist the first phase of process. The second phase containing intervals 30 to 185 is composed of integer and floating point instructions. The third phase, from interval number 186 to 270, has both type of instructions, but its integer units’ utilization is quite higher than previous phase.

Figure 8. a) The throughput-utilization and b) throughput graphs for FFT benchmark with 100K cycle interval length on the first B-type core of multicore processor.

The results for other benchmarks are removed to meet the page limitation of the paper.
Figure 6. a) The throughput-utilization and b) throughput graphs for FFT benchmark with 100K cycle interval length on single A-type core processor.

B. Variable Interval Length In this section the simulation results for variable interval length are discussed. The throughput scatters of cores in multicore processor are represented. Here a phase change is

168

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

assumed to accrue when the difference of the throughput of current interval exceeds 100% of the average throughput of all previous intervals (are shown by circles). If the utilization of five successor intervals are more than 95% or less than 30% then a phase change is accrued (shown by triangles). If the difference between the throughput average of the previous intervals and the throughput average of all intervals including the throughput of the current interval stays between -1 and 1 for more than 75 intervals, which shows no change in the application behavior, then the interval length is multiplied by two (shown by squares), otherwise if it exceeds from this range the interval length is divided by two, because it means that application is more likely to change its behavior and needs to be monitored in shorter intervals. The obtained results are depicted in Figs 9 and 10. Note that when the interval length is multiplied by 2, the throughput of the next interval increases as well. Comparing the results of variable length and the results obtained by applying fix interval length, one can conclude that utilizing the here proposed variable interval length can reduce the profiling overhead more than three times, on average and for all the benchmarks under consideration.
Figure 10. The throughput scatters of variable interval size for a) the firs Atype core and b) the first B-type core for FMM benchmark. The phase changes that are caused bye throughput changes and utilization changes are shown by circles and triangles, respectively. Interval length changes are shown by squares.

REFERENCES
R. Kumar, K. Farkas, N. Jouppi, P. Ranganathan, D. Tullsen, Processor Power Reduction Via Single-ISA Heterogeneous Multi-Core Architectures, In Computer Architecture Letters, Volume 2, April 2003. [2] M. Huang, J. Reneau, J. Torrellas, Positional adaptation of processors: application to energy reduction, In Proc. 2003 International Symposium on Computer Architecture, Jun. 2003, pp. 157-168. [3] A. S. Dhodapkar, J.E. Smith, Managing multi-configuration hardware via dynamic working set analysis, In Proc. 2002 International Symposium on Computer Architecture, May 2002, pp. 233-244. [4] T. Sherwood, E. Perelman, B. Calder, Basic block distribution analysis to find periodic behavior and simulation, In Proc. 2001 International Conference on Parallel Architectures and Compilation Techniques, Sep. 2001, pp. 3-14. [5] T. Sherwood, E. Perelman, G. Hamerly, B. Calder, Automatically characterizing large scale program behavior, In Proc. 2002 International Conference on Architectural Support for Programming Languages and Operating Systems, Oct. 2002, pp. 45-57. [6] M. Aater Suleman, Moinuddin K. Qureshi, and Yale N. Patt, Feedback Driven Threading: Power-Efficient and High-Performance Execution of Multithreaded Workloads on CMPs, In the International Conference on Architectural Support for Programming Language and Operating Systems (ASPLOS) 2008. [7] M. Aater Suleman, Onur Mutlu, Moinuddin K. Qureshi and Yale N. Patt, Accelerating Critical Section Execution with Asymmetric MultiCore Architectures, In the International Conference on Architectural Support for Programming Language and Operating Systems (ASPLOS) 2009. [8] Jonathan A. Winter, David H. Albonesi: Scheduling algorithms for unpredictably heterogeneous CMP architectures. DSN 2008: 42-51. [9] Ali El-Moursy, R. Garg, David H. Albonesi, Sandhya Dwarkadas: Compatible phase co-scheduling on a CMP of multi-threaded processors. IPDPS 2006 [10] Daniel Shelepov and Alexandra Fedorova, Scheduling on Heterogeneous Multicore Processors Using Architectural Signatures, In Proceedings of the Workshop on the Interaction between Operating Systems and Computer Architecture, in conjunction with ISCA-35, Beijing, China, 2008. [11] Alexandra Fedorova, David Vengerov and Daniel Doucette, Operating System Scheduling on Heterogeneous Core Systems, In Proceedings of [1]

Figure 9. The throughput scatters of variable interval size for a) the firs Atype core and b) the first B-type core for FMM benchmark. The phase changes that are caused bye throughput changes and utilization changes are shown by circles and triangles, respectively. Interval length changes are shown by squares.

IV.

CONCLUSIONS

Heterogeneous multicore processors have cores with different level of performances. We can achieve higher throughput and lower power consumption by assigning the incoming process to appropriate core. Moreover, a process may have different resource demands in its different parts. The detection of demand changes can be used in an assignment schedule for the purpose of higher throughput and lower power consumption. In this work we present a theory for detecting demand changes in a process. The detection has been done in dynamic time intervals. The dynamic interval reduces the profiling overheads in the magnitude of three folds.

169

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
the First Workshop on Operating System Support for Heterogeneous Multicore Architectures, at PACT 2007, Brasov, Romania. [12] Alexandra Fedorova, Margo Seltzer and Michael D. Smith, Improving Performance Isolation on Chip Multiprocessors via an Operating System Scheduler, In Proceedings of the Sixteenth International Conference on Parallel Architectures and Compilation Techniques (PACT), Brasov, Romania, September 2007. [13] J. Renau, B. Fraguela, J. Tuck, W. Liu, M. Prvulovic, L. Ceze, S. Sarangi, P. Sack, K. Strauss, P. Montesinos, SESC simulator, January 2005, http://sesc.sourceforge.net. [14] S.C. Woo, M. Ohara, E. Torrie, J. P. Singh, A. Gupta, The SPLASH-2 Programs: Characterization and Methodological Considerations, In Proc. 1995 International Symposium on Computer Architecture, Santa Margherita Ligure, Italy, by the ACM, June 1995, pp. 24-36.

AUTHORS PROFILE Morteza Analoui received his B.S. in Electrical Engineering from Iran University of Science and Technology in 1982. He received his Ph.D. in Electrical and Electronic Engineering from Okayama University in 1990. He is an assistant professor in Iran University of Science and Technology. His research interests include computer networks and Internet, stochastic pattern recognition, grid computing and computer network performance analysis.

Ali Z. Jooya recived his B.S. in Computer Engineering from Azad University in Tehran in 2003. He received his M.Sc. in Computer Engineering from Iran University of Science and Technology in 2009. He worked on dynamic scheduling in heteregenous multicore processors under the supervision of Prof. M. Analoui. His research interest include high performance, low power processor design and moder processor performance evaluation.

170

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Syllable analysis to build a dictation system in Telugu language
N.Kalyani
Assoc. Prof, CSE Dept G.N.I.T.S Hyderabad, India. . .

Dr K.V.N.Sunitha
.Professor & HOD, CSE Dept. G.N.I.T.S Hyderabad, India.

Abstract— In recent decades, Speech interactive systems gained increasing importance. To develop Dictation System like Dragon for Indian languages it is most important to adapt the system to a speaker with minimum training. In this paper we focus on the importance of creating speech database at syllable units and identifying minimum text to be considered while training any speech recognition system. There are systems developed for continuous speech recognition in English and in few Indian languages like Hindi and Tamil. This paper gives the statistical details of syllables in Telugu and its use in minimizing the search space during recognition of speech. The minimum words that cover maximum syllables are identified. This words list can be used for preparing a small text which can be used for collecting speech sample while training the dictation system. The results are plotted for frequency of syllables and the number of syllables in each word. This approach is applied on the CIIL Mysore text corpus which is of 3 million words. Keywords-component; formatting; style; styling; insert (key words)

new language. In country like India which includes 22 officials and a number of unofficial languages, building huge text and speech databases is a difficult task. There are other related works which gained good importance is listed below. Some methods that require manually annotated speech corpora for speech recognition are listed. A method called bootstrapping is proposed by Rabiner et al (1982)[3] which can increase the transcribed data for training the system, for speech recognition. Ljolje & Riley (1991)[4] have used an automatic approach to segmentation and labeling of speech when only the orthographic transcription of speech is available. Kemp&Waibel (1998)[5] used unsupervised training approach for speech recognition for TV broadcasts. Wessel&Ney (2001)[6] have proposed an approach in which a low-cost recognizer trained with one hour of manually transcribed speech is used to recognize 72 hours of unrestricted acoustic data. Lamel et al (2002)[7] have shown that the acoustic models can be initialized using as little as 10 minutes of manually annotated data. There are also few methods that do not require any manually annotated speech corpora for speech recognition. Incremental maximum a posteriori estimation of HMMs is proposed by Gotoh & Hochberg (1995)[8]. This algorithm randomly selects a sub-set of data from the training set, updates the model using maximum a posteriori estimation and this process is iterated until it covers all possible units. Chang et al (2000)[9] proposed a method which extracts articulatoryacoustic phonetic features from each frame of speech signal and then the phone is identified using neural networks. There is an interesting approach proposed by Nagarajan & Murthy (2004)[10] for Indian languages. Their approach focuses on automatically segmenting and transcribing the continuous speech signal into syllable-like units using a group delay based segmentation algorithm without the use of manually segmented and labeled speech corpora. This approach is more appropriate for Indian languages as they are syllable centered. The focus of this paper is to extract the possible syllables from the raw Telugu text corpus. Once the syllable like units is

I.

INTRODUCTION (HEADING 1)

Developing a robust dictation system to transcribe continuous speech signal into a sequence of words is a difficult task, as continuous speech does not have any natural pauses in between words. It is also difficult to make the system robust for speaker variability and the environment conditions. There are many research organizations working on speech with different approach. The conventional method of building a large vocabulary speech recognizer for any language uses a top-down approach to speech recognition (Huang & Acero 1993)[1]. What they mean by top-down is that these recognizers first hypothesize the sentence, then the words that make up the sentence and ultimately the sub-word units that make up the words. This approach requires large speech corpus with sentence or phoneme level transcription of the speech utterances (Thomas Hain et al 2005; Ravishankar 1996) [2]. The transcriptions must also include different speech order so that the recognizer can build models for all the sounds present. It also requires maintaining a dictionary with the phoneme/subword unit transcription of the words and language models to perform large vocabulary continuous speech recognition. The recognizer outputs words that exist in the dictionary. If the system is to be developed for a new language it requires building of a dictionary and extensive language models for the

171

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

obtained they are analyzed to understand their frequency of coverage in words. The final list of words can be prepared considering the words having high probable syllables or identifying the words which contain the maximum syllables. Details of Telugu corpus and Procedure to convert Telugu text to WX notation is introduced in section 2. Algorithm which uses Syllabification rules to syllabify the text is discussed in section 3. A study of results obtained is summarized in section 4 and conclusion and future scope presented in section 5.

There are 41 consonants in the common core inventory. In Unicode Standard 3.0 they begin with 0C15 to 0C39 and 0C1A to 0C2F. The character set for consonants in Telugu is complex and peculiar in their function. These character signs have three or more than three distinct shapes depending on their occurrence • • • Base consonants or Primaries, when they are used as stand alone characters. Pure consonant or hanger, when used with a vowel other than the inherent vowel /a/ Ottulu or Secondary consonant, when used as a constituent of a conjunct

II.

CONVERSION OF TELUGU TEXT TO WX NOTATION.

Telugu is one of the major Scheduled languages of India. It has the second largest number of speakers mainly concentrated in South India. It is the official language of Andhra Pradesh and second widely spoken language in Tamilnadu, Karnataka. There are number of Telugu language speakers have migrated to Mauritius, South Africa, and recently to USA, UK, and Australia. Telugu is often referred as "Italian of the East". The Primary units of Telugu alphabet are syllables, therefore it should be rightly called a syllabic language. There is good correspondence in the written and spoken form of the south Indian languages. Any analysis done on written form would closely relate to spoken form of the language. The Telugu alphabet can be viewed as consisting of more commonly used inventory, a common core, and an overall pattern comprising all those symbols that are used in all domains. The overall pattern consists of 60 symbols, of which 16 are vowels, 3 vowel modifiers, and 41 consonants. Since Indian languages are syllable-timed languages, syllable is considered as the basic unit in this work and analysis is performed to identify the words with syllables with high frequency and words with varying coverage of syllables.

The basic character set for consonants are called as primaries or stand alone characters as they occur in the alphabet. Each of which has an inherent vowel /a/ which often /. This graphic sign is explicitly indicated by sign / indicating the vowel /a/ is normally deleted and replaced with another explicit mark for a different vowel. List of pure consonants carrying explicit secondary vowel /a/ sign and its corresponding WX notation are క -ka గ-ga ఖ -Ka
ఘ -G ఙ -fa

,

Ba
ఱ

Da ణ -Na , త -wa థ -Wa ద -xa ధ -Xa న -na , -rY
మ -ma య

చ -ca ఛ -Ca జ -ja ఝ

-Ja ఞ -Fa ,

ట

-ya ర -ra ల -la

వ -va శ -Sa ష -Ra స -sa హ -ha ళ -lYa

ప -pa ఫ -Pa బ -ba భ -

-ta ఠ -Ta డ -da ఢ -

The Telugu text in Unicode format is converted to WX notation. The conversion is done character by character using Unicode value of the character. If the Unicode of the character is between 0C15 and 0C39 (క to హ ), representation corresponding to Pure consonant is retrieved from WX table. If the Unicode of the character is between 0C3E and 0C4C (ా to

ౌ), the last letter from Pure consonant is removed and secondary vowel representation is added. If Unicode of the character is 0C4D which correspond to stress mark ్, the last letter from the WX notation is removed indicating that the next occurrence of character is secondary consonant. B. Algorithm The algorithm for conversion is given below where englishtext is initialized to null. • • string englishtext=null read the contents and convert into character array o for each character till end of the file do if Unicode of the letter is between 0C15 and 0C60 retrieve the corresponding English character for the Unicode, add to englishtext and increment i by 1

A. Telugu to English letter translation The WX notation of thirteen vowel signs is a,A,i,I,u, అ ,ఆ ,ఇ ,ఈ ,ఉ ,ఊ ,ఋ ,ఎ ,ఏ ,ఐ ,ఒ ,ఓ ,ఔ U,q,eV,e,E,oV,o,O occur as stand alone characters and In UNICODE Standard 3.0., each of these is assigned a hexadecimal code point 0C00-0C7F. When a vowel occurs immediately after a consonant it is represented as a dependent or secondary sign called, guNiMtaM gurtulu. The Telugu alphabet is a syllabic language in which the primary consonant always has an inherent vowel [a] / /. When a consonant is attached with another vowel other than [a] / / then secondary vowel sign is attached to the consonant after removing the inherent vowel /a/. There are exceptions where the primary vowel may be considered as secondary.

172

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

else if Unicode of the letter is between 0C3E and 0C4C remove the last letter from the englishtext, retrieve the corresponding English character for the Unicode, add to englishtext and increment i by 1 else if Unicode of the letter is 0C4D remove last letter from the englishtext else copy the character into English text and increment i by 1 o • • end for

A. Algorithm • Read from the file which has text in WX notation. • Label the characters as consonants and vowels using the following rules o o o o o Any consonant except(y, H, M) followed by y is a single consonant, label it as C Any consonant except (y, r, l, lY, lYY) followed by r is taken as single consonant Consonants like(k, c, t, w, p, g, j, d, x, b, m, R, S, s) followed by l is taken as single consonant. Consonant like (k, c, t, w, p, g, j, d, x, b, R, S, s, r) followed by v is taken as a single consonant. Label the remaining as Vowel (V) or Consonant(C) depending on the set to which it belongs. Store the attribute of the word in terms of (C*VC*)* in temp2 file.

Store in temp file for Syllabification. end III. SYLLABIFICATION •

o

The scripts of Indian languages have originated from the ancient Brahmi script. The basic speech sounds units and basic written form has one to one correspondence. An important feature of Indian language scripts is their phonetic nature. The characters are the orthographic representation of speech sounds. A character in Indian language scripts is close to syllable and can be typically of the following form: C, V, CV, CCV and CVC, where C is a consonant and V is a vowel. There are about 35 consonants and about 15 vowels in Indian languages. The rules required to map the letters to sounds of Indian languages are almost straight forward. All Indian language scripts have common phonetic base. The majority of the speech recognition systems in existence today use an observation space based on a temporal sequence of frames containing short-term spectral information. While these systems have been successful [10, 12], they rely on the incorrect assumption of statistical conditional independence between frames. These systems ignore the segment-level correlations that exist in the speech signal. The high-energy regions in the Short Term Energy function correspond to the syllable nuclei while the valleys at both ends of the nuclei determine the syllable boundaries. The text segmentation is based on the linguistic rules derived from the language. Any syllable based language can be syllabified using these generic rules. To make the text segments exactly equivalent to the speech units. The syllable can be defined as a vowel nucleus supported by consonants on either side, It can be generalized as a C*VC* unit where C is a consonant and V is a vowel. The linguistic rules to extract the syllables segments from a text are generated from spoken Telugu. These rules can be generalized to any syllable centric language. The text is preprocessed to remove any punctuation. The following algorithm divides the word into syllable like units.

For each word in the corpus get its label attribute from temp2 file. o o o o o o o If the first character is a C then the associate it to the nearest Vowel on the right. If the last character is a C then associate it to the nearest Vowel on the left. If sequences correspond to VV then break is as V-V. Else If sequence correspond to VCV then break it as V-CV. Else If sequence correspond to VCCV then break it as VC-CV. Else If sequence correspond to VCCCV then break it as VC-CCV. The strings separated by – are identified as syllable units.

• •

repeat Store the result in output file.

The following Table:1 shows the output obtained for the input in Telugu text in UNICODE
TABLE I. S. No 1. 2. 3. Input
కం ె కంటె ఖరుచ్కంటె లా ాలకు

OUTPUT FOR ALGORITHM 1 AND ALGORITHM 2 Output of Algoritm 1
kaMpeVnIkaMteV KarcukaMteV lABAlaku

Output of Algorithm 2
kaM-peV-nI-kaM-teV Kar-cu-kaM-teV lA-BA-la-ku

173

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

IV.

STATISTICAL ALALYSIS

A. Phoneme Analysis: The following observations are based on a study made on CIIL Mysore Telugu text corpus of 3 million words of running texts in Telugu. This corpus is first cleaned and words are extracted. The Word frequencies are dropped in order to avoid their skewing effect on the results of character frequencies. These words are broken into syllables using the rules of the language and analysis is performed to study the distribution of phonemes and syllables. B. Phoneme Frequency chart: On observing it is found that of Vowels cover nearly 44.98% in total text corpus. It clearly shows that the vowels are the major units in the speech utterance. The vowel modifiers coverage is 3.82% and the consonant coverage is 51.21%. The following Fig 1 gives the details of the analysis.
Phoneme variation chart
percentage 60 40 20 0 consonants vowels vowel modifier 3.82% 51.21% 44.98%

D. Consonant variation chart: Consonants are characterized by significant constriction or obstruction in the pharyngeal and/or oral cavities. Some consonants are voiced and others are unvoiced. Many consonants occur in pairs, that is, they share the same configuration of articulators, and one member of the pair additionally has voicing which the other lacks. Based on the articulators involved and manner of articulation the consonants are classified as Bilabial, Dental Alveolar, Retroflex, Velar and Glottal. The distribution of consonants is shown in the following Fig. 3.
Distribution of Consonants
50.00 45.00 40.00 Percentage 35.00 30.00 25.00 20.00 15.00 10.00 5.00 0.00 Bilabial Dental Alveolar Retroflex Velar Glottal 12.12% 9.10% 12.00% Bilabial Dental Alveolar 19.88% Retroflex Velar Glottal 0.82% 45.62%

consonants vowels vowel modifier

Fig. 3 Consonant distribution chart

Based on the above analysis it is clear that the vowels play a major role in the utterance of speech units. Or the basic unit of utterance is concentric at the vowel which forms the key component in the syllable. It is hence good to build the speech recognition systems considering the syllable as the basic unit. E. Syllable Analysis : It is shown that most of the Indian languages are syllable centric. Syllable boundary in speech signal can be approximately identified it is intended to make a study of developing a speech recognition system at Syllable level. The total distinct syllables observed are 12,378 and the frequency of occurrence of the syllables is plotted in the following chart. The number of Syllables with frequency less than 100 is 11057(12,378 – 1321). It is observed that nearly 4903 syllables have frequency one. This is due to loanwords from English like (Apple, coffee, strength etc.) When these words are written in Telugu it normally takes the same pronunciation. Such kind of words creates a different syllable which may not occur in the native language. Fig 4 shows the count of Syllables with the frequencies in Hundreds.
No. Of Syllables N u m b e r o f S y ll a b l e s 1500 1000 500 0 0.1K 0.2K 0.3K 0.4K 0.5K 0.6K 0.7K 0.8K 0.9K 1K frequency of Syllables 1321 921 712 600 522 469 422 395 355 337

Fig. 1 Phoneme variation chart

C. Vowel variation chart: Vowels occur either in stand alone form or as modifiers and have total coverage of 48%. The vowels are classified based on the position of the articulator and manner of articulation. The vowel classification is given in Table 5 and the distribution of vowels is shown in the Fig 2.
TABLE 5 VOWEL CLASSIFICATION Classification Closed Front (CF) Half Closed Front (HCF) Closed Back (CB) Half Closed Back (HCB) Open Vowels I,i eV, e u,U oV, o a, A

Distribution of Vowels
60 Percentage
41.16% 50.70%

CF HCF CB

40 20 0 CF HCF CB HCB Open
17.85% 18.45% 5.66%

HCB Open

Fig. 2 Vowel distribution chart

Fig. 4 Number of syllables with frequency in the range 100 to 1K.

174

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

The following Fig 5 shows the count of Syllables that have frequency ranging from 1K to 10K.
Number of Syllables

Words count, with Syllable Index 50%, 80% and 100% and cut-off frequency varying in the range from 1K to 10K is shown in Fig.8.
Number of words with different coverage of Syllables

N u m b e r o f S y lla b le s

400 300 200 100 0

337
700,000

216 161 135
N u m b e r o f W o rd s

600,000 500,000 400,000 300,000 200,000 100,000 0 1K 2K 3K 4K 5K 6K 7K 8K 9K 10K Frequency of Syllable 0.5 0.8 1

114

100

88

80

77

71

1K

2K

3K

4K

5K

6K

7K

8K

9K

10K

Frequency of Syllables

Fig. 5 Number of syllables with frequency in the range 1K to 10K

The following Fig 6 shows the count of Syllables that have frequency ranging from 10K to 100K.
Number of Syllables 80 N u m b e r o f S y llb le s 60 40 20 0 10K 20K 30K 40K 50K 60K 70K 80K 90K 100K Frequency of Syllables

Fig. 8 Number of words having 50%, 80% and 100% syllables with syllable frequency in the range 1K to 10K.

71

Words count, with Syllable Index 50%, 80% and 100% and cut-off frequency varying in the range from 10K to 100K is shown in Fig.9. It is observed from the above figures that as the frequency increases the number of words included decreases. Importance of the word depends on the Syllable Index and on the cut-of frequency. It is directly proportional to Syllable Index and cutof frequency.

38 21 11 5 3 3 2 1 0

Number of words with different coverage of Syllables
1000000

Fig. 6 Number of syllables with frequency in the range 10K to 100K
N u m b e r o f w o rd s

100000 10000 1000 100 10 1 10K 20K 30K 40K 50K 60K 70K 80K 90K 100K Frequency of Syllables 0.5 0.8 1

It is observed that there are nearly 71 syllables that have frequency more than 10K. A study is also made in terms of the words which have varying number of syllables with varying frequencies. Here in the following figures, plots are given for words which have syllables with cut-off frequency specified on X axis, Y-axis indicates number of words having the Syllable Index and above cut-off frequency and Syllable Index 0.5, 0.8 and 1.0.. Words count, with Syllable Index 50%, 80% and 100% and cut-off frequency varying in the range from 100 to 1000 is shown in Fig.7.
Number of words with different coverage of Syllables
700,000 N u m b e r o f w o rd s 600,000 500,000 400,000 300,000 200,000 100,000 0 0.1K 0.2K 0.3K 0.4K 0.5K 0.6K 0.7K 0.8K 0.9K 1K frequency of Syllables 0.5 0.8 1

Fig. 9 Number of words having 50%, 80% and 100% syllables with syllable frequency in the range 10K to 100K.

V.

CONCLUSION

This paper explores the details of phonemes and syllables in the text corpus. As there is one to one correspondence between the written form and spoken form of the language, detailed analysis is performed to understand the coverage of different phonemes and syllables. This analysis is useful in selecting good set of words that would cover all possible syllables in large vocabulary. Optimal selection of words depends on the selection strategy applied. Good strategy can be used to obtain limited words, which are useful during recording the speech to train the system. This ultimately improves the performance of the dictation system.

Fig. 7 Number of words having 50%, 80% and 100% syllables with syllable frequency in the range 100 to 1000.

175

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

REFERENCE:
Huang, Acero 1993 Spoken Language Processing. (New Jersy: Prentice Hall) [2] Thomas Hain, et al 2005 Automatic transcription of conversational telephone speech. In IEEE Trans.Speech, Audio Processing 13: 1173– 1185 [3] Rabiner L R Rosenberg A E, Wilpon J G, Zampini T M 1982 A bootstrapping training technique for obtaining demisyllabic reference patterns. In J. Acoust. Soc. Amer. 71: 1588–1595 [4] Ljolje A, Riley M D 1991 Automatic segmentation and labelling of speech. In Proceedings of IEEE Int. Conf. Acoust., Speech, and Signal Processing 1: 473–476 [5] Kemp T, Waibel A 1998 Unsupervised training of a speech recognizer using tv broadcasts. In Proc.of ICSLP 98. Vol. 5 Sydney, Australia, 2207–2210 [6] Wessel F, Ney H 2001 Unsupervised training of acoustic models for large vocabulary continuous speech recognition. In IEEE Workshop on ASRU. 307–310 [7] Lamel L, Gauvain J-L, Adda G 2002 Unsupervised acoustic model training. In Proceedings of IEEE Int. Conf. Acoust., Speech, and Signal Processing. 877–880 [8] Gotoh Y and Hochberg M M 1995 Incremental map estimation of hmms for efficient training and improved performance. In Proceedings of IEEE Int. Conf. Acoust., Speech, and Signal Processing 877–880 [9] Chang S, Sastri L, Greenberg S 2000 Automatic phonetic transcription of sponta- neous speech. In Proceedings of Int. Conf. Spoken Language Processing 4: 330–333 [10] Nagarajan T, Murthy H A 2004 Non-bootstrap approach to segmentation and labelling of continuous speech. In National Conference on Communication. 508–512 AUTHORS PROFILE Dr K.V.N.Sunitha did her B.Tech ECE from Nagarjuna University in 1988, M.Tech Computer Science from REC Warangal in 1993 and Ph.D from JNTU hyderabada in 2006. She has 18 years of Teaching Experience. She has been working in G.Narayanamma Institute of Technology and Science, Hyderabad as HOD CSE Dept. from the inception of the CSE Dept since 2001. She is a recipient of “Academic Excellence Award” by GNITS in 2004.She is awarded “Best computer engineering Teacher Award “ by ISTE in Second Annual Convention held in Feb 2008.She has published more than 35 papers in International & National Journals and Conferences. Authored a book on ”Programming in UNIX and Compiler Design”. She is guiding five PhDs . She is a reviewer for International Journals-JATIT,IJFCA. She is fellow of Institute of engineers &Sr member for IACSIT, International association of CSIT, life member of many technical associations like CSI, IEEE, ACM. She is academic advisory body member for ACE Engg college , Ghatkesar. She is Board of Studies member for UG ang PG programmaes, CSE at G.Pulla Reddy Engg College,Kurnool. N. Kalyani completed B.Tech civil from Osmania University in 1994, M.Tech Computer Science from JNTUH in 2001. She has working experience of 5 years as Design Engineer in R. K. Engineers, Hyderabad and 9 years of teaching for both under and post graduate students. She is presently working as Associate Professor in the Department of Computer Science Engineering, G.Narayanamma Institute of Technology and Science, Hyderabad. She is reciepient of “Academic Excellence Award” by GNITS in 2008. She has published 10 papers in International & National Journals and conferences. She is a life member of CSI & ISTE technical associations. [1]

176

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Sinusoidal Frequency Doublers Circuit With Low Voltage + 1.5 Volt CMOS Inverter
Bancha Burapattanasiri Department of Electronic and Telecommunication Engineering, Engineering Collaborative Research Center Faculty of Engineering, Kasem Bundit University Bangkok, Thailand 10250 .
Abstract—This paper is present sinusoidal frequency doublers circuit with low voltage + 1.5 volt CMOS inverter. Main structure of circuit has three parts that is CMOS inverter circuit, differential amplifier circuit, and square root circuit. This circuit has designed to receive input voltage and give output voltage use few MOS transistor, easy to understand, non complex of circuit, high precision, low error and low power. The Simulation of circuit has MOS transistor functional in active and saturation period. PSpice programmed has used to confirmation of testing and simulation. Keywords-component; sinusoidal frequency doublers, low voltage,CMOS inverter.

low errors and high precision. The second part is differential amplifier circuit has function to squares rule of circuit. The third part is square-rooter circuit has functional to squared of integrated differential amplifier circuit by the relation of MOS transistor in active and saturation period, when the functional three part of circuit put together is able to write diagram box as figure 1. For more understand in this researcher we are separate the part of circuit operation as following.

I.

INTRODUCTION

Sinusoidal frequency doublers is popular in telecommunication for example using instrument processing, or circuit analysis in analog processing [1-2]. The normally, sinusoidal frequency doublers has be the characteristic of tune LC circuit or analog multiplier circuit. A lot of researches presented has the narrow frequency operation period and non suitable for establish of integrated circuit, so the circuit development by multiplier circuit and sinusoidal frequency doublers used op-amp to function, then it able to charge the limited of tune LC circuit, and able to establish integrated circuit too. However, the circuits still have op-amp then the circuits still have big size, high loss of power supply, and used a lot of device. So in this paper is present new choice of sinusoidal frequency doublers and suitable for establish high integrated circuit. Because of the circuit has designed easy to understand, noncomplex, and MOS transistor functional in active and saturation regions for compound to CMOS inverter circuit, Differential amplifier circuit and square-rooter circuit. The new circuits is of MOS transistor, but still have high efficiency, low power supply + 1.5 Volt, high precision low error and low power. From the purpose of research doesn’t want error with circuit, so setting all K of MOS transistors is equal. II. DESIGNATION AND FUNCTIONAL The sinusoidal frequency doublers circuit with low voltage + 1.5 volt has three parts of circuit. In the first part is CMOS inverter circuit has functional to invert input signal high speed,

Figure 1. Diagram is show sinusoidal frequency doublers circuit with low voltage + 1.5 volt CMOS inverter.

A. CMOS Iverter Crcuit From the figure 2 show CMOS inverter circuit component by M1, M2, M3 and M4. So, Sending input signal has to positive and negative phase entrance pin-gate M1 and M2 it will two MOS transistor serration working. To be a result of CMOS inverter designed working in MOS transistor saturation. So we can computable output–input relation of CMOS inverter by

Figure 2. CMOS inverter circuit

177

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

I DM 1 = I DM 3 And I DM 2 = I DM 4

(1) By And

I RL = I DM 5 + I DM 6

(11) (12) (13)

An equation current of MOS transistor saturation working is

I DM 5 = K 5 (Vin − Vss − VT ) 2

I D = K (VGS − VT ) 2 ; (VGS − VT ) < V DS
When

(2)

I DM 6 = K 6 (−Vin − VSS − VT ) 2

K=

ε ins ε O µ n
D

and β = K (W / L)

(3)

Instead of an equation (12) and (13) in to (11) then the result is

Whence (1) equation and CMOS inverter circuit value to result is

IRL = K5(Vin −VSS −VT )2 + K6 (−Vin −VSS −VT )2
From the designing is setting K5=K6=K (4) So

[

][

]

(14)

I DM 1 = − I DM 2
And

I RL = 2 K (Vin ) 2 + (VSS + VT ) 2

[

]

(15)

I DM 1 =

− βp 2

(V
2

in

− V DD − Vtp )

2

(5)

I DM 2 =

βn

(Vin − Vtn )2

(6)

C. Square-Rooter Circuit From the figure 4 is show a square-rooter circuit component M7 and M8 . The circuit designing is setting M7 work in saturation period and M8 work in non-saturation period. The circuit relation show as an equation current [6]

Then

 VDD + Vtp + Vtn (βn / β p )1/ 2   Vin = − 1/ 2   1 + (βn / β p )  

(7)

If a symmetry point of circuit in equation (8)

Vin = Vout = V DD / 2
By So

(8) (9) (10)
Figure 4. Square-rooter circuit

β n = β p and Vtn = −Vtp
Vout = −Vin

I DM 7 = K (VDM 7 − VO − VT ) 2
2  V  I DM 8 = K (VDM 8 − VT )VO − O  2  

(16)

B. Differential Amplifier Circuit From the figure 3 is show a differential amplifier circuit component M5 and M6. Functional of circuit is working when M5 and M6 receive signal from CMOS inverter circuit, and to setting MOS transistor working by square formulation. The relation M5 and M6 current able to show [5]

(17)

From figure 4 if I DM 7 = I DM 8 so, the relation Voltage output Circuit computable by

RL

VDM 8 =

I DM 7 + VO + VT K

(18)

V in

M5 M6

-V in

When instead result of an equation (18) to (17) the new result is

 IDM 7  V 2 IDM 7 = IDM 8 = K  VO +VT −VT VO − O   2  K   
Figure 3. Differential amplifier circuit

(19)

178

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

IDM 7 2 V2 K O +K V0 − IDM 7 = 0 2 K

If the power series of the from

(20)

1 1 1 + X = 1 +   x +   x 2 + .....are employed (25)  2 8
Then the equation (23) can be rewritten as

From equation (20) is able mathematic calculation to finding VO then the result is square root of drain current as (21)

0.732 VO = IDM 7 K

(21)

VO = 1.035(VDC +VmCos2ωt)
III. SIMULATION AND MEASUREMENT RESULT

(26)

D. The Completely Sinusoidal Frequency Doublers Circuit With Low Voltage + 1.5 volt CMOS Inverter.

From the completely sinusoidal frequency doublers circuit with low voltage + 1.5 volt CMOS inverter is abler to confirmation the efficient of simulation circuit by PSpice programmed to analyze and the testing by send input voltage signal as equation (27) setting VDD = 1.5 Volt, VSS = -1.5 Volt and W/L = 1.5/0.15 µm, by sending input voltage signal at equation (27), and to setting W/L = 1.5/0.15 µm, then the result is output signal as figure 6

Vin = 0.1 Sin 2000 πt

(27)

Figure 5. The completely sinusoidal frequency doublers circuit with low voltage + 1.5 volt CMOS inverter.

From the function of three circuits, when we put gathers, so the new circuit is the completely sinusoidal frequency doublers circuit with low voltage + 1.5 volt, it has high precision, low error, used few MOS transistor, low power. When bring equation (10) of CMOS inverter circuit to analyze by setting Vin = Vm sin ωt the output CMOS inverter result is –Vin. While, the differential amplifier circuit integrated to inverter CMOS circuit, and square-rooter circuit integrated to differential amplifier circuit. So, the new result has related to all result of circuit, it is able to write the output circuit equation as
Figure 6. The result from sending input signal as (27)

VO = 0.732

1 (2KV2 ) in K
2 in

(22) (23)

IV.

CONCLUSION

VO =1.035 V

When bring Vin = Vm sin ωt instead in equation (23) and analyze by trigonometric function relate is

Sin 2θ =1−Cos2θ

(24)

From the sinusoidal frequency doublers circuit with low voltage + 1.5 volt CMOS inverter has presented that show noncomplex of working function, dissipation of current source, and few MOS transistor, operating at input and output in voltage mode, high precision, low error. The efficient simulation circuit is able confirm by PSpice program as presentation principle.

179

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

APPENDIX

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

AUTHORS PROFILE

The parameters used in simulation are 0.5 µm CMOS Model obtained through MIETEC [10] as listed in Table I. For aspect ratio (W/L) of MOS transistors used are as follows: 1.5 µm / 0.15 µm for all NMOS transistors; 1.5 µm / 0.15 µm for all PMOS transistors.
TABLE I. CMOS MODEL USED IN THE SIMULATION

---------------------------------------------------------------------------------------------.MODEL CMOSN NMOS LEVEL = 3 TOX = 1.4E-8 NSUB = 1E17 GAMMA = 0.5483559 PHI = 0.7 VTO = 0.7640855 DELTA = 3.0541177 UO = 662.6984452 ETA = 3.162045E-6 THETA = 0.1013999 KP = 1.259355E-4 VMAX = 1.442228E5 KAPPA = 0.3 RSH = 7.513418E-3 NFS = 1E12 TPG = 1 XJ = 3E-7 LD = 1E-13 WD = 2.334779E-7 CGDO = 2.15E-10 CGSO = 2.15E-10 CGBO = 1E-10 CJ = 4.258447E-4 PB = 0.9140376 MJ = 0.435903 CJSW = 3.147465E-10 MJSW = 0.1977689 .MODEL CMOSP PMOS LEVEL = 3 TOX = 1.4E-8 NSUB = 1E17 GAMMA = 0.6243261 PHI = 0.7 VTO = -0.9444911 DELTA = 0.1118368 UO = 250 ETA = 0 THETA = 0.1633973 KP = 3.924644E-5 VMAX = 1E6 KAPPA = 30.1015109 RSH = 33.9672594 NFS = 1E12 TPG = -1 XJ = 2E-7 LD = 5E-13 WD = 4.11531E-7 CGDO = 2.34E-10 CGSO = 2.34E-10 CGBO = 1E-10 CJ = 7.285722E-4 PB = 0.96443 MJ = 0.5 CJSW = 2.955161E-10 MJSW = 0.3184873 ----------------------------------------------------------------------------------------------

Mr.Bancha Burapattanasiri received the bleacher degree in electronic engineering from Kasem Bundit University in 2002 and master degree in Telecommunication Engineering, from King Mongkut’s Institute of Technology Ladkrabang in 2008. He is a lecture of Electronic and Telecommunication Engineering, Faculty of Engineering, Kasem Bundit University, Bangkok, Thailand. His research interests analog circuit design, low voltage, high frequency and high-speed CMOS technology.

ACKNOWLEDGMENT The researchers, we are thank you very much to our parents, who has supporting everything to us. Thankfully to all of professor for knowledge and a consultant, thank you to Miss Suphansa Kansa-Ard for her time and supporting to this research. The last one we couldn’t forget that is Kasem Bundit University, Engineering Faculty for supporting and give opportunity to our to development in knowledge and research, so we are special thanks for everything. REFERENCES
[1] Barker, R.W.J., “Versatile precision full wave Rectifier,” Electron Letts, 5: Vol.13, pp. 143-144, 1977. [2] Barker, R.W.J. and Hart, B.L., “Precision absolute-value circuit technique,” Int. J. Electronics Letts, 3: Vol.66, pp. 445-448, 1989. [3] Toumazou, C. and Lidgey, F.J., “Wide-Band precision rectification,” IEE Proc. G, 1: Vol.134, pp.7-15, 1987. [4] Wang, Z., “Full-wave precision rectification that is performed in current domain and very suitable for CMOS implementation,” IEEE Trans. Circuits and Syst, 6: Part I, Vol. 39, pp.456-462, 1992. [5] Toumazou, C., Lidgey, F.J.and Chattong, S., “High frequency current conveyor precision full-wave rectifier,” Electron. Letts, 10: Vol. 30, pp. 745-746, 1994. [6] Wilson, B. and Mannama, V., “Current-mode rectifier with improved precision,” Electron. Letts, 4: Vol. 31, pp. 247-248, 1995. [7] Surakampontorn, W. and Riewruja, V., “Integrable CMOS sinusoidal frequency doubler and full-wave rectifier,” Int.J. Electronics, Letts, 3: Vol. 73, pp. 627-632, 1992. [8] Traff, H., “Novel Approach to High Speed CMOS Current Comparators,” Electron. Letts, 3: Vol. 28, pp. 310-312, 1992. [9] Monpapassorn, A., “Improved Class AB Full-Wave rectifier,” Thammasat Int. J. Sc. Tech., No. 3, November, Vol. 4, 1999. [10] A. Monpapassorn, K. Dejhan, and F.Cheevasuvit, “CMOS dual output current mode half-wave rectifier,” International Journal of Electronics, Vol. 88, 2001, pp. 1073-1084.

180

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Speech Recognition by Machine: A Review

M.A.Anusuya
Department of Computer Science and Engineering Sri Jaya chamarajendra College of Engineering Mysore, India .

S.K.Katti
Department of Computer Science and Engineering Sri Jayachamarajendra College of Engineering Mysore, India . communication and speech processing has been one of the most exciting areas of the signal processing. Speech recognition technology has made it possible for computer to follow human voice commands and understand human languages. The main goal of speech recognition area is to develop techniques and systems for speech input to machine. Speech is the primary means of communication between humans. For reasons ranging from technological curiosity about the mechanisms for mechanical realization of human speech capabilities to desire to automate simple tasks which necessitates human machine interactions and research in automatic speech recognition by machines has attracted a great deal of attention for sixty years[76]. Based on major advances in statistical modeling of speech, automatic speech recognition systems today find widespread application in tasks that require human machine interface, such as automatic call processing in telephone networks, and query based information systems that provide updated travel information, stock price quotations, weather reports, Data entry, voice dictation, access to information: travel, banking, Commands, Avoinics, Automobile portal, speech transcription, Handicapped people (blind people) supermarket, railway reservations etc. Speech recognition technology was increasingly used within telephone networks to automate as well as to enhance the operator services. This report reviews major highlights during the last six decades in the research and development of automatic speech recognition, so as to provide a technological perspective. Although many technological progresses have been made, still there remains many research issues that need to be tackled. Fig.1 shows a mathematical representation of speech recognition system in simple equations which contain front end unit, model unit, language model unit, and search unit. The recognition process is shown below (Fig .1).

Abstract This paper presents a brief survey on Automatic Speech Recognition and discusses the major themes and advances made in the past 60 years of research, so as to provide a technological perspective and an appreciation of the fundamental progress that has been accomplished in this important area of speech communication. After years of research and development the accuracy of automatic speech recognition remains one of the important research challenges (eg., variations of the context, speakers, and environment).The design of Speech Recognition system requires careful attentions to the following issues: Definition of various types of speech classes, speech representation, feature extraction techniques, speech classifiers, database and performance evaluation. The problems that are existing in ASR and the various techniques to solve these problems constructed by various research workers have been presented in a chronological order. Hence authors hope that this work shall be a contribution in the area of speech recognition. The objective of this review paper is to summarize and compare some of the well known methods used in various stages of speech recognition system and identify research topic and applications which are at the forefront of this exciting and challenging field. Key words: Automatic Speech Recognition, Statistical Modeling, Robust speech recognition, Noisy speech recognition, classifiers, feature extraction, performance evaluation, Data base. I. INTRODUCTION

A. Definition of speech recognition: Speech Recognition (is also known as Automatic Speech Recognition (ASR), or computer speech recognition) is the process of converting a speech signal to a sequence of words, by means of an algorithm implemented as a computer program. 1.2 Basic Model of Speech Recognition: Research in speech processing and communication for the most part, was motivated by people s desire to build mechanical models to emulate human verbal communication capabilities. Speech is the most natural form of human

181

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

1.3 Types of Speech Recognition Speech recognition systems can be separated in several different classes by describing what types of utterances they have the ability to recognize. These classes are classified as the following: Isolated Words: Isolated word recognizers usually require each utterance to have quiet (lack of an audio signal) on both sides of the sample window. It accepts single words or single utterance at a time. These systems have "Listen/Not-Listen" states, where they require the speaker to wait between utterances (usually doing processing during the pauses). Isolated Utterance might be a better name for this class. Connected Words: Connected word systems (or more correctly 'connected utterances') are similar to isolated words, but allows separate utterances to be 'run-together' with a minimal pause between them. Continuous Speech: Continuous speech recognizers allow users to speak almost naturally, while the computer determines the content. (Basically, it's computer dictation). Recognizers with continuous speech capabilities are some of the most difficult to create because they utilize special methods to determine utterance boundaries. Spontaneous Speech: At a basic level, it can be thought of as speech that is natural sounding and not rehearsed. An ASR system with spontaneous speech ability should be able to handle a variety of natural speech features such as words being run together, "ums" and "ahs", and even slight stutters. 1.4 Applications of Speech Recognition: Various applications of speech recognition domain have been discussed in the following table 1. .. (3) Table 1: Applications of speech recognition: Problem Domain Application Input pattern Speech/Telehphone/ Telephone Speech Communication directory enquiry wave Sector/Recognition without operator form assistance Education Sector Teaching students Speech of foreign wave languages to form pronounce vocabulary correctly. Teaching overseas students to pronounce English correctly. Pattern classes Spoken words

Fig.1 Basic model of speech recognition The standard approach to large vocabulary continuous speech recognition is to assume a simple probabilistic model of speech production whereby a specified word sequence, W, produces an acoustic observation sequence Y, with probability P(W,Y). The goal is then to decode the word string, based on the acoustic observation sequence, so that the decoded string has the maximum a posteriori (MAP) probability. ^ . (1) P(W/A)= arg maxw P(W/A) W Using Baye s rule, equation (1) can be written as P(W/A)=P(A/W)P(W) P(A) . (2)

Since P(A) is independent of W, the MAP decoding rule of equation(1) is ^ ^ W=argmaxw P(A/W)P(W)

The first term in equation (3) P(A/W), is generally called the acoustic model, as it estimates the probability of a sequence of acoustic observations, conditioned on the word string. Hence P(A/W) is computed. For large vocabulary speech recognition systems, it is necessary to build statistical models for sub word speech units, build up word models from these sub word speech unit models (using a lexicon to describe the composition of words), and then postulate word sequences and evaluate the acoustic model probabilities via standard concatenation methods. The second term in equation (3) P(W), is called the language model. It describes the probability associated with a postulated sequence of words. Such language models can incorporate both syntactic and semantic constraints of the language and the recognition task.

Spoken words

182

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Enabling students who are physically handicapped and unable to use a keyboard to enter text verbally Narrative oriented research, where transcripts are automatically generated. This would remove the time to manually generate the transcript, and human error. Computer and video games, Gambling, Precision surgery Oven, refrigerators, dishwashers and washing machines High performance fighter aircraft, Helicopters, Battle management, Training air traffic controllers, Telephony and other domains, people with disabilities Robotics

Dictation

Translation

Dictation systems on the market accepts continuous speech input which replaces menu system. It is an advanced application which translates from one language to another.

Speech wave form

Spoken words

Speech wave form

Spoken words

Outside education sector

Speech wave form Speech wave form Speech wave form

Spoken words

1.5 Automatic Speech Recognition system classification: The following tree structure emphasizes the speech processing applications. Depending on the chosen criterion, Automatic Speech Recognition systems can be classified as shown in figure 2.

Domestic sector

Spoken words

Military sector

Spoken words

Artificial Intelligence sector Medical sector

General:

Physically Handicapped

Health care, Medical Transcriptions (digital speech to text) Automated transcription, Telematics, Air traffic control, Multimodal interacting, court reporting, Grocery shops Useful to the people with limited mobility in their arms and hands or for those with sight

Speech wave form Speech wave form

Spoken words Spoken words 1.6 Relevant issues of ASR design: Main issues on which recognition accuracy depends have been presented in the table 2. Table 2: Relevant issues of ASR design Environment Type of noise; signal/noise ratio; working conditions Transducer Microphone; telephone Channel Band amplitude; distortion; echo Speakers Speaker dependence/independence Sex, Age; physical and psychical state Speech styles Voice tone(quiet, normal, shouted); Production(isolated words or continuous speech read or spontaneous speech) Speed(slow, normal, fast) Vocabulary Characteristics of available training data; specific or generic vocabulary;

Speech wave form

Spoken words

Speech wave form

Spoken words

183

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

2. Approaches to speech recognition: Basically there exist three approaches to speech recognition. They are Acoustic Phonetic Approach Pattern Recognition Approach Artificial Intelligence Approach 2.1 Acoustic phonetic approach: The earliest approaches to speech recognition were based on finding speech sounds and providing appropriate labels to these sounds. This is the basis of the acoustic phonetic approach (Hemdal and Hughes 1967), which postulates that there exist finite, distinctive phonetic units (phonemes) in spoken language and that these units are broadly characterized by a set of acoustics properties that are manifested in the speech signal over time. Even though, the acoustic properties of phonetic units are highly variable, both with speakers and with neighboring sounds (the so-called co articulation effect), it is assumed in the acoustic-phonetic approach that the rules governing the variability are straightforward and can be readily learned by a machine. The first step in the acoustic phonetic approach is a spectral analysis of the speech combined with a feature detection that converts the spectral measurements to a set of features that describe the broad acoustic properties of the different phonetic units. The next step is a segmentation and labeling phase in which the speech signal is segmented into stable acoustic regions, followed by attaching one or more phonetic labels to each segmented region, resulting in a phoneme lattice characterization of the speech. The last step in this approach attempts to determine a valid word (or string of words) from the phonetic label sequences produced by the segmentation to labeling. In the validation process, linguistic constraints on the task (i.e., the vocabulary, the syntax, and other semantic rules) are invoked in order to access the lexicon for word decoding based on the phoneme lattice. The acoustic phonetic approach has not been widely used in most commercial applications ([76], Refer fig.2.32. p.81).The following table 3 broadly gives the different speech recognition techniques.

Table 3: Speech Recognition Techniques

2.2 Pattern Recognition approach: The pattern-matching approach (Itakura 1975; Rabiner 1989; Rabiner and Juang 1993) involves two essential steps namely, pattern training and pattern comparison. The essential feature of this approach is that it uses a well formulated mathematical framework and establishes consistent speech pattern representations, for reliable pattern comparison, from a set of labeled training samples via a formal training algorithm. A speech pattern representation can be in the form of a speech template or a statistical model (e.g., a HIDDEN MARKOV MODEL or HMM) and can be applied to a sound (smaller than a word), a word, or a phrase. In the pattern-comparison stage of the approach, a direct comparison is made between the unknown speeches (the speech to be recognized) with each possible pattern learned in the training stage in order to determine the identity of the unknown according to the goodness of match of the patterns. The pattern-matching approach has become the predominant method for speech recognition in the last six decades ([76] Refer fig.2.37. pg.87). A block schematic diagram of pattern recognition is presented in fig.3 below. In this, there exists two methods namely template approach and stochastic approach.

184

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

density HMM, with identity covariance matrices and a slope constrained topology. Although templates can be trained on fewer instances, they lack the probabilistic formulation of full HMMs and typically underperforms HMMs. Compared to knowledge based approaches; HMMs enable easy integration of knowledge sources into a compiled architecture. A negative side effect of this is that HMMs do not provide much insight on the recognition process. As a result, it is often difficult to analyze the errors of an HMM system in an attempt to improve its performance. Nevertheless, prudent incorporation of knowledge has significantly improved HMM based systems. 2.2.1. Template Based Approach: Template based approach [97] to speech recognition have provided a family of techniques that have advanced the field considerably during the last six decades. The underlying idea is simple. A collection of prototypical speech patterns are stored as reference patterns representing the dictionary of candidate s words. Recognition is then carried out by matching an unknown spoken utterance with each of these reference templates and selecting the category of the best matching pattern. Usually templates for entire words are constructed. This has the advantage that, errors due to segmentation or classification of smaller acoustically more variable units such as phonemes can be avoided. In turn, each word must have its own full reference template; template preparation and matching become prohibitively expensive or impractical as vocabulary size increases beyond a few hundred words. One key idea in template method is to derive a typical sequences of speech frames for a pattern(a word) via some averaging procedure, and to rely on the use of local spectral distance measures to compare patterns. Another key idea is to use some form of dynamic programming to temporarily align patterns to account for differences in speaking rates across talkers as well as across repetitions of the word by the same talker. 2.2.2. Stochastic Approach: Stochastic modeling [97] entails the use of probabilistic models to deal with uncertain or incomplete information. In speech recognition, uncertainty and incompleteness arise from many sources; for example, confusable sounds, speaker variability s, contextual effects, and homophones words. Thus, stochastic models are particularly suitable approach to speech recognition. The most popular stochastic approach today is hidden Markov modeling. A hidden Markov model is characterized by a finite state markov model and a set of output distributions. The transition parameters in the Markov chain models, temporal variabilities, while the parameters in the output distribution model, spectral variabilities. These two types of variabilites are the essence of speech recognition. Compared to template based approach, hidden Markov modeling is more general and has a firmer mathematical foundation. A template based model is simply a continuous 2.3. Dynamic Time Warping(DTW): Dynamic time warping is an algorithm for measuring similarity between two sequences which may vary in time or speed. For instance, similarities in walking patterns would be detected, even if in one video, the person was walking slowly and if in another, he or she were walking more quickly, or even if there were accelerations and decelerations during the course of one observation. DTW has been applied to video, audio, and graphics indeed, any data which can be turned into a linear representation can be analyzed with DTW. A well known application has been automatic speech recognition, to cope with different speaking speeds.In general, DTW is a method that allows a computer to find an optimal match between two given sequences (e.g. time series) with certain restrictions. The sequences are "warped" non-linearly in the time dimension to determine a measure of their similarity independent of certain non-linear variations in the time dimension. This sequence alignment method is often used in the context of hidden Markov models. One example of the restrictions imposed on the matching of the sequences is on the monotonicity of the mapping in the time dimension. Continuity is less important in DTW than in other pattern matching algorithms; DTW is an algorithm particularly suited to matching sequences with missing information, provided there are long enough segments for matching to occur. The optimization process is performed using dynamic programming, hence the name. 2.4. Vector Quantization(VQ): Vector Quantization(VQ)[97] is often applied to ASR. It is useful for speech coders, i.e., efficient data reduction. Since transmission rate is not a major issue for ASR, the utility of VQ here lies in the efficiency of using compact codebooks for reference models and codebook searcher in place of more costly evaluation methods. For IWR, each vocabulary word gets its own VQ codebook, based on training sequence of several repetitions of the word. The test speech is evaluated by all codebooks and ASR chooses the word whose codebook yields the lowest distance measure. In basic VQ, codebooks have no explicit time information (e.g., the temporal order of phonetic segments in each word and their relative durations are ignored) , since codebook entries are not ordered and can

185

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

come from any part of the training words. However, some indirect durational cues are preserved because the codebook entries are chosen to minimize average distance across all training frames, and frames, corresponding to longer acoustic segments ( e.g., vowels) are more frequent in the training data. Such segments are thus more likely to specify code words than less frequent consonant frames, especially with small codebooks. Code words nonetheless exist for constant frames because such frames would otherwise contribute large frame distances to the codebook. Often a few code words suffice to represent many frames during relatively steady sections of vowels, thus allowing more codeword to represent short, dynamic portions of the words. This relative emphasis that VQ puts on speech transients can be an advantage over other ASR comparison methods for vocabularies of similar words. 2.5. Artificial Intelligence approach (Knowledge Based approach) The Artificial Intelligence approach [97] is a hybrid of the acoustic phonetic approach and pattern recognition approach. In this, it exploits the ideas and concepts of Acoustic phonetic and pattern recognition methods. Knowledge based approach uses the information regarding linguistic, phonetic and spectrogram. Some speech researchers developed recognition system that used acoustic phonetic knowledge to develop classification rules for speech sounds. While template based approaches have been very effective in the design of a variety of speech recognition systems; they provided little insight about human speech processing, thereby making error analysis and knowledge-based system enhancement difficult. On the other hand, a large body of linguistic and phonetic literature provided insights and understanding to human speech processing. In its pure form, knowledge engineering design involves the direct and explicit incorporation of expert s speech knowledge into a recognition system. This knowledge is usually derived from careful study of spectrograms and is incorporated using rules or procedures. Pure knowledge engineering was also motivated by the interest and research in expert systems. However, this approach had only limited success, largely due to the difficulty in quantifying expert knowledge. Another difficult problem is the integration of many levels of human knowledge phonetics, phonotactics, lexical access, syntax, semantics and pragmatics. Alternatively, combining independent and asynchronous knowledge sources optimally remains an unsolved problem. In more indirect forms, knowledge has also been used to guide the design of the models and algorithms of other techniques such as template matching and stochastic modeling. This form of knowledge application makes an important distinction between knowledge and algorithms Algorithms enable us to solve problems. Knowledge enable the algorithms to work better. This form of knowledge based system enhancement has contributed considerably to the design of all successful strategies reported. It plays an important role in the selection of a suitable input representation, the definition of units of speech, or the design of the recognition algorithm itself.

2.6. Connectionist Approaches (Artificial Neural Networks): The artificial intelligence approach ( [97], Lesser et al. 1975; Lippmann 1987) attempts to mechanize the recognition procedure according to the way a person applies intelligence in visualizing, analyzing, and characterizing speech based on a set of measured acoustic features. Among the techniques used within this class of methods are use of an expert system (e.g., a neural network) that integrates phonemic, lexical, syntactic, semantic, and even pragmatic knowledge for segmentation and labeling, and uses tools such as artificial NEURAL NETWORKS for learning the relationships among phonetic events. The focus in this approach has been mostly in the representation of knowledge and integration of knowledge sources. This method has not been widely used in commercial systems. Connectionist modeling of speech is the youngest development in speech recognition and still the subject of much controversy. In connectionist models, knowledge or constraints are not encoded in individual units, rules, or procedures, but distributed across many simple computing units. Uncertainty is modeled not as likelihoods or probability density functions of a single unit, but by the pattern of activity in many units. The computing units are simple in nature, and knowledge is not programmed into any individual unit function; rather, it lies in the connections and interactions between linked processing elements. Because the style of computation that can be performed by networks of such units bears some resemblance to the style of computation in the nervous system. Connectionist models are also referred to as neural networks or artificial neural networks. Similarly, parallel distributed processing or massively distributed processing are terms used to describe these models. Not unlike stochastic models, connectionist models rely critically on the availability of good training or learning strategies. Connectionist learning seeks to optimize or organize a network of processing elements. However, connectionist models need not make assumptions about the underlying probability distributions. Multilayer neural networks can be trained to generate rather complex nonlinear classifiers or mapping function. The simplicity and uniformity of the underlying processing element makes connectionist models attractive for hardware implementation, which enables the operation of a net to be simulated efficiently. On the other hand, training often requires much iteration over large amounts of training data, and can, in some cases, be prohibitively expensive. While connectionism appears to hold great promise as plausible model of cognition, may question relating to the concrete realization of practical connectionist recognition techniques, still remain to be resolved. 2.7. Support Vector Machine(SVM): One of the powerful tools for pattern recognition that uses a discriminative approach is a SVM[97]. SVMs use linear and nonlinear separating hyper-planes for data classification. However, since SVMs can only classify fixed length data vectors, this method cannot be readily applied to task involving variable length data classification. The variable

186

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

length data has to be transformed to fixed length vectors before SVMs can be used. It is a generalized linear classifier with maximum-margin fitting functions. This fitting function provides regularization which helps the classifier generalized better. The classifier tends to ignore many of the features. Conventional statistical and Neural Network methods control model complexity by using a small number of features ( the problem dimensionality or the number of hidden units). SVM controls the model complexity by controlling the VC dimensions of its model. This method is independent of dimensionality and can utilize spaces of very large dimensions spaces, which permits a construction of very large number of non-linear features and then performing adaptive feature selection during training. By shifting all non-linearity to the features, SVM can use linear model for which VC dimensions is known. For example, a support vector machine can be used as a regularized radial basis function classifier. 2.8. Taxonomy of Speech Recognition: Existing techniques for speech recognition have been represented diagrammatically in the following figure 4.

automatic creation of acoustic models for these sounds without the need for an excessive amount of training data, and they should exhibit statistics which are largely invariant across speakers and speaking environment. 3.1.Various methods for Feature Extraction in speech recognition are broadly shown in the following table 4. Table 4: feature extraction methods Method Principal Component Analysis(PCA) Property Non linear feature extraction method, Linear map; fast; eigenvector-based Comments Traditional, eigenvector based method, also known as karhuneu-Loeve expansion; good for Gaussian data. Better than PCA for classification;

Linear Discriminant Analysis(LDA)

Independent Component Analysis (ICA)

Non linear feature extraction method, Supervised linear map; fast; eigenvector-based Non linear feature extraction method, Linear map, iterative nonGaussian Static feature extraction method,10 to 16 lower order coefficient,

Blind course separation, used for de-mixing nonGaussian distributed sources(features)

Linear Predictive coding

Cepstral Analysis 3. Feature Extraction: In speech recognition, the main goal of the feature extraction step is to compute a parsimonious sequence of feature vectors providing a compact representation of the given input signal. The feature extraction is usually performed in three stages. The first stage is called the speech analysis or the acoustic front end. It performs some kind of spectro temporal analysis of the signal and generates raw features describing the envelope of the power spectrum of short speech intervals. The second stage compiles an extended feature vector composed of static and dynamic features. Finally, the last stage( which is not always present) transforms these extended feature vectors into more compact and robust vectors that are then supplied to the recognizer. Although there is no real consensus as to what the optimal feature sets should look like, one usually would like them to have the following properties: they should allow an automatic system to discriminate between different through similar sounding speech sounds, they should allow for the

Mel-frequency scale analysis

Static feature extraction method, Power spectrum Static feature extraction method, Spectral analysis

Used to represent spectral envelope Spectral analysis is done with a fixed resolution along a subjective frequency scale i.e. Mel-frequency scale.

Filter bank analysis Mel-frequency cepstrum (MFFCs) Kernel based feature extraction method

Filters tuned required frequencies Power spectrum is computed by performing Fourier Analysis Non linear transformations,

Dimensionality reduction leads to better classification and it is used to

187

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Wavelet

Better time resolution than Fourier Transform

remove noisy and redundant features, and improvement in classification error It replaces the fixed bandwidth of Fourier transform with one proportional to frequency which allow better time resolution at high frequencies than Fourier Transform

Fig.4a. Classification example Once a feature selection or classification procedure finds a proper representation, a classifier can be designed using a number of possible approaches. In practice, the choice of a classifier is a difficult problem and it is often based on which classifier(s) happen to be available, or best known, to the user. The three different approaches are identified to design a classifier. The simplest and the most intuitive approach to classifier design is based on the concept of similarity: patterns that are similar should be assigned to the same class. So, once a good metric has been established to define similarity, patterns can be classified by template matching or the minimum distance classifier using a few prototypes per class. The choice of the metric and the prototypes is crucial to the success of this approach. In the nearest mean classifier, selecting prototypes is very simple and robust; each pattern class is represented by a single prototype which is the mean vector of all the training patterns in that class. More advanced techniques for computing prototypes are vector quantization [154, 155] and Learning Vector Quantization [156], and the data reduction methods associated with the one-nearest neighbor decision rule (1-NN) such as editing and condensing [157]. The most straightforward 1-NN rule can be conveniently used as a benchmark for all the other classifiers since it appears to always provide a reasonable classification performance in most applications. Further, as the 1-NN classifier does not require any user-specified parameters (except perhaps the distance metric used to find the nearest neighbor, but Euclidean distance is commonly used), its classification results are implementation independent. In many classification problems, the classifier is expected to have some desired invariant properties. An example is the shift invariance of characters in character recognition, a change in a character's location should not affect its classification. If the preprocessing or the representation scheme does not normalize the input pattern for this invariance, then the same character may be represented at multiple positions in the feature space. These positions define a one-dimensional subspace. As more invariants are considered, the dimensionality of this subspace correspondingly increases. Template matching or the nearest mean classifier can be viewed as finding the nearest subspace [158]. The second main concept used for designing pattern classifiers is based on the probabilistic approach. The optimal Bayes decision rule (with the 0/1 loss function) assigns a pattern to the class with the maximum posterior probability. This rule can be modified to take into account, costs

Dynamic feature extractions i)LPC ii)MFCCs

Spectral subtraction Cepstral mean subtraction RASTA filtering Integrated Phoneme subspace method

Acceleration and delta coefficients i.e. II and III order derivatives of normal LPC and MFCCs coefficients Robust Feature extraction method Robust Feature extraction For Noisy speech A transformation based on PCA+LDA+ICA

Higher Accuracy than the existing methods

4. Classifiers [149]: In speech recognition a supervised pattern classification system is trained with labeled examples; that is, each input pattern has a class label associated with it. Pattern classifiers can also be trained in an unsupervised fashion. For example in a technique known as vector quantization, some representation of the input data is clustered by finding implicit groupings in the data. The resulting table of cluster centers is known as a codebook, which can be used to index new vectors by finding the cluster center that is closest to the new vectors. For the case of speech, fig.4a. shows an extreme case of some vowels represented by their formant frequencies F1 and F2. The vowels represented are, as pronounced in the words bot(/a/) and boot (/u/). Notice that they fall into nice groupings.

188

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

associated with different types of classifications. For known class conditional densities, the Bayes decision rule gives the optimum classifier, in the sense that for given prior probabilities, loss function and class-conditional densities, no other decision rule will have a lower risk (i.e., expected value of the loss function, for example, probability of error). If the prior class probabilities are equal and a 0/1 loss function, the Bayes decision rule and the maximum likelihood decision rule exactly coincide. In practice, the empirical Bayes decision rule, or plug-in rule, is used. The estimates of the densities are used in place of the true densities. These density estimates are either parametric or nonparametric. Commonly used parametric models are multivariate Gaussian distributions [159] for continuous features, binomial distributions for binary features, and multi-normal distributions for integer-valued (and categorical) features. A critical issue for Gaussian distributions is the assumption made about the covariance matrices. If the covariance matrices for different classes are assumed to be identical, then the Bayes plug-in rule, called Bayes normal-linear, provides a linear decision boundary. On the other hand, if the covariance matrices are assumed to be different, the resulting Bayes plug-in rule, which we call Bayes-normal-quadratic, provides a quadratic decision boundary. In addition to the commonly used maximum likelihood estimator of the covariance matrix, various regularization techniques [160] are available to obtain a robust estimate in small sample size situations and the leave-one-out estimator is available for minimizing the bias [161]. 5. Performance of speech recognition systems: The performance of speech recognition systems is usually specified in terms of accuracy and speed. Accuracy may be measured in terms of performance accuracy which is usually rated with word error rate (WER), whereas speed is measured with the real time factor. Other measures of accuracy include Single Word Error Rate (SWER) and Command Success Rate (CSR). Word Error Rate(WER):Word error rate is a common metric of the performance of a speech recognition or machine translation system. The general difficulty of measuring performance lies in the fact that the recognized word sequence can have a different length from the reference word sequence (supposedly the correct one). The WER is derived from the Levenshtein distance, working at the word level instead of the phoneme level. This problem is solved by first aligning the recognized word sequence with the reference (spoken) word sequence using dynamic string alignment. Word error rate can then be computed as:

D is the number of the deletions, I is the number of the insertions, N is the number of words in the reference. When reporting the performance of a speech recognition system, sometimes word recognition rate (WRR) is used instead:

..(10) where H is N-(S+D), the number of correctly recognized words. 6. Literature Survey of speech recognition: ( year vise): 6.1 1920-1960s: In the early 1920s machine recognition came into existence. The first machine to recognize speech to any significant degree commercially named, Radio Rex (toy) was manufactured in 1920[165].Research into the concepts of speech technology began as early as 1936 at Bell Labs. In 1939, Bell Labs demonstrated a speech synthesis machine (which simulates talking) at the World Fair in New York. Bell Labs later abandoned efforts to develop speech-simulated listening and recognition; based on an incorrect conclusion that artificial intelligence would ultimately be necessary for success. The earliest attempts to devise systems for automatic speech recognition by machine were made in 1950s, when various researchers tried to exploit the fundamental ideas of acoustic phonetics. During 1950s[1], most of the speech recognition systems investigated spectral resonances during the vowel region of each utterance which were extracted from output signals of an analogue filter bank and logic circuits. In 1952, at Bell laboratories, Davis, Biddulph, and Balashek built a system for isolated digit recognition for a single speaker [2]. The system relied heavily on measuring spectral resonances during the vowel region of each digit. In an independent effort at RCA Laboratories in 1956, Olson and Belar tried to recognize 10 distinct syllables of a single talker, as embodied in 10 monosyllabic words [3]. The system again relied on spectral measurements (as provided by an analog filter bank) primarily during vowel regions. In 1959, at University College in England, Fry and Denes tried to build a phoneme recognizer to recognize four vowels and nine consonants [4]. They used a spectrum analyzer and a pattern matcher to make the recognition decision. A novel aspect of this research was the use of statistical information about allowable sequences of phonemes in English ( a rudimentary form of language syntax) to improve overall phoneme accuracy for words consisting of two or more phonemes. Another effort of note in this period was the vowel recognizer of Forgie and Forgie constructed at

.(9) where S is the number of substitutions,

189

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

MIT Licoln laboratories in 1959 in which 10 vowels embedded in a /b/-vowel/t/ format were recognized in a speaker independent manner [5]. Again a Filter bank analyzer was used to provide spectral information and a time varying estimate of the vocal tract resonances was made to deicide which vowel was spoken. 6.2 1960-1970: In the 1960s several fundamental ideas in speech recognition surfaced and were published. In the 1960s since computers were still not fast enough, several special purpose hardware were built [6]. However, the decade started with several Japanese laboratories entering the recognition arena and building special purpose hardware as part of their systems. On early Japanese system, described by Suzuki and Nakata of the Radio Research Lab in Tokyo, was a hardware vowel recognizer [7]. An elaborate filter bank spectrum analyzer was used along with logic that connected the outputs of each channel of the spectrum analyzer (in a weighted manner) to a vowel decision circuit, and majority decisions logic scheme was used to choose the spoken vowel. Another hardware effort in Japan was the work of Sakai and Doshita of kyoto University in 1962, who built a hardware phoneme recognizer [7]. A hardware speech segmented was used along with a zero crossing analysis of different regions of the spoken input to provide the recognition output. A third Japanese effort was the digit recognizer hardware of Nagata and coworkers at NEC Laboratories in 1963[8]. This effort was perhaps most notable as the initial attempt at speech recognition at NEC and led to a long and highly productive research program. One of the difficult problems of speech recognition exists in the non uniformity of time scales in speech events. In the 1960s three key research projects were initiated that have had major implications on the research and development of speech recognition for the past 20 years. The first of these projects was the efforts of Martin and his colleagues at RCA Laboratories, beginning in the late 1960s, to develop realistic solutions to the problems associated with non-uniformity of time scales in speech events. Martin developed a set of elementary time normalization methods, based on the ability to reliably detect speech starts and ends, that significantly reduce the variability of the recognition scores[9]. Martin ultimately developed the method and founded one of the first speech recognition companies, Threshold Technology, which was built, marketed and was sold speech recognition products. At about the same time, in the Soviet Union, Vintsyuk proposed the use of dynamic programming methods for time aligning a pair of speech utterances(generally known as Dynamic Time Warping(DTW) [10]),including algorithms for connected word recognition.. Although the essence of the concepts of dynamic time warping, as well as rudimentary versions of the algorithms for connected word recognition, were embodied in Vintsyuk s work, it was largely unknown in the West and did not come to light until the early 1980 s; this was long after the more formal methods were proposed and implemented by others. At the same time in an independent effort in Japan Sakoe and Chiba at NEC Laboratories also

started to use a dynamic Programming technique to solve the non uniformity problems[11].A final achievement of note in the 1960s was the pioneering research of Reddy in the field of continuous speech recognition by dynamic tracking of phonemes [12]. Reddy s research eventually spawned a long and highly successful speech recognition research program at Carnegie Mellon University (to which Reddy moved in the late 1960s) which, to this today, remains a world leader in continuous speech recognition systems. 6.3 1970-1980: In the 1970s speech recognition research achieved a number of significant milestones. First the area of isolated word or discrete utterance recognition became a viable and usable technology based on fundamental studies by Velichko and Zagoruyko in Russia[13], Cakoe and Chiba in Japan[14], and Itakura in the united States. The Russian studies helped the advance use of pattern recognition ideas in speech recognition; the Japanese research showed how dynamic programming methods could be successfully applied; and Itakura s research showed how the ideas of linear predictive coding (LPC), which had already been successfully used in low bit rate speech coding, could be extended to speech recognition systems through the use of an appropriate distance measure based on LPC spectral parameters[15].Another milestone of the 1970s was the beginning of a longstanding, highly successful group effort in large vocabulary speech recognition at IBM in which researchers studied three distinct tasks over a period of almost two decades, namely the New Raleigh language [16] for simple database queries, the laser patent text language [17] for transcribing laser patents, and the office correspondent tasks called Tangora[18], for dictation of simple memos. Finally, at AT&T Bell Labs, researchers began a series of experiments aimed at making speech recognition systems that were truly speaker independent [19]. To achieve this goal a wide range of sophisticated clustering algorithms were used to determine the number of distinct patterns required to represent all variations of different words across a wide user population. This research has been refined over a decade so that the techniques for creating speaker independent patterns are now well understood and widely used. An ambitious speech understanding project was funded by the defence Advanced Research Projects Agencies(DARPA), which led to many seminal systems and technology[20]. One of the demonstrations of speech understanding was achieved by CMU in 1973 there Heresay I system was able to use semantic information to significantly reduce the number of alternatives considered by the recognizer.CMU s Harphy system[21] was shown to be able to recognize speech using a vocabulary of 1,011 words with reasonable accuracy. One of the particular contributions from the Harpy system was the concept of graph search, where the speech recognition language is represented as a connected network derived from lexical representations of words, with syntactical production rules and word boundary rules. The Harpy system was the first to take advantage of a finite state network (FSN) to reduce computation and efficiently determine the closest matching

190

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

string. Other systems developed under the DARPA s speech understanding program included CMU s Hearsay II and BBN s HWIM (Hear what I Mean) systems[20]. The approach proposed by Hearsay II of using parallel asynchronous proceses that simulate the component knowledge sources in a speech system was a pioneering concept. A global blackboard was used to integrate knowledge from parallel sources to produce the next level of hypothesis. 6.4 1980-1990: Just as isolated word recognition was a key focus of research in the 1970s, the problems of connected word recognition was a focus of research in the 1980s. Here the goal was to create a robust system capable of recognizing a fluently spoken string of words(eg., digits) based on matching a concatenated pattern of individual words. Moshey J. Lasry has developed a featurebased speech recognition system in the beginning of 1980. Wherein his studies speech spectrograms of letters and digits[97].A wide variety of the algorithm based on matching a concatenated pattern of individual words were formulated and implemented, including the two level dynamic programming approach of Sakoe at Nippon Electric Corporation (NEC)[22],the one pass method of Bridle and Brown at Joint Speech Research Unit(JSRU) in UK[23], the level building approach of Myers and Rabiner at Bell Labs [24], and the frame synchronous level building approach of Lee and Rabiner at Bell Labs[25]. Each of these optimal matching procedures had its own implementation advantages, which were exploited for a wide range of tasks. Speech research in the 1980s was characterized by a shift in technology from template based approaches to statistical modeling methods especially the hidden Markov model approach [26,27]. Although the methodology of hidden Markov modeling (HMM) was well known and understood in a few laboratories(Primarily IBM, Institute for Defense Analyses (IDA), and Dargon systems), it was not until widespread publication of the methods and theory of HMMs, in the mid1980, that the technique became widely applied in virtually, every speech recognition research laboratory in the world. Today, most practical speech recognition systems are based on the statistical frame work developed in the 1980s and their results, with significant additional improvements have been made in the 1990s. a) Hidden Markov Model(HMM): HMM is one of the key technologies developed in the 1980s, is the hidden Markov model(HMM) approach [28,29,30]. It is a doubly stochastic process which as an underlying stochastic process that is not observable (hence the term hidden), but can be observed through another stochastic process that produces a sequence of observations. Although the HMM was well known and understood in a few laboratories (primarily IBM, Institute for Defense Analysis (IDA) and Dragon Systems), it was not until widespread publication of the methods and theory of HMMs in the mid-1980s that the technique became widely applied in virtually every speech recognition research laboratory in the world. In the early 1970s, Lenny Baum of

Princeton University invented a mathematical approach to recognize speech called Hidden Markov Modeling (HMM). The HMM pattern-matching strategy was eventually adopted by each of the major companies pursuing the commercialization of speech recognition technology (SRT).The U.S. Department of Defense sponsored many practical research projects during the 70s that involved several contractors, including IBM, Dragon, AT&T, Philips and others. Progress was slow in those early years. b) Neural Net: Another new technology that was reintroduced in the late 1980s was the idea of applying neural networks to problems in speech recognition. Neural networks were first introduced in the 1950s, but they did not prove useful initially because they had many practical problems. In the 1980s however, a deeper understanding of the strengths and limitations of the technology was achieved, as well as, understanding of the technology to classical signal classification methods. Several new ways of implementing systems were also proposed [33,34,35]. c) DARPA Program: Finally, the 1980s was a decade in which a major impetus was given to large vocabulary, continuous speech recognition systems by the Defense Advanced Research Projects Agency (DARPA) community, which sponsored a large research program aimed at achieving high word accuracy for a 1000 word continuous speech recognition, database management task. Major research contributions resulted from efforts at CMU(notably the well known SPHINX system)[36], BBN with the BYBLOS system[37], Lincoln Labs[38], SRI[39], MIT[40], and AT&T Bell Labs[41]. The SPHINX system successfully integrated the statistical method of HMM with the network search strength of the earlier Harpy system. Hence, it was able to train and embed context dependent phone models in a sophisticated lexical decoding network. The DARPA program has continued into the 1990s, with emphasis shifting to natural language front ends to the recognizer and the task shifting to retrieval of air travel information. At the same time, speech recognition technology has been increasingly used within telephone networks to automate as well as enhance operator services. 6.5 1990-2000s: In the 1990s a number of innovations took place in the field of pattern recognition. The problem of pattern recognition, which traditionally followed the framework of Bayes and required estimation of distributions for the data, was transformed into an optimization problem involving minimization of the empirical recognition error [42]. This fundamental paradigmatic change was caused by the recognition of the fact that the distribution functions for the speech signal could not be accurately chosen or defined and the Bayes decision theory becomes inapplicable under these circumstances. Fundamentally, the objective of a recognizer design should be to achieve the least recognition error rather than provide the

191

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

best fitting of a distribution function to the given (known)data set as advocated by the Bayes criterion. This error minimization concept produced a number of techniques such as discriminative training and kernel based methods. As an example of discriminative training, the Minimum Classification Error(MCE) criterion was proposed along with a corresponding Generalized Probabilistic Descent(GPD) training algorithm to minimize an objective function which acts to approximate the error rate closely[43]. Another example was the Maximum Mutual Information (MMI) criterion. In MMI training, the mutual information between the acoustic observation and its correct lexical symbol averaged over a training set is maximized. Although this criterion is not based on a direct minimization of the classification error rate and is quite different from the MCE based approach, it is well founded in information theory and possesses good theoretical properties. Both the MMI and MCE can lead to speech recognition performance superior to the maximum likelihood based approach [43]. A key issue[82] in the design and implementation of speech recognition system is how to properly choose the speech material used to train the recognition algorithm. Training may be more formally defined as supervised learning of parameters of primitive speech patterns ( templates, statistical models, etc.,) used to characterize basic speech units ( e.g. word or subword units), using labeled speech samples in the form of words and sentences. It also discusses two methods for generating training sets. The first, uses a nondeterministic statistical method to generate a uniform distribution of sentences from a finite state machine represented in digraph form. The second method, a deterministic heuristic approach, takes into consideration the importance of word ordering to address the problem of co articulation effects that are necessary for good training. The two methods are critically compared. a) DARPA program: The DARPA program continued into the 1990s, with emphasis shifting to natural language front ends to the recognizer. The central focus also shifted to the task of retrieving air travel information, the Air Travel Information Service (ATIS) task. Later the emphasis was expanded to a range of different speech-understanding application areas, in conjunction with a new focus on transcription of broadcast news (BN) and conversational speech. The Switchboard task is among the most challenging ones proposed by DARPA; in this task speech is conversational and spontaneous, with many instances of so-called disfluencies such as partial words, hesitation and repairs. The BN transcription technology was integrated with information extraction and retrieval technology, and many application systems, such as automatic voice document indexing and retrieval systems, were developed. A number of human language technology projects funded by DARPA in the 1980s and 1990s further accelerated the progress, as evidenced by many papers published in the proceedings of the DARPA Speech and Natural Language/Human Language Workshop. The research describes the development of activities for speech recognition

that were conducted in the 1990s[83], at Fujitsu Laboratories Limited. Also, it is focused on extending the functions and performance of speech recognition technologies developed in the 1980s. Advnaces in small implementations of speech recognition, recognition of continuous speech, and recognition of speech in noisy environments, have been described. b) HMM : A weighted hidden markov model HMM algorithm and a subspace projection algorithm are proposed in[109], to address the discrimination and robustness issues for HMM based speech recognition. Word models were constructed for combining phonetic and fenonic models[110] A new hybrid algorithm based on combination of HMM and learning vector were proposed in [111]. Learning Vector Quantisation[112] (LVQ) method showed an important contribution in producing highly discriminative reference vectors for classifying static patterns. The ML estimation of the parameters via FB algorithm was an inefficient method for estimating the parameters values of HMM. To over come this problem paper[113] proposed a corrective training method that minimized the number of errors of parameter estimation. A novel approach [114] for a hybrid connectionist HMM speech recognition system based on the use of a Neural Network as a vector qantiser. showed the important innovations in training the Neural Network. Next the Vector Quantization approach showed much of its significance in the reduction of Word error rate. MVA[115] method obtained from modified Maximum Mutual Information(MMI) is shown in this paper. Nam Soo Kim et.al., have presented various methods for estimating a robust output probability distribution(PD) in speech recognition based on the discrete Hidden Markov Model(HMM) in their paper[118].An extension of the viterbi algorithm[120] made the second order HMM computationally efficient when compared with the existing viterbi algorithm. In paper[123] a general stochastic model that encompasses most of the models proposed in the literature, pointing out similarities of the models in terms of correlation and parameter time assumptions, and drawing analogies between segment models and HMMs have been described. An alternative VQ[124] method in which the phoneme is treated as a cluster in the speech space and a Gaussian model was estimated for each phoneme. The results showed that the phoneme-based Gaussian modeling vector quantization classifies the speech space more effectively and significant improvements in the performance of the DHMM system have been achieved. The trajectory folding phenomenon in HMM model is overcome by using Continuous Density HMM which significantly reduced the Word Error Rate over continuous speech signal as been demonstrated by[125]. A new hidden Markov model[127] showed the integration of the generalized dynamic feature parameters into the model structure was developed and evaluated using maximum-likelihood (ML) and minimum-classification-error (MCE) pattern recognition approaches. The authors have designed the loss function for minimizing error rate specifically for the new model, and derived an analytical form of the gradient of the loss function

192

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

that enables the implementation of the MCE approach. Authors extend[128] previously proposed quasi-Bayes adaptive learning framework to cope with the correlated continuous density hidden Markov models (HMM s) with Gaussian mixture state observation densities to implement the correlated mean vectors to be updated using successive approximation algorithm. Paper [130] investigates the use of Gaussian selection (GS) to increase the speed of a large vocabulary speech recognition system. The aim of GS is to reduce the load by selecting the subset of Gaussian component likelihoods for a given input vector, which also proposes new techniques for obtaining good Gaussian subsets or shortlists a novel framework of online hierarchical transformation of hidden Markov model (HMM) parameters[133], for adaptive speech recognition. Its goal is to incrementally transform (or adapt) all the HMM parameters to a new acoustical environment even though most of HMM units are unseen in observed adaptation data. The theoretical frame work[117] for Bayesian adaptive training of the parameters of discrete hidden markov model(DHMM) and semi continuous HMM(SCHMM) with Gaussian mixture state observation densities were proposed. The proposed MAP algorithms discussed in [117] are shown to be effective especially in the cases in which the training or adaptation data are limited. c) Robust speech recognition: Various techniques were investigated to increase the robustness of speech recognition systems against the mismatch between training and testing conditions, caused by background noises, voice individuality, microphones, transmission channels, room reverberation, etc. Major techniques include, the maximum likelihood linear regression (MLLR) [44], the model decomposition [45], parallel model composition (PMC) [46], and the structural maximum a posteriori (SMAP) method [47] for robust speech recognition. The paper by Mazin G.Rahim et.al[116] presents a signal bias removal (SBR) method based on maximum likelihood estimation for the minimization of the undesirable effects which occur in telephone speech recognition system such as ambient noise, channel distortions etc.,. A maximum likelihood (ML) stochastic matching approach to decrease the acoustic mismatch between a test utterances, and a given set of speech models was proposed in [121] to reduce the recognition performance degradation caused by distortions in the test utterances and/or the model set. A new approach to an auditory model for robust speech recognition for noisy environments was proposed in [129] . The proposed model consists of cochlear bandpass filters and nonlinear operations in which frequency information of the signal is obtained by zero-crossing intervals. Compared with other auditory models, the proposed auditory model is computationally efficient, free from many unknown parameters, and able to serve as a robust front-end for speech recognition in noisy environments. Uniform distribution, is adopted to characterize the uncertainty of the mean vectors of the CDHMM s in [131].The paper proposed two methods, namely, a model compensation

technique based on Bayesian predictive density and a robust decision strategy called Viterbi Bayesian predictive classi fication are studied. The proposed methods are compared with the conventional Viterbi decoding algorithm in speakerindependent recognition experiments on isolated digits and TI connected digit strings (TIDIGITS), where the mismatches between training and testing conditions are caused by: 1) additive Gaussian white noise, 2) each of 25 types of actual additive ambient noises, and 3) gender difference. A novel implementation of a minimax decision rule for continuous density hidden Markov model-based robust speech recognition was proposed in [133]. By combining the idea of the mini-max decision rule with a normal Viterbi search, authors derive a recursive mini-max search algorithm, where the mini-max decision rule is repetitively applied to determine the partial paths during the search procedure. d) Noisy speech recognition: Not much work has been done on noisy speech recognition in this decade. One of the important methods called minimum mean square error (MMSE) estimate of the filter log energies, introducing a significant improvement over existing algorithms were proposed by Adoram Erell and et.al. [98].A model based spectral estimation algorithm is derived that improves the robustness of SR system to additive noise. The algorithm is tailored for filter bank based system, where the estimation should seek to minimize the distortion as measured by the recognizer s distance [99]. Minor work has been done in the area of noisy robust speech recognition. A model based spectral estimation algorithm has been derived in [112] which improves the robustness of the speech recognition system to additive noise. This algorithm is tailored for filter bank based systems where the estimation should seek to minimize the distortions as measured by the recognizers distance metric. The aim of this correspondence [126] is to present a robust representation of speech based on AR modeling of the causal part of the autocorrelation sequence. In noisy speech recognition, this new representation achieves better results than several other related techniques. 6.6. 2000-2009: a) General: Around 2000, a variational Bayesian (VB) estimation and clustering techniques were developed[71]. Unlike Maximum Likelihood, this VB approach is based on a posterior distribution of parameters. Giuseppe Richardi[73] have developed the technique to solve the problem of adaptive learning, in automatic speech recognition and also proposed active learning algorithm for ASR. In 2005, some improvements have been worked out on Large Vocabulary Continuous Speech Recognition[74] system on performance improvement. In 2007, the difference in acoustic features between spontaneous and read speech using a large scale speech data base i.e, CSJ have been analyzed[79]. Sadaoki Furui [81] investigated SR methods that can adapt to speech variation using a large number of models trained based on

193

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

clustering techniques. In 2008, the authors[87] have explored the application of Conditional Random Field(CRF) to combine local posterior estimates provided by multilayer perceptions corresponding to the frame level prediction of phone and phonological attributed classes. De-wachter et.al.[100], attempted to over-come the time dependencies, problems in speech recognition by using straight forward template matching method. Xinwei Li et.al.[105], proposed a new optimization method i.e., semi definite programming(SDP) to solve the large margin estimation(LME) problem of continuous density HMM(CDHMM) in speech recognition. Discriminate training of acoustic models for speech recognition was proposed under Maximum mutual information(MMI)[107]. Around 2007 Rajesh M.Hegde et.al, [106], proposed an alternative method for processing the Fourier transform phase for extraction speech features, which process the group delay feature(GDF) that can be directly computed for the speech signal. b) DARPA program: The Effective Affordable Reusable Speech-to-Text (EARS) program was conducted to develop speech-to-text (automatic transcription) technology with the aim of achieving substantially richer and much more accurate output than before. The tasks include detection of sentence boundaries, fillers and disfluencies. The program was focusing on natural, unconstrained human speech from broadcasts and foreign conversational speech in multiple languages. The goal was to make it possible for machines to do a much better job of detecting, extracting, summarizing and translating important information, thus enabling humans to understand what was said by reading transcriptions instead of listening to audio signals [48, 49]. c) Spontaneous speech recognition: Although read speech and similar types of speech, e.g. news broadcasts reading a text, can be recognized with accuracy higher than 95% using state-of-the-art of speech recognition technology, and recognition accuracy drastically decreases for spontaneous speech. Broadening the application of speech recognition depends crucially on raising recognition performance for spontaneous speech. In order to increase recognition performance for spontaneous speech, several projects have been conducted. In Japan, a 5-year national project Spontaneous Speech: Corpus and Processing was conducted [50]. A world-largest Technology spontaneous speech corpus, Corpus of Spontaneous Japanese (CSJ) consisting of approximately 7 millions of words, corresponding to 700 hours of speech, was built, and various new techniques were investigated. These new techniques include flexible acoustic modeling, sentence boundary detection, pronunciation modeling, acoustic as well as language model adaptation, and automatic speech summarization [51]. The three analyses on the effects of spontaneous speech on continuous speech recognition performance are described in [93] viz., (1) spontaneous speech effects significantly degrade recognition performance, (2)

fluent spontaneous speech yields word accuracies equivalent to read speech, and (3) using spontaneous speech training data. These can significantly improve the performance for recognizing spontaneous speech. It is concluded that word accuracy can be improved by explicitly modeling spontaneous effects in the recognizer, and by using as much spontaneous speech training data as possible. Inclusion of read speech training data, even within the task domain, does not significantly improve performance. d) Robust speech recognition: To further increase the robustness of speech recognition systems, especially for spontaneous speech, utterance verification and confidence measures, are being intensively investigated [52]. In order to have intelligent or human-like interactions in dialogue applications, it is important to attach to each recognized event a number that indicates how confidently the ASR system can accept the recognized events. The confidence measure serves as a reference guide for a dialogue system to provide an appropriate response to its users. To detect semantically, significant parts and reject irrelevant portions in spontaneous utterances, a detection based approach has recently been investigated [53]. The combined recognition and verification strategy work well especially for ill-formed utterances. In order to build acoustic models more sophisticated than conventional HMMs, the dynamic Bayesian network has recently been investigated [54]. Around 2000, a QBPC[56], systems were developed to find the unknown and mismatch between training and testing conditions. A DCT fast subspace techniques[60] has been proposed to approximate the KLT for autoregressive progress. A novel implementation of a mini-max decision rule for continuous density HMM-based Robust speech recognition is developed by combining the idea of mini-max decision rule with a normal viterbi search. Speech signal modeling techniques well suited to high performance and robust isolated word recognition have been contributed[61,63]. The first robust Large vocabulary continuous speech recognition that uses syllable-level acoustic unit of LVCSR on telephone bandwidth speech is described in [64]. In 2003, a novel regression based Bayesian predictive classification(LRBPC[69]) was developed for speech Hidden markov model. Walfgang Rchichal[62] has described the methods of improving the robustness and accurancy of the acoustic modeling using decision tree based state tying. Giuluva Garau et.al.[85], investigated on Large vocabulary continuous speech recognition. Xiong Xiao[92] have shown a novel technique that normalizes the modulation spectra of speech signal. Kernel based nonlinear predictive coding[101] procedure, that yields speech features which are robust to nonstationary noise contaminated speech signal. Features maximally in sensitive to additive noise are obtained by growth transformation of regression functions that span a reproducing a kernel Hilbert space (RKHS). Soundararajan [103] proposed a supervised approach using regression trees to learn non linear transformation of the uncertainty from the

194

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

linear spectral domain to the cepstral domain. Experiments are conducted on Aurora-4 Database. e) Multimodal speech recognition: Humans use multimodal communication when they speak to each other. Studies in speech intelligibility have shown that having both visual and audio information increases the rate of successful transfer of information, especially when the message is complex or when communication takes place in a noisy environment. The use of the visual face information, particularly lip information, in speech recognition has been investigated, and results show that using both types of information gives better recognition performances than using only the audio or only the visual information, particularly, in noisy environment. Jerome R., have developed Large Vocabulary Speech Recognition with Multi-span Statistical Language Models [55] and the work done in this paper characterizes the behavior of such multi span modeling in actual recognition. A novel subspace modeling is presented in [84], including selection approach for noisy speech recognition. In subspace modeling, authors have developed a factor analysis representation of noisy speech i.e., a generalization of a signal subspace representation. They also explored the optimal subspace selection via solving the hypothesis test problems. Subspace selection via testing the correlation of residual speech, provides high recognition accuracies than that of testing the equivalent eigen-values in the minor subspace. Because of the environmental mismatch between training and test data severely deteriorates recognition performance. Jerome R. et.al.[55], have contributed large vocabulary speech recognition with multispan statistical language model. f) Modeling Techniques: Eduardo et.al.[56], introduced a set of acoustic modeling and decoding techniques for utterance verification(UV) in HMM based Continuous Speech Recognition .Lawerence K et.al.[58], discuss regarding HMM models for Automatic speech recognition which rely on high dimension feature vectors for summarizing the short time properties of speech. These have been achieved using some parameters choosen in two ways, namely i) to maximize the likelihood of observed speech signals, or ii) to minimize the number of classification errors. Dat Tat Tran[75] have proposed various models namely, i) the FE-HMM,NC-FE-HMM,FE-GMM,NC-FEGMM,FE-VQ and NC-FE-VQ in the FE approach, ii) the FCM-HMM, NC-FCM-HMM,FCM-GMM and NC-FCMGMM in the FCM approach and iii) the hard HMM and GMM as the special models of both FE and FCM approaches for speech recognition. A new statistical approach namely the probabilistic union model for Robust speech recognition involving partial, unknown frequency[67] band corruption are introduced by Ji Ming et.al. Jen Tzung et.al.[69], have surveyed a series of model selection approaches with a presentation of a novel predictive information criterion for HMM selection. Yang Liu et.al.[95], have shown that in a metadata detection scheme in speech recognition

discriminative models outperform generative than predominant HMM approaches. Alba Sloin et.al. have presented a discriminative training algorithm, that uses support vector machines(SVM) to improve the classification of discrete and continuous output probability hidden markov models. The algorithm presented in the paper[119] uses a set of maximum likelihood (ML) trained HMM models as a baseline system, and an SVM training scheme to rescore the results of the baseline HMMs. The experimental results given in that paper reduces the error rate significantly compared to standard ML training. Paper[140] presents a discriminative training algorithm that uses support vector machines(SVMs) to improve the classification of discrete and continuous output probability hidden markov models(HMMs). The algorithm uses a set of maximum likelihood (ML) trained HMM models as a baseline system, and an SVM training scheme to rescore the results of the baseline HMMs. Paper[142], proposes a Fuzzy approach to the hidden Markov model (HMM) method called the fuzzy HMM for speech and speaker recognition as an application of fuzzy expectation maximizing algorithm in HMM. This fuzzy approach can be applied to EM-style algorithms such as the Baum- Welch algorithm for hidden Markov models, the EM algorithm for Gaussian mixture models in speech and speaker recognition. Equation and how estimation of discrete and continuous HMM parameters based on this two algorithm is explained and performance of two methods of speech recognition for one hundred words is surveyed . This paper showed better results for the fuzzy HMM, compared with the conventional HMM. A novel method to estimate continuous-density hidden Markov model (CDHMM) for speech recognition [143] is, according to the principle of maximizing the minimum multi-class separation margin. The approach is named large margin HMM. First, they showed that this type of large margin HMM estimation problem can be formulated as a constrained mini-max optimization problem. Second, they propose to solve this constrained mini-max optimization problem by using a penalized gradient descent algorithm, where the original objective function, i.e., minimum margin, is approximated by a differentiable function and the constraints are cast as penalty terms in the objective function. Ultimately paper showed that the large margin training method yields significant recognition error rate reduction even on top of some popular discriminative training methods. In the work[145], techniques for recognizing phonemes automatically by using Hidden Markov Models (HMM) were proposed. The features input to the HMMs will be extracted from a single phoneme directly rather than from a string of phonemes forming a word. Also feature extraction techniques are compared to their performance in phoneme-based recognition systems. They also describe a pattern recognition approach developed for continuous speech recognition. Modeling dynamic structure of speech[146] is a novel paradigm in speech recognition research within the generative modeling framework, and it offers a potential to overcome limitations of the current hidden Markov modeling approach. Analogous to structured language models where syntactic

195

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

structure is exploited to represent long-distance relationships among words [5], the structured speech model described in this paper make use of the dynamic structure in the hidden vocal tract resonance space to characterize long-span contextual influence among phonetic units. The paper[147], discusses two novel HMM based techniques that segregate a speech segment from its concurrent background. The first method can be reliably used in clean environments while the second method, which makes use of the wavelets denoising technique, is effective in noisy environments. These methods have been implemented and they showed the superiority over other popular techniques, thus, indicating that they have the potential to achieve greater levels of accuracy in speech recognition rates. Paper[162], is motivated by large margin classifiers in machine learning. It proposed a novel method to estimate continuous-density hidden Markov model (CDHMM) for speech recognition according to the principle of maximizing the minimum multi-class separation margin. The approach is named as large margin HMM. First, it shows this type of large margin HMM estimation problem can be formulated as a constrained mini-max optimization problem. Second, it proposes to solve this constrained mini-max optimization problem by using a penalized gradient descent algorithm, where the original objective function, i.e., minimum margin, is approximated by a differentiable function and the constraints are cast as penalty terms in the objective function. The new training method is evaluated in the speakerindependent isolated E-set recognition and the TIDIGITS connected digit string recognition tasks. Experimental results clearly show that the large margin HMMs consistently outperform the conventional HMM training methods. It has been consistently observed that the large margin training method yields significant recognition error rate reduction even on top of some popular discriminative training methods. Despite their known weaknesses, hidden Markov models (HMMs) have been the dominant technique for acoustic modeling in speech recognition for over two decades. Still, the advances in the HMM framework have not solved its key problems: it discards information about time dependencies and is prone to overgeneralization. Paper[163], has attempted to overcome the above problems by relying on straightforward template matching. It showed the decrease in word error rate with 17% compared to the HMM results. In automatic speech recognition, hidden Markov models (HMMs) are commonly used for speech decoding, while switching linear dynamic models (SLDMs) can be employed for a preceding modelbased speech feature enhancement. These model types are combined[164] in order to obtain a novel iterative speech feature enhancement and recognition architecture. It is shown that speech feature enhancement with SLDMs can be improved by feeding back information from the HMM to the enhancement stage. Two different feedback structures are derived. In the first, the posteriors of the HMM states are used to control the model probabilities of the SLDMs, while in the second they are employed to directly influence the estimate of the speech feature distribution. Both approaches lead to improvements in recognition accuracy both on the AURORA-

2 and AURORA-4 databases compared to non-iterative speech feature enhancement with SLDMs. It is also shown that a combination with uncertainty decoding further enhances performance. g) Noisy speech recognition: In 2008[89], a new approach for speech feature enhancement in the log spectral domain for noisy speech recognition is presented. A switching linear dynamic model (SLDM) is explored as a parametric model for the clean speech distribution. The results showed that the new SLDM approach can further improve the speech feature enhancement performance in terms of noise robust recognition accuracy. Jen-Tzung et.al.[69], present a novel subspace modeling and selection approach for noisy speech recognition. Jianping Dingebal [89] have presented a new approach for speech feature enhancement in the large spectral domain for NSR. Xiaodong[108] propose a novel approach which extends the conventional GMHMM by modeling state emission(mean and variance) as a polynomial function of a continuous environment dependent variable. This is used to improve the recognition performance in noisy environments by using multi condition training. Switching Linear dynamical system(SLDC)[102], is a new model that combines both the raw speech signal and the noise was introduced in the year 2008. This was tested using isolated digit utterance corrupted by Gaussian noise. Contrary to Autoregressive HMMs,SLDC s outperforms a state of the art feature based HMM. Mark D.Skowronski[104], proposed echo state network classifier by combining ESN with state machine frame work for noisy speech recognition. In the paper[144], authors propose a novel approach which extends the conventional Gaussian mixture hidden Markov model (GMHMM) by modeling state emission parameters (mean and variance) as a polynomial function of a continuous environment-dependent variable. At the recognition time, a set of HMMs specific to the given value of the environment variable is instantiated and used for recognition. The maximum-likelihood (ML) estimation of the polynomial functions of the proposed variable-parameter GMHMM is given within the expectation-maximization (EM) framework. h) Data driven approach: A new approach [134], for deriving compound words from a training corpus was proposed. The motivation for making compound words is because under some assumptions, speech recognition errors occur less frequently in longer words were discussed along with the accurate modeling . They have also introduced a measure based on the product between the direct and the reverse bi-gram probability of a pair of words for finding candidate pairs in order to create compound words. Paper[135] surveys a series of model selection approaches and presents a novel predictive information criterion (PIC) for hidden Markov model (HMM) selection. The approximate Bayesian using Viterbi approach is applied for PIC selection of the best HMMs providing the largest prediction information for generalization of future data. Authors have developed a

196

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

top-down prior/posterior propagation algorithm for estimation of structural hyper-parameters and they have showed the evaluation of continuous speech recognition(data driven) using decision tree HMMs, the PIC criterion outperforms ML and MDL criteria in building a compact tree structure with moderate tree size and higher recognition rate. A method of compensating for nonlinear distortions in speech representation caused by noise was proposed which is based on the histogram equalization. Paper[138], introduces the data driven signal decomposition method based on the empirical mode decomposition(EMD) technique. The decomposition process uses the data themselves to derive the base function in order to decompose the one-dimensional signal into a finite set of intrinsic mode signals. The novelty of EMD is that the decomposition does not use any artificial data windowing which implies fewer artifacts in the decomposed signals. The results show that the method can be effectively used in analyzing non-stationary signals. 7. Speech Databases: Speech databases have a wider use in Automatic Speech Recognition. They are also used in other important applications like, Automatic speech synthesis, coding and analysis including speaker and language identification and verification. All these applications require large amounts of recoded database. Different types of databases that are used for speech recognition applications are discussed along with its taxonomy.
Taxonomy of Existing Speech Databases: The intra-speaker and inter-speaker variability are important parameters for a speech database. Intra-speaker variability is very important for speaker recognition performance. The intraspeaker variation can originate from a variable speaking rate, changing emotions or other mental variables, and in environment noise. The variance brought by different speakers is denoted inter-speaker variance and is caused by the individual variability in vocal systems involving source excitation, vocal tract articulation, lips and/or nostril radiation. If the inter-speaker variability dominates the intra-speaker variability, speaker recognition is feasible. Speech databases are most commonly classified into single-session and multisession. Multi-session databases allow estimation of temporal intra-speaker variability. According to the acoustic environment, databases are recorded either in noise free environment, such as in the sound booth, or with office/home noise. Moreover, according to the purpose of the databases, some corpora are designed for developing and evaluating speech recognition, for instance TIMIT, and some are specially designed for speaker recognition, such as SIVA, Polycost and YOHO. Many databases were recorded in one native language of recording subjects; however there are also multi-language databases with non-native language of speakers, in which case, the language and speech recognition become the additional use of those databases. Main database characteristics:

Table-6 represents the characteristics of main databases used in speech recognition.

Abbreviations: QR: Quiet Room Ofc: Office RF:Radio Frequency
Table 6: Database Characteristics:

7.1. Resource Management Complete Set 2.0: The DARPA Resource Management Continuous Speech Corpora (RM) consists of digitized and transcribed speech for use in designing and evaluating continuous speech recognition systems. There are two main sections, often referred to as RM1 and RM2. RM1 contains three sections, SpeakerDependent (SD) training data, Speaker-Independent (SI) training data and test and evaluation data. RM2 has an additional and larger SD data set, including test material. All RM material consists of read sentences modeled after a naval resource management task. The complete corpus contains over 25,000 utterances from more than 160 speakers representing a variety of American dialects. The material was recorded at 16KHz, with 16-bit resolution, using a Sennheiser HMD-414 headset microphone. All discs conform to the ISO-9660 data format. 7.1.1. Resource Management SD and SI Training and Test Data (RM1): The Speaker-Dependent (SD) Training Data contains 12 subjects, each reading a set of 600 "training sentences," two "dialect" sentences and ten "rapid adaptation" sentences, for a total of 7,344 recorded sentence utterances. The 600 sentences designated as training cover 97 of the lexical items in the corpus. The Speaker-Independent (SI) Training Data contains 80 speakers, each reading two "dialect" sentences plus 40 sentences from the Resource Management text corpus, for a total of 3,360 recorded sentence utterances. Any given sentence from a set of 1,600 Resource Management sentence

197

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

texts were recorded by two subjects, while no sentence was read twice by the same subject. RM1 contains all SD and SI system test material used in 5 DARPA benchmark tests conducted in March and October of 1987, June 1988 and February and October 1989, along with scoring and diagnostic software and documentation for those tests. Documentation is also provided outlining the use of the Resource Management training and test material at CMU in development of the SPHINX system. Example output and scored results for state-of-the-art speaker-dependent and speaker-independent systems (i.e. the BBN BYBLOS and CMU SPHINX systems) for the October 1989 benchmark tests are included. 7.1.2.Extended Resource Management Speaker-Dependent Corpus (RM2): This set forms a speaker-dependent extension to the Resource Management (RM1) corpus. The corpus consists of a total of 10,508 sentence utterances (two male and two female speakers each speaking 2,652 sentence texts). These include the 600 "standard" Resource Management speaker-dependent training sentences, two dialect calibration sentences, ten rapid adaptation sentences, 1,800 newly-generated extended training sentences, 120 newly-generated development-test sentences and 120 newly-generated evaluation-test sentences. The evaluation-test material on this disc was used as the test set for the June 1990 DARPA SLS Resource Management Benchmark Tests (see the Proceedings). The RM2 corpus was recorded at Texas Instruments. The NIST speech recognition scoring software originally distributed on the RM1 "Test" Disc was adapted for RM2 sentences. 7.2.TIMIT: TIMIT is a corpus of phonemically and lexically transcribed speech of American English speakers of different sexes and dialects. Each transcribed element has been delineated in time. TIMIT was designed to further acoustic-phonetic knowledge and automatic speech recognition systems. It was commissioned by DARPA and worked on by many sites, including Texa Instrument (TI) and Masachusetts Institute of Technology (MIT), hence the corpus' name. There is also a telephone bandwidth version called NTIMIT (Network TIMIT). The TIMIT corpus of read speech is designed to provide speech data for acoustic-phonetic studies and for the development and evaluation of automatic speech recognition systems. Although it was primarily designed for speech recognition, it is also widely used in speaker recognition studies, since it is one of the few databases with a relatively large number of speakers. It is a single-session database recorded in a sound booth with fixed wideband headset. TIMIT contains broadband recordings of 630 speakers of eight major dialects of American English, each reading ten phonetically rich sentences. The TIMIT corpus includes timealigned orthographic, phonetic and word transcriptions as well as a 16-bit, 16kHz speech waveform file for each utterance.

Corpus design was a joint effort among the Massachusetts Institute of Technology (MIT), SRI International (SRI) and Texas Instruments, Inc. (TI). There are numerous corpora for speech recognition. The most popular bases are: TIMIT and its derivatives, Polycost, and YOHO.
7.2.1. TIMIT and Derivatives: The derivatives of TIMIT are: CTIMIT, FFMTIMIT, HTIMIT, NTIMIT, VidTIMIT. They were recorded by playing different recording input devices, such as telephone handset lines and cellular telephone handset, etc. TIMIT and most of the derivatives are single-session, and are thus not optimal for evaluating speaker recognition systems because of lack of intra-speaker variability. VidTIMIT is an exception, being comprised of video and corresponding audio recordings of 43 subjects. It was recorded into three sessions with around one week delay between each session. It can be useful for research involving automatic visual or audio-visual speech recognition or speaker verification. 7.3 TI46 database: The TI46 corpus was designed and collected at Texas Instruments(TI). The speech was produced by 16 speakers, 8 females and 8 males, labeled f1-f8 and m1-m8 respectively, consisting of two vocabularies TI-20 and TI-alphabet. The TI-20 vocabulary contains the ten digits from 0 to 9 and ten command words: enter, erase, go, help, no, robot, stop, start, and yes. The TI alphabet vocabulary contains the names of the 26 letters of the alphabet from a to z. For each vocabulary item each speaker produced 10 tokens in a single training session and another two tokens in each of 8 testing sessions. 7.4 SWITCHBOARD: SWITCHBOARD is a large multi-speaker corpus of telephone conversations. Although designed to support several types of speech and language research, its variety of speakers, speech data, telephone handsets, and recording conditions make SWITCHBOARD a rich source for speaker verification experiments of several kinds. Collected at Texas Instruments with funding from ARPA, SWITCHBOARD includes about 2430 conversations averaging 6 minutes in length; in other terms, over 240 hours of recorded and transcribed speech, about 3 million words, spoken by over 500 speakers of both sexes from every major dialect of American English. The data is 8 kHz, 8-bit mu-law encoded, with the two channels interleaved in each audio. In addition to its volume, SWITCHBOARD has a number of unique features contributing to its value for telephone-based speaker identification technology development. SWITCHBOARD was collected without human intervention, under computer control. From human factors perspective, automation guards against the intrusion of experimenter bias, and guarantees a degree of uniformity throughout the long period of data collection. The protocols were further intended to elicit natural and spontaneous speech by the participants.

198

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Each transcript is accompanied by a time alignment i.e, which estimates the beginning time and duration of each word in the transcript in centi-seconds. The time alignment was accomplished with supervised phone-based speech recognition, as described by Wheatley et al. The corpus is therefore capable of supporting not only purely textindependent approaches to speaker verification, but also those which make use of any degree of knowledge of the text, including phonetics. SWITCHBOARD has both depth and breadth of coverage for studying speaker characteristics. Forty eight people participated 20 times or more; this yields at least an hour of speech, enough for extensive training or modeling and for repeated testing with unseen material. Hundreds of others participated ten times or less, providing a pool large enough for many open-set experiments. The participants demographics, as well as the dates, times, and other pertinent information about each phone call, are recorded in relational database tables. Except for personal information about the callers, these tables are included with the corpus. The volunteers who participated provided information relevant to studies of voice, dialect, and other aspects of speech style, including age, sex, education, current residence, and places of residence during formative years. The exact time and the area code of origin of each call is provided, as well as a means of telling which calls by the same person came from different telephones.

4) 5)

Distance -based methods Maximum likelihood approach Isolated word recognition Small vocabulary Context Independent units Clean speech recognition Single speaker recognition Monologue recognition Read speech recognition Single modality(audio signal only) Hardware recognizer Speech signal is assumed as quasi stationary in the traditional approaches. The feature vectors are extracted using FFT and wavelet methods etc.,.

Likelihood based methods Discriminative approach MCE/GPD and MMI e.g.

6) 7) 8)

Continuous speech recognition, Large vocabulary Context dependent units

9) 10) 11) 12) 13)

Noisy/telephone recognition

speech

Speaker-independent/adaptive recognition Dialogue/Conversation recognition Spontaneous speech recognition Multimodal(audio/visual)speech recognition Software recognizer Data driven approach does not posses this assumption i.e. signal is treated as nonlinear and nonstationary. In this features are extracted using Hilbert Haung Transform using IMFs.[141]

7.5. Air Travel Information System(ATIS): The ATIS database is commonly used for the evaluation of word error performances of the Automatic Speech Recognition. ATIS is based on a realistic application environment and is a good simulation of spontaneous conversation. 8. Summary of the technology progress: In the last 60 years, especially in the last three decades, research in speech recognition has been intensively carried out world wide, spurred on by advances in signal processing algorithms, architectures and hardware. The technological progress in the 60 years can be summarized in the table 7[137]. Table 7: Summary of the technological progress in the last 60 years Sl.No. 1) Past Template matching Filter bank/spectral resonance Heuristic time normalization Present(new) Corpus-based statistical modeling, e.g. HMM and n grams Cepstral features, Kernel based function, group delay functions DTW/DP matching

14) 15)

2)

9. Gap between machine and human speech recognition: What we know about human speech processing is still very limited, and we have yet to witness a complete and worthwhile unification of the science and technology of speech. In 1994,

3)

199

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Moore [96] presented the following 20 themes which is believed to be an important to the greater understanding of the nature of speech and mechanisms of speech pattern processing in general: How important is the communicative nature of speech? Is human-human speech communication relevant to human machine communication by speech? Speech technology or speech science?(How can we integrate speech science and technology). Whither a unified theory? Is speech special? Why is speech contrastive? Is there random variability in speech? How important is individuality? Is disfluency normal? How much effort does speech need? What is a good architecture (for speech processes)? What are suitable levels of representation? What are the units? What is the formalism? How important are the physiological mechanisms? Is time-frame based speech analysis sufficient? How important is adaptation? What are the mechanisms for learning? What is speech good for? How good is speech. After more than 10 years, we still do not have clear answers to these 20 questions. 10. Discussions and Conclusions: Speech is the primary, and the most convenient means of communication between people. Whether due to technological curiosity to build machines that mimic humans or desire to automate work with machines, research in speech and speaker recognition, as a first step toward natural human-machine communication, has attracted much enthusiasm over the past five decades. we have also encountered a number of practical limitations which hinder a widespread deployment of application and services. In most speech recognition tasks, human subjects produce one to two orders of magnitude less errors than machines. There is now increasing interest in finding ways to bridge such a performance gap. What we know about human speech processing is very limited. Although these areas of investigations are important the significant advances will come from studies in acousticphonetics, speech perception, linguistics, and psychoacoustics. Future systems need to have an efficient way of representing, storing, and retrieving knowledge required for natural conversation. This paper attempts to provide a comprehensive survey of research on speech recognition and to provide some year wise progress to this date. Although significant progress has been made in the last two decades, there is still work to be

done, and we believe that a robust speech recognition system should be effective under full variation in: environmental conditions, speaker variability s etc. Speech Recognition is a challenging and interesting problem in and of itself. We have attempted in this paper to provide a comprehensive cursory, look and review of how much speech recognition technology progressed in the last 60 years. Speech recognition is one of the most integrating areas of machine intelligence, since, humans do a daily activity of speech recognition. Speech recognition has attracted scientists as an important discipline and has created a technological impact on society and is expected to flourish further in this area of human machine interaction. We hope this paper brings about understanding and inspiration amongst the research communities of ASR. ACKNOWLEDGMENTS The authors remain thankful to Dr.G.Krishna (Rtd. Prof. IISc, Bangalore)and Dr.M.Narashima Murthuy, Prof. & Chairman, Dept. of CS & Automation, IISc., Bangalore, for their useful discussions and suggestions during the preparation of this technical paper.

REFERENCES [1]. Sadaoki Furui, 50 years of Progress in speech and Speaker Recognition Research , ECTI Transactions on Computer and Information Technology,Vol.1. No.2 November 2005. [2].K.H.Davis, R.Biddulph, and S.Balashek, Automatic Recognition of spoken Digits, J.Acoust.Soc.Am., 24(6):637-642,1952. [3]. H.F.Olson and H.Belar, Phonetic Typewriter , J.Acoust.Soc.Am.,28(6):1072-1081,1956. [4]. D.B.Fry, Theoritical Aspects of Mechanical speech Recognition , and P.Denes, The design and Operation of the Mechanical Speech Recognizer at Universtiy College London, J.British Inst. Radio Engr., 19:4,211-299,1959. [5]. J.W.Forgie and C.D.Forgie, Results obtained from a vowel recognition computer program , J.A.S.A., 31(11),pp.1480-1489.1959. [6]. J.Suzuki and K.Nakata, Recognition of Japanese Vowels Preliminary to the Recognition of Speech , J.Radio Res.Lab37(8):193-212,1961. [7]. T.Sakai and S.Doshita, The phonetic typewriter, information processing 1962 , Proc.IFIP Congress, 1962. [8]. K.Nagata, Y.Kato, and S.Chiba, Spoken Digit Recognizer for Japanese Language , NEC Res.Develop., No.6,1963. [9].T.B.Martin, A.L.Nelson, and H.J.Zadell, Speech Recognition b Feature Abstraction Techniques , Tech.Report AL-TDR-64-176,Air Force Avionics Lab,1964. [10].T.K.Vintsyuk, Speech Discrimination by Dynamic Programming , Kibernetika, 4(2):81-88,Jan.-Feb.1968. [11]. H.Sakoe and S.Chiba, Dynamic programming algorithm

200

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

optimization for spoken word recognition ,IEEE Trans. Acoustics, Speech, Signal Proc., ASSP-26(1).pp.4349,1978. [12].D.R.Reddy, An Approach to Computer Speech Recognition by Direct Analysis of the Speech Wave , Tech.Report No.C549, Computer Science Dept., Stanford Univ., September 1966. [13].V.M.Velichko and N.G.Zagoruyko, Automatic Recognition of 200 words , Int.J.Man-Machine Studies,2:223,June 1970. [14].H.Sakoe and S.Chiba, Dynamic Programming Algorithm Optimization for Spoken Word Recognition ,IEEE Trans.Acoustics, Speech, Signal Proc.,ASSP-26(1):4349,February 1978. [15].F.Itakura, Minimum Prediction Residula Applied to Speech Recognition ,IEEE Trans.Acoustics, Speech,Signal Proc., ASSP-23(1):67-72,February 1975. [16].C.C.Tappert,N.R.Dixon, A.S.Rabinowitz, and W.D.Chapman, Automatic Recognition of Continuous Speech Utilizing Dynamic Segmentation, Dual Classification, Sequential Decoding and Error Recover , Rome Air Dev.Cen, Rome, NY,Tech.Report TR-71146,1971. [17].F.Jelinek, L.R.Bahl, and R.L.Mercer, Design of a Lingusistic Statistical Decoder for the Recognition of Continuous Speech , IEEE Trans.Information Theory,IT21:250-256,1975. [18].F.Jelinek, The Development of an Experimental Discrete Dictation Recognizer , Proc.IEEE,73(11):1616624,1985. [19].L.R.Rabiner, S.E.Levinson, A.E.Rosenberg, and J.G.Wilpon, Speaker Independent Recognition of Isolated Words Using Clustering Techniques , IEEE Trans. Acoustics, Speech, Signal Proc., ASSP-27:336349,August 1979. [20].D.Klatt, Review of the ARPA Speech understanding project , J.A.S.A. 62(6),pp.1324-1366,1977. [21].B.Lowrre, The HARPY speech understanding system , Trends in Speech Recognition, W.Lea,Ed., Speech Science Pub., pp.576-586,1990. [22].H.Sakoe, Two Level DP Matching A Dynamic Programming Based Pattern Matching Algorithm for Connected Word Recognition , IEEE Trans. Acoustics, Speech, Signal Proc., ASSP-27:588-595, December 1979. [23].J.S.Bridle and M.D.Brown, Connected Word Recognition Using whole word templates , Proc. Inst.Acoust.Autumn Conf.,25-28,November 1979. [24]. C.S.Myers and L.R.Rabiner, A Level Building Dynaimc Time Warping Algorithm for Connected Word Recognition , IEEE Trans. Acoustics, Speech Signal Proc., ASSP-29:284-297,April 1981. [25].C.H.Lee and L.R.Rabiner, A Frame Synchronous Network Search Algorithm for Connected Word Recognition , IEEE Trans. Acoustics, Speech, Signal Proc., 37(11):1649-1658, November 1989. [26].J.Ferguson, Ed., Hidden Markov Models for Speech,

IDA,Princeton, NJ,1980. [27].L.R.Rabiner, A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition , Proc.IEEE,77(2):257-286,February 1989. [28].J.Ferguson, Ed., Hidden Markov models for speech, IDA, Princeton, NJ,1980. [29].L.R.Rabiner, A tutorial on hidden Markov models and selected applications in speech recognition , Proc.IEEE,77(2),pp.257-286,1989. [30].L.R.Rabiner and B.H.Juang, Fundamentals of Speech Recognition, Prentice-Hall, Englewood Cliff, New Jersy, 1993. [31].S.Furui, Speaker-independent isolated word recognition using dynamic features of speech spectrum , IEEE Trans. Acoustics, Speech, Signal Processing, ASSP34,pp.52-59,1986. [32].F.Jelinek, The development of an experimental discrete dictation recognizer , Proc. IEEE,73(11),pp.16161624,1985. [33].S.Katagiri, Speech pattern recognition using neural networks , W.Chou and B.H.Juang (Eds.) Pattern Recognition in Speech and Language Processing, CRC Press, pp.115-147,2003. [34].R.P.Lippmann, An introduction to computing with neural nets , IEEE ASSP Mag., 4(2),pp.4-22,April 1987. [35].A.Weibel, et.al., Phoneme recognition using time-delay neural networks , IEEE Trans. Acoustics, Speech, Signal Proc.,37,pp.393-404,1989. [36].K.F.Lee, et.al, An overview of the SPHINX speech recognition system , Proc.ICASSP, 38,pp.600-610,1990. [37].Y.L.Chow, et.al. BYBLOS, the BBN continuous speech recognition system , Proc.ICASSP, pp.89-92,1987. [38].M.Weintraub et al., Linguistic constraints in hidden Markov model based speech recognition , Proc.ICASSP, pp.699-702,1989. [39].D.B.Paul, The Lincoln robust continuous speech recognizer , Proc.ICASSP,449-452,1989. [40].V.Zue, et.al., The MIT summit speech recognition system, a progress report , Proc.DARPA Speech and Natural Language Workshop, pp.179-189,1989. [41].C.H.Lee, etc.al., Acoustic modeling for large vocabulary speech rcognition ,Computer Speech and Language, 4,pp.127-165,1990. [42].B.H.Juang and S.Furui, Automatic speech recognition and understanding: A first step toward natural human machine communication , Proc.IEEE,88,8,pp.11421165,2000. [43].K.P.Li and G.W.Hughes, Talker differences as they appear in correlation matrices of continuous speech spectra , J.A.S.A., 55,pp.833-837,1974. [44].C. J. Leggetter and P. C. Woodland, Maximum likelihood linear regression for speaker pp.adaptation of continuous density hidden Markov models, Computer Speech and Language, 9, 171-185, 1995. [45].A. P. Varga and R. K. Moore, Hidden Markov model decomposition of speech and noise, Proc. ICASSP, pp. 845-848, 1990.

201

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

[46].M. J. F. Gales and S. J. Young, Parallel model combination for speech recognition in noise, Technical Report, CUED/FINFENG/ TR135, 1993. [47].K. Shinoda and C. H. Lee, A structural Bayes approach to speaker adaptation, IEEE Trans. Speech and Audio Proc., 9, 3, pp. 276-287, 2001. [48].Y. Liu, et. al., Structural metadata research in the EARS program, Proc. ICASSP, V-957, 2005. [49].H. Soltau, et. al., The IBM 2004 conversational transcription, Proc. ICASSP, I-205-208, 2005. [50].S. Furui, Recent progress in corpus-based spontaneous speech recognition, IEICE Trans. Inf. & Syst., E88-D, 3, pp. 366-375, 2005. [51].S. Furui, Speech-to-text and speech-to-speech summarization of spontaneous speech, IEEE Trans. Speech & Audio Processing, 12, 4, pp. 401-408, 2004. [52]. C. H. Lee, Statistical confidence measures and their applications, Proc. ICSP, pp.1021-1028,2001. [53]. T. Kawahara, et al., Key-phrase verification for flexible speech understanding , IEEE Trans.Speech & Audio Proc., 558-568, 1998. [54]. C. S. Myers and L. R. Rabiner, A level building for connected word recognition, IEEE Trans. Acoustics, Speech, Signal Proc., ASSP-29, pp.284- 297, 1981. [55]. Jerome R. Bellegarda, Large Vocabulary Speech Recognition with Multispan statistical Language Models , 76 IEEE Transactions On Speech & Audio Processing, Vol. 8, No. 1, January 2000. [56].Eduardo Lleida Et.al. Utterance Verification In Decoding And Training Procedures , IEEE Transactions On Speech And Audio Processing, Vol. 8, No. 2, March 2000. [57]. Qiang Huo and Chin-Hui Lee, A Bayesian Predictive Classification Approach to Robust Speech Recognition , IEEE Transactions On Speech & Audio Processing, Vol. 8, No. 2, March 2000. [58]. Lawrence K. Saul and Mazin G. Rahim, Maximum Likelihood and Minimum Classification Error Factor Analysis for Automatic Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 8, No. 2, March 2000. [59]. Linkai Bu et.al, Perceptual Speech Processing and Phonetic Feature Mapping for Robust Vowel Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 8, No. 2, March 2000. [61]. Jun Huang and Yunxin Zhao, A DCT-Based Fast Signal Subspace Technique for Robust Speech Recognition ,IEEE Transactions On Speech And Audio Processing, Vol. 8, No. 6, November 2000. [62]. Hui Jiang et.al., A Minimax Search Algorithm for Robust Continuous Speech Recognition ,IEEE Transactions On Speech And Audio Processing, Vol. 8, No. 6, November 2000. [63]. Wolfgang Reichl and Wu Chou, Robust Decision Tree State Tying for Continuous Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 8, No. 5, September 2000.

[64]. Montri Karnjanadecha and Stephen A. Zahorian, Signal Modeling for High-Performance Robust Isolated Word Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 9, No. 6, September 2001. [65]. Aravind Ganapathiraju et.al., Syllable-Based Large Vocabulary Continuous Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 9, No. 4, May 2001. [66]. George Saon and Mukund Padmanabhan, Data-Driven Approach to Designing Compound Words for Continuous Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 9, No. 4, May 2001. [67]. Alain Biem et.al, An Application of Discriminative Feature Extraction to Filter-Bank-Based Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 9, No. 2, February 2001. [68]. Ji Ming et.al., Robust Speech Recognition Using Probabilistic Union Models , IEEE Transactions On Speech And Audio Processing, Vol. 10, No. 6, September 2002. [69]. Rita Singh, Bhiksha Raj et. al., Automatic Generation of Subword Units for Speech Recognition Systems , IEEE Transactions On Speech And Audio Processing, Vol. 10, No. 2, February 2002. [70]. Jen-Tzung Chien, Linear Regression Based Bayesian Predictive Classification for Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 11, No. 1, January 2003. [71]. Mohamed Afify and Olivier Siohan, Sequential Estimation With Optimal Forgetting for Robust Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 12, No. 1, January 2004. [72]. Shinji Watanabe, Variational Bayesian Estimation and Clustering for Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 12, No. 4, July 2004. [73]. Mohamed Afify, Feng Liu, Hui Jiang, A New Verification-Based Fast-Match for Large Vocabulary Continuous Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 13, No. 4, July 2005. [74]. Giuseppe Riccardi,, Active Learning: Theory and Applications to Automatic Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 13, No. 4, July 2005. [75]. Frank Wessel and Hermann Ney , Unsupervised Training of Acoustic Models for Large Vocabulary Continuous Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 13, No. 1, January 2005 . [76]. Dat Tat Tran, Fuzzy Approaches to Speech and Speaker Recognition , A thesis submitted for the degree of Doctor of Philosophy of the university of Canberra. [77]. Lawrence Rabiner, Biing Hwang Juang, Fundamental of Speech Recognition , Copyright 1993 by AT&T. [78]. Dennis Norris, Merging information in speech recognition: Feedback is never necessary , Behavioral

202

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

and brain Sciences,2000. [79]. Simon Kinga and Joe Frankel, Recognition ,Speech production knowledge in automatic speech recognition , Journal of Acoustic Society of America,2006. [80]. Masanobu Nakamura, Koji Iwano, and Sadaoki Furui, The Effect Of Spectral Space Reduction In Spontaneous Speech On Recognition Performances ,2007. [81]. Jen-Tzung Chien et.al., Predictive Hidden Markov Model Selection for Speech Recognition ,IEEE Transactions On Speech And Audio Processing, Vol. 13, No. 3, May 2005. [82]. Sadaoki Furui, Tomohisa Ichiba et.al., Cluster-based Modeling for Ubiquitous Speech Recognition , Department of Computer Science Tokyo Institute of Technology, Interspeech 2005. [83]. Michael K.Brown et.al, Training set Design for Connected Speech Recognition , IEEE Transaction on Signal Processing, Vol.39.No.6.June 1991. [84]. Shinta Kimura, Advances in Speech Recognition Technologies , FUJITSU, Sci.Tech. J.,35,2,pp,202211,December 1999. [85]. Jen-TZung et.al, Factor Analysed Subspace modeling And Selection IEEE transactions on Audio, speech and Language Processing,,Vol.16,No.1 Jan 2008. [86]. Giulua Garau & Stere rebaks, Combining spectral representation for large vocabulary continuous speech recognition , IEEE Transactions On Audio, Speech And Language Processing, Vol.16,No.1 Jan 2008. [87]. Jeih-weih, Constructing modulation frequency domain based Features for Robust speech recognition , Vol.16,No.1 Jan 2008. [88]. Jeremy Morris, Conditional Random fields for Integrating local discriminative Classifiers , IEEE Transactions on Audio, speech and Language Processing,Vol.16,No.1 Jan 2008. [89].Chia-Yu et.al., Histogram based quantization for Roboust and/or Distributed speech recognition , IEEE Transactions On Audio, Speech And Language Processing, Vol.16,No.1 Jan 2008. [90]. Jianping Dingebal, Feature enhancement for noisy speech recognition with a time-variant linear predictive HMM structure , IEEE transactions on Audio,Speech and Language Processing, Vol.16,No.1 Jan 2008. [91]. Shahramkhadivi and Hermann Ney, Integrating of Speech recognition and machine translation in computer Assisted Translation , IEEE Transactions On Audio,Speech And Language Processing, Vol.16,No.1 Jan 2008. [92]. Kevin M.Indrebo et.al, Minimum mean squared error estimation of mel-frequency cepstral co-efficients using a Novel Distortion model , IEEE Transactions On Audio,Speech And Language Processing, Vol.16,No.1 Jan 2008. [93]. Xiong Xiao, Normalisaton of the speech modulation spectra for robust speech recognition , IEEE transactions on Audio, Speech and Language Processing, Vol.16,No.1 Jan 2008.

[94]. John Butzberger , Spontaneous Speech Effects In Large Vocabulary Speech Recognition Applications , SRI International Speech Research and Technology Program Menlo Park, CA 94025. [95]. Satya Dharanipragada et.al, Gaussian mixture models with covariance s or Precisions in shared multiple subspaces, , IEEE Transactions on Audio, Speech and Language Processing, V.14,No.4,July 2006. [96]. Yang Liu et.al,. Enriching Speech Recognition with Automatic Detection of sentence Boundaries an disfluencies , IEEE Transactions on Audio,Speech and Language processing, V.14,No.4,July 2006. [97]. R.K.Moore, Twenty things we still don t know about speech , Proc.CRIM/ FORWISS Workshop on Progress and Prospects of speech Research an Technology , 1994. [98]. Alex Weibel and Kai-Fu Lee, Reading in Speech Recognition ,Morgan Kaufmann Publishers, Inc. San Mateo, California,1990. [99]. Adoram Erell et.al., Energy conditioned spectral estimation for Recognition of noisy speed , IEEE Transactions on Audio, Speech and Language processing, Vol.1,No.1, Jan 1993. [100].Adoram Erell et.al, Filter bank energy estimation Using mixture and morkov models for Recognition of Noisy speech , IEEE Transactions on Audio, Speech And Language processing, Vol.1,No.1 Jan 1993. [101].Mathias De-Wachter et.al., Template based continuous speech recognition ,IEEE transactions on Audio, speech and Language processing, Vol.15,No.4, May 2007. [102].Shantanu Chakrabarthy et.al., Robust speech feature extraction by Growth transformation in Reproducing Kernel Hilbert space , IEEE Transactions on Audio, Speech and Language processing, Vol.15,No.6, June 2007. [103].Bertrand Mesot, Switching Linear Dynamical Systems for Noise Robust speech recognition , IEEE transactions on Audio, speech and Language processing, Vol.15,No.6, May 2007. [104].Soundararajan et.al., Transforming Binary unCertainties for Robust Speech Recognition , IEEE Transactions on Audio, Speech and Language processing, Vol.15,No.6, July 2007. [105].Mark D.Skowronski et.al., Noise Robust Automatic Speech Recognition using a predictive Echo state Network , IEEE transactions on Audio, Speech and Language processing, Vol.15,No.5, June 2007. [106].Xinwei Li et.al., Solving large HMM Estimation via Semi-definite programming , IEEE Transactions on Audio, speech and Language processing, Vol.15,No.8, December 2007. [107].Rajesh M Hegde et.al., SWignificance of the Modified Group Delay feature , IEEE, Transactions On Audio, Speech And Language Processing, Vol.15,No.1, Jan 2007. [108].Scott Axelrod et.al., Discriminative Estimation of subspace constrained Gaussian mixture models for Speech recognition , IEEE Transactions on Audio,

203

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Speech and Language processing, Vol.15,No.1, Jan 2007. [108].Xiaodong Lui, et.al., A study of variable parameter Gaussian mixture HMM modeling fro Noisy speech recognition , IEEE Transactions on Audio, Speech and Language processing, Vol.15,No.1, Jan. 2007. [109].Keh-Yih Su et.al., Speech Recognition using weighted HMM and subspace Projection Approaches , IEEE Transactions on Audio, Speech and Language. [110].L.R.Bahl et.al, A method for the construction of Acoustic Markov Models for Words , IEEE Transactions on Audio, Speech and Language processing, Vol.1,No.4, Oct.1993. [111].Shigeru Katagiri et.al., A New hybrid algorithm for speech recognition based on HMM segmentation and learning Vector quantization , IEEE Transactions on Audio, Speech and Language processing Vol.1,No.4, [112].Adoram Erell et.al., Filter bank energy estimation using mixture and Markov models for Recognition of Nosiy Speech , IEEE Transactions on Audio, Speech and Language processing Vol.1,No.1, Jan.1993. [113].Lalit R.Bahl et.al, Estimating Hidden Markov Model Parameters So as to maximize speech recognition Accuracy , IEEE Transactions on Audio, Speech and Language processing Vol.1,No.1, Jan.1993. [114].Gerhard Rigoll, Maximum Mutual Information Neural Networks for Hybrid connectionist-HMM speech Recognition Systems , IEEE Transactions on Audio, Speech and Language processing Vol.2,No.1, Part II,Jan.1994. [115].Antonio M.Peinado et.al., Discriminative Codebook Design Using Multiple Vector quantization in HMM based speech recognizers , IEEE Transactions on Audio, Speech and Language processing Vol.4,No.2, March.1996. [116].Mazin G.Rahim et.al., Signal Bias Removal by maximum Likelihood Estimation for Robust Telephone Speech Recognition ,IEEE Transactions on Audio, Speech and Language processing Vol.4,No.1, Jan.1996. [117].Qiang Huo et.al, Bayesian Adaptive Learning of the parameters of Hidden Markov model for speech recognition ,IEEE Transactions on Audio, Speech and Language processing Vol.3,No.5, Sept..1995. [118].Nam Soo Kim et.al., On estimating Robust probability Distribution in HMM in HMM based speech recognition , IEEE Transactions on Audio, Speech and Language processing Vol.3,No.4, July1995. [119].Alba Sloin et.al., Support Vecotr Machine Training For improved Hidden markov modeling ,IEEE Transactions on Audio, Speech and Language processing,Vol.56,No.1, Jan. 2008. [120].Jean Francois, Automatic Word Recognition Based on Second Order Hidden Markov Models , IEEE Transactions on Audio, Speech and Language processing Vol.5,No.1, Jan.1997. [121].Ananth Sankar, A maximum likelihood approach to stochastic matching for Robust speech recognition ,IEEE Transactions on Audio, Speech and Language processing

Vol.4,No.3, May.1996. [122].David Burshtein, Robust parametric modeling of durations in Hidden Markov models , IEEE Transactions on Audio, Speech and Language processing,Vol.4,No.3, May.1996. [123].Mari Ostendorf et.al., From HMM s to segment Models: A unified view of stochastic Modeling for Speech Recognition , IEEE Transactions on Audio, Speech and Language processing Vol.4,No.5, Septemeber1996. [124].Yaxin Zhang et.al., Phoneme-Based Vector Quantization in a Discrete HMM Speech Recognizer , IEEE Transactions On Speech And Audio Processing, Vol. 5, No. 1,January 1997. [125].Yifan Gong, Stochastic Trajectory Modeling and Sentence Searching for continuous Speech Recognition ,Ieee Transactions On Speech And Audio Processing, Vol. 5,No. 1, January 1997. [126].Javier Hernando and Climent Nadeu, Linear Prediction of the One-Sided Autocorrelation Sequence for Noisy Speech Recognition ,IEEE Transactions On Speech And Audio Processing, Vol. 5, No. 1, January 1997. [127].Rathinavelu Chengalvarayan , Use of Generalized Dynamic Feature Parameters for Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 5,No. 3, May 1997. [128]. Qiang Huo, On-Line Adaptive Learning of the Correlated Continuous Density Hidden Markov Models for Speech Recognition ,IEEE Transactions On Speech And Audio Processing, Vol. 6, No. 4, July 1998. [129].Doh-Suk Kim, Auditory Processing of Speech Signals for Robust Speech Recognition in Real-World Noisy Environments ,IEEE Transactions On Speech And Audio Processing, Vol. 7, No. 1, January 1999. [131].Mark J. F. Gales, Katherine M. Knill, et.al., StateBased Gaussian Selection in Large Vocabulary Continuous Speech Recognition Using HMM s ,IEEE Transactions On Speech And Audio Processing, Vol. 7, No. 2, March 1999. [132].Hui Jiang, Keikichi Hirose , Robust Speech Recognition Based on a Bayesian Prediction Approach ,IEEE Transactions On Speech And Audio Processing, Vol. 7, No. 4, July 1999. [133].Jen-Tzung Chien, Online Hierarchical Transformation Of Hidden Markov Models for Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol.7,No. 6, November 1999. [134]. Hui Jiang, A Minimax Search Algorithm for Robust Continuous Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 8, No. 6, November 2000. [135].George Saon and Mukund Padmanabhan, Data-Driven Approach to Designing Compound Words for continuous Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 9,No. 4, May,2001. [136].Jen-Tzung Chien, Predictive Hidden Markov Model

204

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Selection for Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 13, No. 3, May 2005. [137].Angel de la Torre , Histogram Equalization of Speech Representation for Robust Speech Recognition , IEEE Transactions On Speech And Audio Processing, Vol. 13, No. 3, May 2005. [138].S.Furui, Fifty years of progress in speech and Speaker recognition , Proc.148th ASA meeting, 2004. [139].Pornchai Chanyagorn, Data driven signal Decomposition Method Proc.of the 2005 IEEE International Conf. on Information acquisition, Pgs.464466,July,2005. [140].Andras Kocsor et.al., Kernel Based Feature Extraction with a Speech Technologoy Application , IEEE Transactions on Signal Processing,Vol.52, NO.8, Aug.2004. [141].Albs Sloin et.al., Support Vector Machine Training for Improved Hidden Markov Model , IEEE Transactions on Signal Processing, Vol.56,NO.1,Jan.2008. [142].N.E.Huang et.al.,, The empricial mode decomposition and the Hilbert spectrum for nonlinear and nonstationary timeseries analysisis , Proc. Of the Royal society , London, UK,Vol.454,pp.903-995,1998. [143].Asghar.Taheri , Mohammad Reza Tarihi et.al. , Fuzzy Hidden Markov Models for Speech Recognition on based FEM Algorithm ,Transactions On Engineering, Computing And Technology V4 February 2005 ,ISSN, 1305-5313. [144].Hui Jiang, et.al., Large Margin Hidden Markov Models for Speech Recognition ,IEEE Transactions On Audio,Speech, And Language Processing, Vol. 14, No. 5,September 2006. [145].Xiaodong Cui et.al., A Study of Variable-Parameter Gaussian Mixture Hidden Markov Modeling for Noisy Speech Recognition IEEE Transactions On Audio, Speech, And Language Processing, Vol. 15, No. 4, May 2007. [146].Anant G. Veeravalli et.al, Phoneme Recognition Using Hidden Markov Models , Abstract Submitted to Huntsville Simulation Conference 2004. [147].Li Deng et.al, Structured Speech Modeling ,IEEE Transactions On Audio, Speech, And Language Processing, Vol. 14, No. 5, September 2006. [148].Waleed H. Abdulla, HMM-based techniques for speech segments extraction , Scientific Programming 10 (2002) 221 239 ,IOS Press. [149].Anil K.Jain, et.al., Statistical Pattern Recognition:A Review , IEEE Transactions on Pattern Analysis and Machine Intelligence ,Vol.22, No.1, January 2000. [150].B.H.Juang, The Past, Present, and Future of Speech Processing , IEEE Signal Processing Magazine,May 1998. [151].L.Rabiner, B.Juang, S.Levinson, and M.Sondhi, Recent developments in the application of hidden Markov models to speaker-independent isolated word recognition , In Proc.of IEEE Trans.ASSP, 34(1):52-

59,1986. [152].M.M.Sondhi, New methods of pitch extraction , IEEE Trans.Audio and Electroacoustics,AU-16(2):262-266, June 1968. [153].B.Juang, W.Chou, and C.H.Lee., Minimum classification error rate methods for speech recognition , IEEE Trans.Speech and Audio Proc. T-SAP,5(3):257265, May 1997. [154].L.R.Rabiner, A tutorial on hidden Markov models and selected applications in speech recognition , Proc.IEEE 77(2):257-286.February 1989. [155].K.L.Oehler and R.M.Gray, Combining Image compression and Classification Using vector quantization , IEEE Trans. Pattern Analyssis and Machine Intelligence, Vol.17, no.5,pp.461-473,1995. [156].Q.B.Xie, C.A.Laszlo, and R.K.Ward, Vector Quantisation Technique for Nonparametric Classifier Design , IEEE Trans. Pattern Analysis and Machine Intelligence, vol.15, no.12, pp.1,326-1,330,1993. [157].T.Kohonen, Self-Organizing Maps. Springer Series in Information Sciences, vol.30,Berlin, 1995. [158].P.A.Devijver and J.Kittler, Pattern Recognition:A Statistical Approach , London, Prentice Hall, 1982. [159].E.Oja, Subspace Methods of Pattern Recognition , Letchworth, HeHertfordshire,England:Research Studies Press, 1983. [160].K.Fukunaga, Introduction to Statistical Pattern Recognition , second, Newyork:Academic Press, 1990. [161].J.H.Friedman, Exploratory Projection Pursuit , J.Am.Statistical Assoc.,Vol.84,pp.165-175,1989. [162].P.Hoffbeck and D.A.Landgrebe, Covariance Matrix Estimation and Classification with Limited Training Data, IEEE Trans. Pattern Analysis and Machine Intelligence, vol.18, no.7,pp.763-767, July 1996. [163].Hui Jiang, Xinwei Li, and Chaojun Liu, Large Margin Hidden Markov Models for Speech Recognition ,IEEE Transactions On Audio, Speech, And Language Processing, Vol. 14, No. 5, September 2006. [164].Mathias De Wachter et.al., Template-Based Continuous Speech recognition IEEE Transactions On Audio, Speech, And Language Processing, Vol. 15, No. 4, May,2007 1377 [165].Stefan Windmann and Reinhold Haeb-Umbach, Approaches to Iterative Speech Feature Enhancement and Recognition , IEEE Transactions On Audio, Speech, And Language Processing, Vol. 17, No. 5, July 2009. [166]. David, E., and Selfridge, O., Eyes and ears for computers, Proc.IRE 50:1093-1101,1962.

205

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

An Extension for Combination of Duty Constraints in Role-Based Access Control

Ali Hosseini
ICT Group, E-Learning Center, Iran University of Science and Technology Tehran, Iran .
Abstract—Among access control models, Role-Based Access Control (RBAC) is very useful and is used in many computer systems. Static Combination of Duty (SCD) and Dynamic Combination of Duty (DCD) constraints have been introduced recently for this model to handle dependent roles. These roles must be used together and can be considered as a contrary point of conflicting roles. In this paper, we propose several new types of SCD and DCD constraints. Also, we introduce strong dependent roles and define new groups of SCD constraints for these types of roles as SCD with common items and SCD with union items. In addition, we present an extension for SCD constraints in the presence of hierarchy. Keywords- Role-Based Access Control (RBAC); Combination of Duty (CD); Static combination of Duty (SCD); Dynamic Combination of Duty (DCD); Dependent Roles.

Mohammad Abdollahi Azgomi (Corresponding Author)
School of Computer Engineering, Iran University of Science and Technology Tehran, Iran . complete business job), history-based SD (i.e. no user is allowed to perform all the actions in a business task in the same target or collection of targets). In [7], formal definitions of SD constraints are not provided. Gligor et al. formally defined these constraints and also presented them with more details [8]. For example, SSD constraint was defined in two forms of Strict SSD (SSSD) and one step SSSD (1sSSSD) [8]. SSSD means that conflicting roles are not permitted to perform more than one operation on objects. 1sSSSD means that each two distinct roles in a set of conflicting roles are not permitted to perform operations on the same object. SD constraints are an important topic in RBAC model and many researchers have investigated their aspects and issues. SD constraints force conflicting items such as roles, permissions and so on to be used separately. But, we should not focus only on conflicting items; because, there are dependent items in our environments, which must be used together. In [9], Hosseini and Azgomi introduced Combination of Duty (CD) constraints which handle dependent roles. They proposed Static Combination of Duty (SCD) and Dynamic Combination of Duty (DCD). SCD means that a user must be assigned to dependent roles. DCD means that a user must activate dependent roles. SCD and DCD constraints are not enough to support a wide range of dependent roles. Therefore, it is necessary to define more CD constraints. Also, CD constraints and dependent roles can be considered as a contrary point of SD constraints and conflicting roles, respectively. Hence, we can declare a new CD constraint corresponding to each SD constraint. In this paper, we propose completely the following CD constraints: • Two types of SCD, • SCD with common items, • SCD with union items, • SCD, SCD with common items and SCD with union, items in the presence of hierarchy, and • Five types of DCD. Two types of SCD constraints distribute dependent roles between a set of users. SCD with common items constraints are a strict version of SCD types. Here, common items must be assigned to dependent roles. SCD with union items constraints are an intermediate version

I.

INTRODUCTION

Role-Based Access Control (RBAC) model is accepted well for its numerous advantages such as policy neutrality and efficient access control management [1, 2, 3]. Furthermore, the concept of role is associated with the notion of functional role in an organization. Hence, RBAC models provide intuitive support for expressing organizational access control policies. RBAC has been introduced by Ferraiolo and Kuhn in 1992 and then has been completed by Sandhu et al. in 1996 [4, 3]. In 2001, National Institute of Standards and Technology (NIST) provided a standard for RBAC model [1]. In this model, Static Separation of Duty (SSD) and Dynamic Separation of Duty (DSD) constraints have been defined. Also, SSD has been extended in the presence of hierarchy. SSD means that no user must be assigned to conflicting roles. DSD means that no user must activate conflicting roles within the session. Ahn and Sandhu proposed Role-based Separation of duty Language (RSL99) in the context of RBAC model and then extended this language to Role-based constraints Language (RCL 2000) [5, 6]. Their language can specify Separation of Duty (SD) constraints for conflicting permissions and conflicting users. Simon and Zurko classified completely SD constraints in the role-based environments [7]. They proposed objectbased SD (i.e. no user may act upon a target that the user has previously acted upon), operational SD (i.e. no user may assume a set of roles that have the capability for a

206

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

between SCD with common items and SCD constraints. In RBAC model, roles can inherit other roles. Therefore, we extend SCD constraints in this manner. Five types of DCD constraints distribute dependent roles between a set of sessions and a set of users. By using these new CD constraints, we can specify powerful, exact and flexible policies related to dependent roles. The remainder of this paper is organized as follows. In Section II, the RBAC model and its required detail are mentioned. In section III, new CD constraints are formally defined. Finally, we conclude the paper and mention some future works in Section IV. II. THE NIST RBAC MODEL

RoleOpsOnOb(r, ob)={op: OPS| ((op, ob), r) ∈ PA}

(8)

In 2001, NIST provided a standard for RBAC model [1]. RBAC model is defined in terms of three components: core, hierarchy and constraint. Each component has some elements. Also, some functions are defined in NIST RBAC model. In RBAC model, core includes six elements, called USERS, ROLES, OBS, OPS, PRMS and SESSIONS, which are the sets of users, roles, objects, operations, permissions and sessions, respectively [1]. User is defined as a human being and role is a job function in an organization. Permission is an approval to perform an operation (such as read or write) on one or more objects. Users and permissions are assigned to roles. UA and UP are sets of user-role and permission-role assignments, respectively. PRMS, UA and PA can be formally defined as follows:
PRMS= 2 OPS ×OBS
UA ⊆ USERS × ROLES PA ⊆ PRMS × ROLES

Each session is associated with a single user and each user is associated with one or more sessions. A user can activate a subset of roles assigned to him/her within a session. The following are some functions related to session: • UserSessions function returns a set of sessions associated with the user. • SessionUser function returns the user who is owner of the session. • SessionRoles function returns a set of roles activated within the session. • ActivatedRoles function returns a set of roles activated by the user. The above functions can be formally defined as follows:
UserSessions(u) → 2SESSIONS (9)

∀u1, u2 ∈ USERS: UserSessions(u1) I UserSessions(u2)= ∅

SessionUser(s)={u:USERS|UserSession(u) I {s}={s}}
SessionRoles(s) ⊆ AssignedRoles(SessionUser(s)

(10) (11)
(12)

ActivatedRoles(u)=

U SessionRoles(s)
s∈UserSessions(u )

(1) (2) (3)

We explain some required functions related to core as follows: • AssignedRoles function returns a set of roles assigned to the user. • RolePrms function returns a set of permissions assigned to the role. • RoleObs function returns a set of objects assigned as permissions to the role. • RoleOps function returns a set of operations assigned as permissions to the role. • RoleOpsOnOb function returns a set of operations which can perform on an object by the role. The above functions can be formally defined as follows:
AssignedRoles(u)={r: ROLES | (u, r) ∈ UA}
RolePrms(r)={p: PRMS | (p, r) ∈ PA} RoleObs(r)={ob: OBS| ∃ op ∈ OPS: ((op, ob), r) ∈ PA} RoleOps(r)={op: OPS| ∃ ob ∈ OBS: ((op, ob), r) ∈ PA}

(4)
(5) (6) (7)

A Role Hierarchy (RH) is a subset of ROLES × ROLES and defines a seniority relation between roles, whereby senior roles inherit permissions of their juniors, and junior roles inherit user membership of their seniors. This relation between roles r1 and r2 is denoted by r1 f r2 or (r1, r2) ∈ RH, whereby r1 and r1 are senior and junior roles, respectively. We use the term authorized to refer to both assigned and inherited. Now, we revise the hierarchal version of the above functions for core as follows: • AuthorizedRoles function returns a set of roles, which are authorized for the user. • RoleHPrms function returns a set of permissions, which are authorized for the role. • RoleHObs function returns a set of objects, which are authorized as permissions for the role. • RoleHOps function returns a set of operations, which are authorized as permissions for the role. • RoleHOpsOnHOb function returns a set of operations, which can be performed on an object by the role in the presence of hierarchy. The above functions can be formally defined as follows. In the following definitions, we have ∀ r ∈ ROLES: (r, r) ∈ RH.
AuthorizedRoles(u)= {r:ROLES| ∃r ′ ∈ ROLES: (13)

(r ′, r) ∈ RH ∧ (u, r ′) ∈ UA}
RoleHPrms(r)= {p:PRMS| ∃r ′ ∈ ROLES:
(14)

207

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

(r , r ′) ∈ RH ∧ (p, r ′) ∈ PA} RoleHObs(r)= {ob: OBS| ∃r ′ ∈ ROLES, op ∈ OPS: (r , r ′) ∈ RH ∧ ((op, ob), r ′) ∈ PA} RoleHOps(r)= {op: OPS| ∃r ′ ∈ ROLES, ob ∈ OBS: (r , r ′) ∈ RH ∧ ((op, ob), r ′) ∈ PA} RoleHOpsOnHOb(r, ob)={op: OPS| ∃r ′ ∈ ROLES: (r , r ′) ∈ RH ∧ ((op, ob), r ′) ∈ PA}
(17) (16) (15)

Constrained RBAC includes three types of SD constraints as SSD, SSD in the presence of hierarchy and DSD. Definition 1. Separation of Duty (SD) is defined as a pair (rs, rn), which rs is a set of conflicting roles and rn is a natural number greater than or equal to 2 and less than or equal to the cardinality of rs. Definition 2. Static Separation of Duty (SSD) is a subset of SD, such that each user must be assigned to less than rn conflicting roles. SSD is formally defined as follows, where N is a set of natural numbers:
SSD ⊆ 2ROLES × N (18)

∀ (rs, rn) ∈ SSD, u ∈ USERS: |AssignedRoles(u) I rs|<rn
Definition 3. Static Separation of Duty in the presence of Hierarchy (SSDH) is a subset of SD, such that each user must be authorized for less than rn conflicting roles. SSDH is formally defined as follows:
SSDH ⊆ 2ROLES × N (19)

roles must be used separately. But we do not deal with only such roles. There is another group of roles, which are called dependent roles. These roles are as a contrary point of conflicting roles and must be used together. Because of dependency, if we assign them to distinct users, then we will waste much time to collaborate users. Therefore, using dependent roles separately is not desired and efficient and may be entirely impossible. The importance of dependent roles undergoes a rising trend, because the new methods and tools are generated continuously, which automate the works. Hence, the required time for doing tasks are decreased. Therefore, more roles can be assigned to one person while it was not possible in the past. Organizations are very interested in this topic, because they are looking for the ways to reduce their expenses for better competition. It is obvious that they prefer to assign dependent roles instead of independent roles to one user. Therefore, some policies appear which are based on dependent roles and we need various constraints for specifying these policies. However, there are not enough constraints that are related to dependent roles. In a previous paper [9], we introduced Combination of Duty (CD) constraints which handle dependent roles. CD is defined as a pair (rs, rn) which rs is a set of dependent roles and rn is a natural number greater than or equal to 1 and less than the cardinality of rs. Also, we defined Static Combination of Duty (SCD) and Dynamic Combination of Duty (DCD). We repeat the formal definitions of these constraints in the subsections A and E. Also, we define other types of CD constraints in the following subsections. A. Various types of static combination of duty In this subsection, first we present the formal definition of SCD constraint as appeared in [9]. Then, we propose two other types of SCD constraints. We rename the original SCD as SCDtype I or shortly SCDI and call the new constraints as SCDII and SCDIII. Definition 5. Type I of Static Combination of Duty (SCDI) means that each user must be assigned to no or more than rn dependent roles. SCDI can be formally defined as follows:
SCDI ⊆ 2ROLES × N (21)

∀ (rs, rn) ∈ SSDH, u ∈ USERS: |AuthorizedRoles(u) I rs|<rn
Definition 4. Dynamic Separation of Duty (DSD) is a subset of SD, such that less than rn conflicting roles must be activated within the session. DSD is formally defined as follows:
DSD ⊆ 2ROLES × N (20)

∀ (rs, rn) ∈ DSD, s ∈ SESSIONS: |SessionRoles(s) I rs|<rn
It is worth to mention that the above formal definitions differ from that appeared in [1] for NIST RBAC model, but they have the same meanings. The above definitions are given in order to be compatible with the new constraints we present in the next section. III. DEFINITIONS OF EXTENDED COMBINATION OF DUTY CONSTRAINTS

∀ (rs, rn) ∈ SCDI, u ∈ USERS:
|AssignedRoles(u) I rs|=0 ∨ |AssignedRoles(u) I rs|>rn

In RBAC model and its extensions, which we introduced in section I, the conflicting roles have been investigated by numerous researchers. Hence, various SD constraints have been defined to handle them. Conflicting

Example 1. SCDI = {({r1, r2, r3, r4}, 2)}, AssignedRoles(u1)={r1, r2, r3}, AssignedRoles(u2)={r5}, AssignedRoles(u3)={r1} u1 can satisfy the constraint because u1 is assigned to more than two dependent roles. u2 can satisfy the constraint, because u2 is assigned to no dependent role. u3 cannot satisfy the constraint, because u3 is not assigned to enough dependent roles (i.e. more than two). As mentioned earlier, dependent roles must be used together. SCDI focuses dependency on each user.

208

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Therefore, he/she must be assigned to more than rn dependent roles. We define other versions, which distribute dependency between a set of users. Therefore, this set instead of each user must be assigned to more than rn dependent roles. Definition 6. Type II of Static Combination of Duty (SCDII) means that the user u can be assigned to less than or equal to rn dependent roles if the following conditions are satisfied: • There is the set us of users assigned to less than or equal to rn dependent roles. • u and us are assigned to more than rn dependent roles. We explain the expression “less than or equal to rn” of the first condition that u and us must need together to satisfy the dependency relation. If us is assigned to more than rn dependent roles then u will not have any effect in this satisfaction. SCDII can be formally defined as follows:
SCDII ⊆ 2ROLES × N (22)

n
∃ us1, us2, …, usn ⊆ USERS:

U usi =USERS

∧ ( ∀ usi,

i =1 usj such that i ≠ j: usi I usj= ∅ ) ∧ ( ∀ usi: | (
Roles(x)) I rs|=0 ∨ ( (

U Assigned
x∈usi

AssignedRoles( x)) I rs|>rn ∧ x∈usi

U

( ∀ uss ⊂ usi such that |uss|=|usi|-1: | ( (x)) I rs| ≤ rn)))

U AssignedRoles
x∈uss

∀ (rs, rn) ∈ SCDII:
∀ u ∈ USERS ,such that 0<|AssignedRoles(u) I rs| ≤ rn: ∃ us ⊆ USERS: | (
|(

U AssignedRoles( x)) I rs| ≤ rn
x∈us

∧

AssignedRoles( x )) I rs|>rn x∈us U{u}

U

Example 2. SCDII={({r1, r2, r3, r4}, 2)}, AssignedRoles(u1)={r1}, AssignedRoles(u2)={r2, r3}, AssignedRoles(u3)={r2}, AssignedRoles(u4)={r3}, AssignedRoles(u5)={r1, r2, r3}. u5 is assigned to more than two dependent roles. Hence, this user does not need other users to satisfy SCDII. u1 and u2 are not assigned to enough dependent roles. It is true because {u1, u2} is assigned to enough dependent roles. Also, u1, u3 and u4 are not assigned to enough dependent roles. It is true because {u1, u3, u4} is assigned to enough dependent roles. As observed, u1 is a common user of {u1, u2} and {u1, u3, u4}. Therefore, we define SCDIII that forces the sets to be distinct: Definition 7. Type III of Static Combination of Duty (SCDIII) means that the user u can be assigned to less than or equal to rn dependent roles if the following conditions are satisfied: • There is the set us of users assigned to less than or equal to rn dependent roles. • u and us are assigned to more than rn dependent roles. • u and us do not have this relation with other users. SCDIII can be formally defined as follows:
SCDIII ⊆ 2ROLES × N (23)

Example 3. SCDIII={({r1, r2, r3, r4}, 2)} Step 1: AssignedRoles(u1)={r1}, AssignedRoles(u2)={r2, r3}, AssignedRoles(u3)={r2}, AssignedRoles(u4)={r3}. us1={u1, u2} is assigned to more than two dependent roles and can satisfy SCDIII. Users of this set such as u1 cannot cooperate with other users such as u3 and u4. Hence, we cannot consider us2={u1, u3, u4}. Also, {u3, u4} is not assigned to enough dependent roles. We have two ways to satisfy SCDIII. In the first way, u3 or u4 is assigned to r1 or r4 (i.e. AssignedRoles(u3)={r1, r2}). Therefore, us2={u3, u4} can satisfy SCDIII. In the second way, another user such u5 is assigned to r1 or r4 (i.e. AssignedRoles(u5) ={r1}). Therefore, us2={u3, u4, u5} can satisfy SCDIII. Step 2: AssignedRoles(u1)={r1}, AssignedRoles(u2)={r2}, AssignedRoles(u3) ={r3, r4}. us1={u1, u2, u3} cannot satisfy SCDIII, because a subset of us1 such as {u2, u3} is assigned to more than rn dependent roles. Also, we cannot consider us1={u1, u3} and us2={u2, u3}, because these two sets have a common user (i.e. u3). If u3 is de-assigned from r4, then us1={u1, u2, u3} can satisfy SCDIII. B. Static Combination of Duty with common objects, operations and permissions In this subsection, we define new constraints as SCD with common items, which are strict versions of SCD types. These constraints force strong dependency between roles. Here, common items such as objects, operations and permissions must be assigned to dependent roles. We can define four kinds of these constraints for each type of SCD as follows: • SCD with common objects. • SCD with common operations. • SCD with common objects and operations. • SCD with common permissions. Common items are specified in two ways. In the first way, a set of common items is determined. We use obs, ops and prms, which are sets of common objects, operations and permissions, respectively. These sets cannot be empty and must have some members. In the second way, the number of common items is determined.

∀ (rs, rn) ∈ SCDIII:

209

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

rssopson(ob) =

I RoleOpsOnOb(r , ob)).
r∈rss

SCDCOBI ⊆ 2ROLES × N × (2OBS||N)

(24)

∀ (rs, rn, obs||obn) ∈ SCDCOBI, u ∈ USERS:
|rss|=0 ∨ (|rss|>rn ∧ (obs ⊆ rssobs || obn ≤ |rssobs|)) ,such that rss=AssignedRoles(u) I rs ∧ rssobs = RoleObs(r ). r rss

I

SCDCOB-OP forces to perform all common operations on each common object. But it may be needed to perform some operations on object1, some operations on object2 and so on. This demand can be fulfilled by SCD with common permissions. We mention that the permission is an approval to perform an operation on an object. Definition 11. Static Combination of Duty with Common Permissions (SCDCPRMSI) means that each user must be assigned to no or more than rn dependent roles. Also, in the second state, the intersection of permissions assigned to these roles must include the set prms or have more than or equal to prmn members. SCDCPRMSI can be formally defined as follows:
SCDCPRMSI ⊆ 2ROLES × N × (2PRMS||N) (27)

Definition 9. Type I of Static Combination of Duty with Common Operations (SCDCOPI) means that each user must be assigned to no or more than rn dependent roles. Also, in the second state, the intersection of operations assigned as permissions to these roles must include the set ops or have more than or equal to opn members. SCDCOPI can be formally defined as follows:
SCDCOPI ⊆ 2ROLES × N × (2OPS||N) (25)

∀ (rs, rn, prms||prmn) ∈ SCDCPRMSI, u ∈ USERS:
|rss|=0 ∨ (|rss|>rn ∧ (prms ⊆ rssprms || prmn ≤ |rssprms|)) ,such that rss=AssignedRoles(u) I rs ∧ rssprms = RolePrms(r ). r rss

∀ (rs, rn, ops||opn) ∈ SCDCOPI, u ∈ USERS:
|rss|=0 ∨ (|rss|>rn ∧ (ops ⊆ rssops || opn ≤ |rssops|)) ,such that rss=AssignedRoles(u) I rs ∧
rssops = RoleOps(r ). r rss

I

We join SCDCOB and SCDCOP together to define a stricter version than them. In this version, common operations must perform on common objects. Definition 10. Type I of Static Combination of Duty with Common Objects and Operations (SCDCOB-OPI) means that each user must be assigned to no or more than rn dependent roles. Also, in the second state, the intersection of objects assigned as permissions to these roles must include the set obs or have more than or equal to obn members. In addition, the intersection of operations, which can be performed on each common object by these roles must include the set ops or have more than or equal

Example 4. There are the following specifications. RolePrms(r1)={(ob1, op1), (ob1, op2), (ob2, op1), (ob2, op2)} RolePrms(r2)={(ob1, op1), (ob1, op2), (ob2, op1), (ob2, op2), (ob3,op3)}, RolePrms(r3)={(ob1, op1), (ob3, op3)}, RolePrms(r4)={(ob1, op1), (ob2, op2), (ob4, op4)}, RolePrms(r5)={(ob3, op3), (ob4, op4)}, RolePrms(r6)={(ob1, op1), (ob1, op2), (ob1, op3) (ob2, op1), (ob2, op2), (ob2, op3)}, As a result, we have: RoleObs(r1)={ob1, ob2}, RoleObs(r2)={ob1, ob2, ob3}, RoleObs(r3)={ob1, ob3}, RoleObs(r4)={ob1, ob2, ob4}, RoleObs(r5)={ob3, ob4}, RoleObs(r6)={ob1, ob2}, RoleOps(r1)={op1, op2}, RoleOps(r2)={op1, op2, op3}, RoleOps(r3)={op1, op3}, RoleOps(r4)={op1, op2, op4},

210

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

∈

∈

We use obn, opn and prmn, which are the number of common objects, operations and permissions, respectively. These parameters cannot be zero and must be a natural number. In the formal definitions, we use them as (obs||obn), (ops||opn) and (prms||prmn). The symbol “||” between two expressions means that only one of the two expressions must be considered. Also, we say that “||” may be used for more times in a definition or paragraph. For all “||”, only the left expressions or only the right expressions must be considered. We define SCDI with common items. Definitions of other types are similar to them. Definition 8. Type I of Static Combination of Duty with Common Object (SCDCOBI) means that each user must be assigned to no or more than rn dependent roles. Also, in the second state, the intersection of objects assigned as permissions to these roles must include the set obs or have more than or equal to obn members. SCDCOBI can be formally defined as follows:

to opn members. SCDCOB-OPI can be formally defined as follows:
SCDCOB-OPI ⊆ 2ROLES × N × (2OBS||N) × (2OPS||N) (26)

∀ (rs, rn, obs||obn, ops||opn) ∈ SCDCOB-OPI, u ∈ USERS:
|rss|=0 ∨ (|rss|>rn ∧ (obs ⊆ rssobs || obn ≤ |rssobs|) ∧ ( ∀ ob ∈ obs: ops ⊆ rssopson(ob) || opn ≤ |rssopson(ob)|)) ,such that rss=AssignedRoles(u) I rs ∧ rssobs =

I RoleObs(r )
r rss

∧

∈

∈

I

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

RoleOps(r5)={op3, op4}, RoleOps(r6)={op1, op2, op3}, RoleOpsOnOb(r1, ob1)={op1, op2}, RoleOpsOnOb(r1, ob2)={op1, op2}, RoleOpsOnOb(r2, ob1)={op1, op2}, RoleOpsOnOb(r2, ob2)={op1, op2}, RoleOpsOnOb(r2, ob3)={op3}, RoleOpsOnOb(r3, ob1)={op1}, RoleOpsOnOb(r3, ob3)={op3}, RoleOpsOnOb(r4, ob1)={op1}, RoleOpsOnOb(r4, ob2)={op2}, RoleOpsOnOb(r4, ob4)={op4}, RoleOpsOnOb(r5, ob3)={op3}, RoleOpsOnOb(r5, ob4)={op4}, RoleOpsOnOb(r6, ob1)={op1, op2, op3}, RoleOpsOnOb(r6, ob2)={op1, op2, op3}.

object by these roles (i.e. {op1, op2} on ob1 and {op1, op2} on ob2) includes ops. Step 5: SCDCPRMSI={({r1, r2, r3, r4}, 2, {(ob1, op1), (ob2, op2)})}. AssignedRoles(u2)={r1, r2, r4} can satisfy the constraint, because the intersection of permissions assigned to these roles (i.e. {(ob1, op1), (ob2, op2)}) includes prms (i.e. prms ={(ob1, op1), (ob2, op2)}). C. Static Combination of Duty with union objects, operations and permissions SCD with common items constraints focus on each dependent role and force common items to be assigned to it. In this subsection, we define new constraints as SCD with union items, which focus on a set of dependent roles and forces the items to be assigned to this set. These constraints can be considered as an intermediate version between SCD with common items and SCD constraints. Similar to SCD with common items, there are four kinds of SCD with union items constraints for each type of SCD as follows: • SCD with union objects. • SCD with union operations. • SCD with union objects and operations. • SCD with union permissions. We define SCDI with union objects and operations. Definitions of other kinds and types are similar to it. Definition 12. Type I of Static Combination of Duty with Union Objects and Operations (SCDUOB-OPI) means that each user must be assigned to no or more than rn dependent roles. Also, in the second state, union of objects assigned as permissions to these roles must include the set obs or have more than or equal to obn members. In addition, union of operations, which can be performed on each union object by these roles must include the set ops or have more than or equal to opn members. SCDUOB-OPI can be formally defined as follows:
SCDUOB-OPI ⊆ 2ROLES × N × (2OBS||N) × ( 2OPS||N) (28)

Example 5: There are the following specifications. RolePrms(r1)={(ob1, op1), (ob2, op1)}, RolePrms(r2)={(ob1, op2)}, RolePrms(r3)={(ob3, op4)}, RolePrms(r4)={(ob1, op3), (ob2, op2)}, As a result, we have:

211

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

∈

Step 1: SCDCOBI={({r1, r2, r3, r4}, 2, {ob1, ob2})}. AssignedRoles(u1)={r1, r2, r3} cannot satisfy the constraint, because the intersection of objects assigned to these roles (i.e. {ob1}) does not include obs (i.e. obs={ob1, ob2}). AssignedRoles(u2)={r1, r2, r4} can satisfy the constraint, because the intersection of objects assigned to these roles (i.e. {ob1, ob2}) includes obs. Step 2: SCDCOBI={({r1, r3, r5}, 1, 1)}. AssignedRoles(u3)={r1, r3} can satisfy the constraint, because the intersection of objects assigned to these roles (i.e. {ob1}) has obn member (i.e. obn=1). AssignedRoles(u4)={r3, r5} can satisfy the constraint, because the intersection of objects assigned to these roles (i.e. {ob3}) has obn member. Step 3: SCDCOPI={({r1, r2, r3, r4}, 2, {op1, op2})}. AssignedRoles(u1)={r1, r2, r3} cannot satisfy the constraint, because the intersection of operations assigned to these roles (i.e. {op1}) does not include ops (i.e. ops={op1, op2}). AssignedRoles(u2)={r1, r2, r4} can satisfy the constraint, because the intersection of operations assigned to these roles (i.e. {op1, op2}) includes ops. Step 4: SCDCOB-OPI={({r1, r2, r3, r4}, 2, {ob1, ob2}, {op1, op2})}. AssignedRoles(u2)={r1, r2, r4} cannot satisfy the constraint, because although the intersection of objects assigned to these roles (i.e. {ob1, ob2}) includes obs (i.e. obs={ob1, ob2}), but the intersection of operations which can be performed on each common object by these roles (i.e. {op1} on ob1 and {op2} on ob2) does not include ops (i.e. ops={op1, op2}). AssignedRoles(u5)={r1, r2, r6} can satisfy the constraint, because the intersection of objects assigned to these roles (i.e. {ob1, ob2}) includes obs. Also, the intersection of operations, which can be performed on each common

∀ (rs, rn, obs||obn, ops||opn) ∈ SCDUOB-OPI, u ∈ USERS:
|rss|=0 ∨ (|rss|>rn ∧ (obs ⊆ rssobs || obn ≤ |rssobs|) ∧ ( ∀ ob ∈ obs: ops ⊆ rssopson(ob) || opn ≤ |rssopson(ob)|)) ,such that rss=AssignedRoles(u) I rs ∧ RoleObs(r ) ∧ r rss rssopson(ob) = RoleOpsOnOb(r , ob)). r∈rss rssobs =

U

U

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

RoleObs(r1)={ob1, ob2}, RoleObs(r2)={ob1}, RoleObs(r3)={ob3}, RoleObs(r4)={ob1, ob2}, RoleOps(r1)={op1}, RoleOps(r2)={op2}, RoleOps(r3)={op4}, RoleOps(r4)={op2, op3}, RoleOpsOnOb(r1, ob1)={op1}, RoleOpsOnOb(r1, ob2)={op1}, RoleOpsOnOb(r2, ob1)={op2}, RoleOpsOnOb(r3, ob3)={op4}, RoleOpsOnOb(r4, ob1)={op3}, RoleOpsOnOb(r4, ob2)={op2}. Step 1: SCDUOBI={({r1, r2, r3, r4}, 2, {ob1, ob2})}. AssignedRoles(u1)={r1, r2, r3} can satisfy the constraint, because the union of objects assigned to these roles (i.e. {ob1, ob2, ob3}) includes obs (i.e. obs={ob1, ob2}). Step 2: SCDUOPI={({r1, r2, r3, r4}, 2, {op1, op2})}, AssignedRoles(u1)={r1, r2, r3} can satisfy the constraint, because the union of operations assigned to these roles (i.e. {op1, op2, op4}) includes ops (i.e. ops={op1, op2}). Step 3: SCDUOB-OPI={({r1, r2, r3, r4}, 2, {ob1, ob2}, {op1, op2})}. AssignedRoles(u2)={r1, r2, r3} cannot satisfy the constraint, because although the union of objects assigned to these roles (i.e. {ob1, ob2, ob3}) includes obs (i.e. obs={ob1, ob2}) but the union of operations which can perform on each union object by these roles (i.e. {op1, op2} on ob1 and {op1} on ob2) does not include ops (i.e. ops={op1, op2}). AssignedRoles(u5)={r1, r2, r4} can satisfy the constraint, because the union of objects assigned to these roles (i.e. {ob1, ob2}) includes obs. Also, the union of operations which can perform on each union object by these roles (i.e. {op1, op2, op3} on ob1 and {op1, op2} on ob2) includes ops. Step 4: SCDUPRMSI={({r1, r2, r3, r4}, 2, {(ob1, op1), (ob2, op2)})}. AssignedRoles(u2)={r1, r2, r3} cannot satisfy the constraint, because the union of permissions assigned to these roles (i.e. {(ob1, op1), (ob1, op2), (ob2, op1), (ob3, op4)}) does not include prms (i.e. prms={(ob1, op1), (ob2, op2)}). AssignedRoles(u2)={r1, r2, r4} can satisfy the constraint, because the union of permissions assigned to these roles (i.e. {(ob1, op1), (ob1, op2), (ob1, op3), (ob2, op1), (ob2, op2)}) includes prms.

must be authorized for no or more than rn dependent roles. SCDHI can be formally defined as follows:
SCDHI ⊆ 2ROLES × N (29)

∀ (rn, n) ∈ SCDHI, u ∈ USERS:
|AuthorizedRoles(u) I rs|=0 ∨ |AuthorizedRoles(u) I rs|>rn

Definition 14. Type II of Static Combination of Duty in the presence of Hierarchy (SCDHII) means that the user u can be authorized for less than or equal to rn dependent roles if the following conditions are satisfied. • There is the set us of users authorized for less than or equal to rn dependent roles. • u and us are authorized for more than rn dependent roles. SCDHII can be formally defined as follows:
SCDHII ⊆ 2ROLES × N
∀ (rs, rn) ∈ SCDHII:

(30)

∀ u ∈ USERS ,such that 0<|AuthorizedRoles(u) I rs| ≤ rn: ∃ us ⊆ USERS: | (
AuthorizedRoles( x)) I rs| ≤ rn ∧ x∈us |( AuthorizedRoles( x)) I rs|>rn x∈us U{u}

U

U

Definition 15. Type I of Static Combination of Duty in the presence of Hierarchy with Common Objects and Operations (SCDHCOB-OPI) means that each user must be authorized for no or more than rn dependent roles. Also, in the second state, the intersection of objects authorized as permissions for these roles must include the set obs or have more than or equal to obn members. In addition, the intersection of operations which can be performed on each object by these roles in the presence of hierarchy must include the set ops or have more than or equal to opn members. SCDHCOB-OPI can be formally defined as follows:
SCDHCOB-OPI ⊆ 2ROLES × N × (2OBS||N) × (2OPS||N) (31)

∀ (rs, rn, obs||obn, ops||opn) ∈ SCDHCOB-OPI, u ∈ USERS:
|rss|=0 ∨ (|rss|>rn ∧ (obs ⊆ rsshobs || obn ≤ |rsshobs|) ∧ ( ∀ ob ∈ obs: ops ⊆ rsshopsonh(ob) || opn ≤ |rsshopsonh(ob)|)) ,such that rss=AuthorizedRoles(u) I rs ∧ RoleHObs(r ) ∧ r rss rsshopsonh(ob) = RoleHOpsOnHOb(r , ob)). r ∈rss rsshobs =

I

Definition 16. Type I of Static Combination of Duty in the presence of Hierarchy with Union Objects and Operations (SCDHUOB-OPI) means that each user must

212

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

∈

D. Static Combination of Duty in the presence of hierarchy In this subsection, we revise the previous constraints in the presence of hierarchy, which we already explained in section II. We define hierarchal versions of SCDI, SCDII, SCDCOB-OPI and SCDUOB-OPI. Definitions of the other constraints are similar to them. Definition 13. Type I of Static Combination of Duty in the presence of Hierarchy (SCDHI) means that each user

I

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

be authorized for no or more than rn dependent roles. Also, in the second state, the union of objects authorized as permissions for these roles must include the set obs or have more than or equal to obn members. In addition, the union of operations which can be performed on each object by these roles in the presence of hierarchy must include the set ops or have more than or equal to opn members. SCDHUOB-OPI can be formally defined as follows:
SCDHUOB-OPI ⊆ 2ROLES × N × (2OBS||N) × ( 2OPS||N) (32)

∀ (rs, rn, obs||obn, ops||opn) ∈ SCDHUOB-OPI, u ∈ USERS:
|rss|=0 ∨ (|rss|>rn ∧ (obs ⊆ rsshobs || obn ≤ |rsshobs|) ∧ ( ∀ ob ∈ obs: ops ⊆ rsshopsonhob || opn ≤ |rsshopsonh(ob)|)) ,such that rss=AuthorizedRoles(u) I rs ∧ RoleHObs(r ) ∧ r∈rss RoleHOpsOnHOb(r , ob)). rsshopsonh(ob) = r ∈rss rsshobs =

U

U

Example 6. There are following specifications. RH={(r3, r2)}. RolePrms(r1)={(ob1, op1), (ob2, op1)}, RolePrms(r2)={(ob1, op1), (ob2, op2)}, RolePrms(r3)={(ob1, op4)}, RolePrms(r4)={(ob1, op2)}. As a result, we have: RoleObs(r1)={ob1, ob2}, RoleObs(r2)={ob1, ob2}, RoleObs(r3)={ob1}, RoleObs(r4)={ob1}, RoleHObs(r3)={ob1, ob2}, RoleOps(r1)={op1}, RoleOps(r2)={op1, op2}, RoleOps(r3)={op4}, RoleOps(r4)={op2}, RoleHOps(r3)={op1, op2, op4}, RoleOpsOnOb(r1, ob1)={op1}, RoleOpsOnOb(r1, ob2)={op1}, RoleOpsOnOb(r2, ob1)={op1}, RoleOpsOnOb(r2, ob2)={op2}, RoleOpsOnOb(r3, ob1)={op4}, RoleOpsOnOb(r3, ob2)={}, RoleOpsOnOb(r4, ob1)={op2}, RoleHOpsOnHOb(r3, ob1)={op1, op4}, RoleHOpsOnHOb(r3, ob2)={op2}. The hierarchal versions of functions are equal to original functions for roles, which are not senior of other roles such as r1, r2 and r4. For example, RoleHObs(r1)=RoleObs(r1) and RoleHOpsOnHOb(r1)=RoleObsOnOb(r1). Step 1: Part 1: SCDI={({r1, r2, r3, r4}, 2)}.

AssignedRoles(u1)={r1, r3} cannot satisfy the constraint, because this user is not assigned to enough dependent roles (i.e. more than 2). Part 2: SCDHI={({r1, r2, r3, r4}, 2)}. AssignedRoles(u1)={r1, r3} can satisfy the constraint, because this user is authorized for more than two dependent roles (i.e. {r1, r2, r3}). Step 2: Part 1: SCDCOBI={({r1, r2, r3, r4}, 2, {ob1, ob2})}. AssignedRoles(u1)={r1, r3} cannot satisfy the constraint, because this user is not assigned to enough dependent roles. Also, the intersection of objects assigned to these roles (i.e. {ob1}) does not include obs (i.e. obs={ob1, ob2}). Part 2: SCDHCOBI={({r1, r2, r3, r4}, 2, {ob1, ob2})}. AssignedRoles(u1)={r1, r3} can satisfy the constraint, because this user is authorized for more than two dependent roles and the intersection of objects authorized for these roles (i.e. {ob1, ob2}) includes obs. Step 3: Part 1: SCDUOB-OPI={({r1, r2, r3, r4}, 2, {ob1, ob2}, {op1, op2})}. AssignedRoles(u2)={r1, r3, r4} cannot satisfy the constraint, because the union of operations which can be performed on each object by these roles (i.e. {op1, op2, op4} on ob1 and {op1} on ob2) does not include ops (i.e. ops={op1, op2}). Part 2: SCDHUOB-OPI={({r1, r2, r3, r4}, 2, {ob1, ob2}, {op1, op2})}. AssignedRoles(u2)={r1, r3, r4} can satisfy the constraint, because the union of objects assigned to these roles (i.e. {ob1, ob2}) includes obs (i.e. obs={ob1, ob2}) and the union of the operations, which can be performed on each object by these roles in the presence of hierarchy (i.e. {op1, op2, op4} on ob1 and {op1, op2} on ob2) includes ops. E. Various types of dynamic combination of duty In this subsection, first we present the formal definition of DCD constraint which was already defined in [9]. This constraint is in the base of each session. Therefore, we rename it as DCDS-type I or shortly DCDS-I. We define another version which is based on each user. This version is called as DCDU-I. Then, we propose two other types of each constraint as DCDS-II, DCDS-III, DCDU-II and DCDU-III. Definition 17. Type S-I of Dynamic Combination of Duty (DCDS-I) means that no or more than rn dependent roles must be activated within the session. DCDS-I can be formally defined as follows:
DCDS-I ⊆ 2ROLES × N (33)

∀ (rs, rn) ∈ DCDS-I, s ∈ SESSIONS:
|SessionRoles(s) I rs|=0 ∨ |SessionRoles(s) I rs|>rn

Example 7. DCDS-I={({r1, r2, r3, r4}, 2)}.

213

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

SessionRoles(s1)={r1, r2, r3}, SessionRoles(s2)={r5}, SessionRoles(s3)={r1}. s1 can satisfy the constraint, because more than two dependent roles are activated within it. s2 can satisfy the constraint, because none of the dependent roles are activated within it. s3 cannot satisfy the constraint, because enough dependent roles are not activated within it (i.e. more than two). Definition 18. Type U-I of Dynamic Combination of Duty (DCDU-I) means that no or more than rn dependent roles must be activated by the user. DCDU-I can be formally defined as follows:
DCDU-I ⊆ 2ROLES × N (34)

∀ (rs, rn) ∈ DCDU-I, u ∈ USERS:
|ActivatedRoles(u) I rs|=0 ∨ |ActivatedRoles(u) I rs|>rn

Example 8. DCDU-I={({r1, r2, r3, r4}, 2)}, UserSessions(u1)={s1, s2, s3, s4}. SessionRoles(s1)={r1}, SessionRoles(s2)={r2}, SessionRoles(s3)={}, SessionRoles(s4)={r2, r3}. As a result, we have: ActivatedRoles(u1)={r1, r2, r3}. Enough dependent roles are not activated within each session of u1. But, enough dependent roles are activated by u1. Therefore, DCDU-I is satisfied. As observed, DCDU-I distribute dependency between the sessions of each user. DCDS-I and DCDU-I focus dependency on each session and user, respectively. Hence, more than rn dependent roles must be activated within/by each session/user. We define other versions which distribute dependency between a set of sessions and a set of users, respectively. Therefore, more than rn dependent roles must be activated within/by these sets instead of each session and user. Definition 19. Type S-II of Dynamic Combination of Duty (DCDS-II) means that less than or equal to rn dependent roles can be activated within the session s if following conditions are satisfied. • There is the set ss of sessions that less than or equal to rn dependent roles are activated within it. • More than rn dependent roles are activated within s and ss. SCDS-II can be formally defined as follows:
DCDS-II ⊆ 2ROLES × N (35)

Example 9. DCDS-II={({r1, r2, r3, r4}, 2)}, UserSessions(u1)={s1, s2, s3, s4}, UserSessions(u2)={s5, s6}, UserSessions(u3)={s7, s8, s9}, AssignedRoles(s1)={r1}, AssignedRoles(s5)={r2, r3}, AssignedRoles(s2)={r2}, AssignedRoles(s9)={r3}, AssignedRoles(s7)={r1, r2, r3}. More than two dependent roles are activated with s7. Hence, this session does not need other sessions to satisfy DCDS-II. Enough dependent roles are not activated within s1 and s5. It is true, because enough dependent roles are activated within {s1, s5}. Also, enough dependent roles are not activated within s1, s2 and s9. It is true, because enough dependent roles are activated within {s1, s2, s9}. Definition 20. Type U-II of Dynamic Combination of Duty (DCDU-II) means that the user u can activate less than or equal to rn dependent roles if the following conditions are satisfied. • There is the set us of users which activates less than or equal to rn dependent roles. • u and us activate more than rn dependent roles. DCDU-II can be formally defined as follows:
DCDU-II ⊆ 2ROLES × N (36)

∀ (rs, rn) ∈ DCDU-II:
∀ u ∈ USERS ,such that 0<|ActivatedRoles(u) I rs| ≤ rn: ∃ us ⊆ USERS: | (
ActivatedRoles( x)) I rs| ≤ rn ∧ x∈us ActivatedRoles( x )) I rs|>rn |( x∈us U{u}

U

U

∀ (rs, rn) ∈ DCDS-II:
∀ s ∈ SESSIONS ,such that 0<|SessionRoles(s) I rs| ≤ rn: ∃ ss ⊆ SESSIONS: | (
|(

U SessionRoles( x)) I rs| ≤ rn
x∈ss

∧

SessionRoles( x)) I rs|>rn x∈ss U{s}

U

Example 10. DCDU-II={({r1, r2, r3, r4}, 2)}, ActivatedRoles(u1)={r1, r2}, ActivatedRoles(u2)={r2, r3}, ActivatedRoles(u3)={r1, r2, r3}, ActivatedRoles(u4)={r3}. u3 activates more than two dependent roles. Hence, this user does not need other users to satisfy DCDU-II. u1 and u2 do not activate enough dependent roles. It is true, because {u1, u2} activates enough dependent roles. Also, u1 and u4 do not activate enough dependent roles. It is true, because {u1, u4} activates enough dependent roles. As showed in example 9, s1 is a common session of {s1, s5} and {s1, s2, s9}. Also, in example 10, u1 is a common user of {u1, u2} and {u1, u4}. Therefore, we define DCDS-III and DCDU-III which force the sets to be distinct. Definition 21. Type S-III of Dynamic Combination of Duty (DCDS-III) means that less than or equal to rn dependent roles can be activated within the session s if the following conditions are true. • There is the set ss of sessions that less than or equal to rn dependent roles are activated within it. • More than rn dependent roles are activated within s and ss. • s and ss do not have this relation with other sessions.

214

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

DCDS-III can be formally defined as follows:
DCDS-III ⊆ 2ROLES × N (37)

∀ (rs, rn) ∈ DCDS-III:
n

∃ ss1, ss2, …, ssn ⊆ SESSIONS:

U ssi =SESSIONS
∨ ((

∧

i =1 ( ∀ usi, usj such that i ≠ j: usi I usj= ∅ ) ∧ ( ∀ ssi: | (

Example 12. DCDU-III={({r1, r2, r3, r4}, 2)}, ActivatedRoles(u1)={r1, r2}, ActivatedRoles(u2)={r2, r3}, ActivatedRoles(u3)={r1, r2}, ActivatedRoles(u4)={r3}. us1={u1, u2}, us2={u3, u4}. us1 and us2 do not have common members. Also, they activate more than two dependent roles. Therefore, they can satisfy DCDU-III. IV. CONCLUSIONS In this paper, we proposed various types of combination of duty constraints for handling different aspects of dependent roles. Type I of SCD and DCD focus dependency on each user and session. Type II and III of SCD and DCD constraints distribute dependency between a set of users and a set of sessions. SCD with common items and SCD with union items are strict versions of SCD. The former forces common items to be assigned to each dependent role and the latter forces union items to be assigned to a set of dependent roles. Hierarchal version of SCD constraints considers seniority relation between roles as well as dependency relation. Also, we explain that the usage of dependent roles undergoes a rising trend. We want to do more research about dependent roles in the future works. REFERENCES
[1] D. F. Ferraiolo, et al., “Proposed NIST standard for role-based access control,” ACM Trans. on Information and System Security, vol. 4, no. 3, pp. 224–274, 2001. L. Giuri, “Role-based access control: a natural approach,” Proc. of the 1st ACM Workshop on Role-Based Access Control, 1996. R. S. Sandhu, et al., “Role-based access control models,” IEEE Computer, vol. 29, no. 2, pp. 38–47, 1996. D. F. Ferraiolo and D. R. Kuhn, “Role-based access control,” Proc. of the 15th National Computer Security Conf., pp. 554–563, 1992. G. Ahn and R. Sandhu, “The RSL99 language for role-based separation of duty constraints,” Proc. of the 4th ACM Workshop on Role-Based Access Control, pp. 43–54, 1999. G. Ahn and R. Sandhu, “Role-based authorization constraints specification,” ACM Trans. Information and System Security, vol. 3, no. 4, pp. 207–226, 2000. R. Simon and M. E. Zurko, “Separation of duty in role-based environments,” Proc. of the 10th IEEE Workshop on Computer Security Foundations, 1997. V. D. Gligor, S. I. Gavrila, and D. Ferraiolo, “On the formal definition of separation-of-duty policies and their composition,” Proc. of the 1998 IEEE Symposium on Security and Privacy, pp. 172–183, 1998. A. Hosseini and M. Abdollahi Azgomi, “Combination of duty and historical constraints in role-based access control,” Proc. 6th International Conf. on Innovations in Information Technology (Innovations'09), UAEU, Al Ain, Emirates, Dec. 15-17, 2009.

U SessionRoles( x)) I rs|=0
|(

U Session

x∈ss i x∈ss i Role(x)) I rs|>rn ∧ ( ∀ sss ⊂ usi such that |sss|=|ssi|-1:

U SessionRoles( x)) I rs| ≤ rn)))
x∈sss

Example 11. DCDS-III={({r1, r2, r3, r4}, 2)}, AssignedRoles(s1)={r1}, AssignedRoles(s5)={r2, r3}, AssignedRoles(s2)={r2}, AssignedRoles(s9)={r3}, AssignedRoles(s3)={r1}. ss1={s1, s5}, ss2={s2, s3, s9}. ss1 and ss2 do not have common members. Also, more than two dependent roles are activated within them. Therefore, they can satisfy DCDS-III. Definition 22. Type U-III of Dynamic Combination of Duty (DCDU-III) means that the user u can activate less than or equal to rn dependent roles if following conditions are satisfied. • There is the set us of users which activates less than or equal to rn dependent roles. • u and us activate more than rn dependent roles. • u and us do not have this relation with other users. DCDU-III can be formally defined as follows:
DCDU-III ⊆ 2ROLES × N (38)

[2] [3] [4] [5]

∀ (rs, rn) ∈ DCDU-III:
n

[6]

∃ us1, us2, …, usn ⊆ USERS:

U usi =USERS
i =1

∧ ( ∀ usi,

[7]

usj such that i ≠ j: usi I usj= ∅ ) ∧ ( ∀ usi: | ( Roles(x)) I rs|=0 ∨ ( (

U Activated
x∈usi
n

[8]

U ActivatedRoles(x)) I rs|>r
x∈usi

∧
[9]

( ∀ uss ⊂ usi such that |uss|=|usi|-1: |

U ActivatedRoles

x∈uss (x)) I rs| ≤ rn)))

215

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

An Improved Approach to High Level Privacy Preserving Itemset Mining
Rajesh Kumar Boora
. ($ Corresponding author)
Abstract—Privacy preserving association rule mining has triggered the development of many privacy-preserving data mining techniques. A large fraction of them use randomized data distortion techniques to mask the data for preserving. This paper proposes a new transaction randomization method which is a combination of the fake transaction randomization method and a new per-transaction randomization method. This method distorts the items within each transaction and ensures a higher level of data privacy in comparison to the previous approaches. The pertransaction randomization method involves a randomization function to replace the item by a random number guarantying privacy within the transaction also. A tool has also been developed to implement the proposed approach to mine frequent itemsets and association rules from the data guaranteeing the anti-monotonic property.

Ruchi Shukla$

A. K. Misra

Computer Science and Engineering Department Motilal Nehru National Institute of Technology, Allahabad, India – 211004

implication of the form X => Y where X, Y ⊂ I are sets of items called itemsets and X ∩ Y = Φ. Association rule mining finds the frequent itemsets of a data based on two measurements: Support and Confidence. Definition 1: Let I be a set of n items: I = {a1, a2, …, an}. Let T be a sequence of N transactions: T = {t1, t2, …, tn} where each transaction ti is a subset of I. Given an itemset A ⊆ I, the support of A is defined as suppT (A) = # {t ∈ T | A ⊆ T } .
N

(1)

Keywords; Data Mining, Privacy, Randomization, Association
Rules.

If suppT (A) ≥ Smin, then A is a frequent itemset in T, where Smin is a user-defined parameter called minimum support [3]. Definition 2: The confidence for an association rule X => Y is the ratio of the number of transactions that contain X ∪ Y to the number of transactions that contain X. Confidence of an association rule X => Y = # { X ∪ Y } . (2) #{X } In this work, a new approach to privacy preserving data mining has been proposed. This approach is a mixture of the fake transaction randomization method and a new proposed per-transaction randomization method. The fake transaction randomization method adds fake transactions randomly in between the real transactions [2]. This approach provides good mining results with small probability of error and guaranteed data privacy. However, in recent years, most of the research has focused on privacy preserving data mining using the pertransaction randomization method. Hereby, a different pertransaction randomization approach is proposed, which includes a randomization function to distort each item of every transaction and does not influence the support of any itemset. Besides, reconstruction function applied on the distorted data produces near absolute accurate items. A tool has also been developed to implement the new approach.

I.

INTRODUCTION

Data mining deals with the problem of discovering unknown patterns from data. It includes building models on data, presenting statistical summary of data in human understandable form, deciding upon strategies based on the extracted information. The main consideration in privacy preserving data mining is the sensitive nature of raw data. The data miner, while mining for aggregate statistical information about the data, should not be able to access data in its original form with all the sensitive information. This call for more sophisticated techniques in privacy preserving data mining that intentionally modify data to hide sensitive information, but still preserve the inherent statistics of the data important for mining purpose. Randomization is the only effective approach to preserve the privacy in a system with one data miner and multiple data providers. The discovery of interesting association relationships among huge amounts of business transaction records can help catalog design, crossmarketing, loss-leader analysis, and other business decision making processes. The main goal of the association rule mining is to find out associations or correlations between the items of the particular data involved in the mining. An association rule is an

216

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

II.

RELATED WORK

Many recent publications on privacy have focused on the perturbation model. A perturbation based approach was proposed by Agrawal and Srikant which built a decision-tree classifier from training data [1]. References [4] and [6] presented the problem of association rule mining, where transactions are distributed across multiple sites. References [1] and [7] mine the perturbed data instead of original data and [8] used cryptographic techniques to preserve the privacy in distributed scenarios. Reference [9] presented that the randomization approach can be used to determine web demographics, while the cryptographic approach can be used to support inter enterprise data mining [2]. Reference [10] proposed an algebraic technique to preserve the privacy, which results in identifying association rules accurately but discloses less private information. A growing body of literature exists on different approaches of privacy preserving data mining. Some of these approaches adopted for privacy preserving data mining are briefly summarized below: A. P3P and Secure Database P3P covers the system and architecture design perspectives of privacy preserving data mining. It does not involve any development of algorithm for data mining on sensitive data. P3P provides a way for web site owners to encode their privacy policies in a standard XML format so that users can check against their privacy preferences to decide whether or not to release their personal data to the web site. Koike [11] provides a detailed survey of current P3P implementations. The basic P3P architecture is client based, i.e. privacy of client is defined at web-client end. As opposed to the client-centric implementations, the author in [12] proposed a server-centric architecture for P3P. The work presented in [13] addressed the problem of enforcing the web sites act according to their stated privacy policies by identifying the technical challenges and founding principles in designing a Hippocratic database. B. Secure Multi-Party Computation Secure Multi-Party Computation (SMC) is the problem of evaluating a function of two or more parties’ secret inputs. Each party finally holds a share of the function output. No extra information specific to a party is revealed except what is implied by the party’s own inputs and outputs [14]. The work described in [16] proposed a paradigm of information sharing across private databases based on cryptographic protocols. Compared with the brute force circuit schema, this algorithm is much faster. The work in [17] described several secure multi-party computation based algorithms that can support privacy preserving data mining, e.g., secure sum, secure set union, secure size of set intersection and secure scalar product. The computation is secure if the view of each site during the execution of the protocol can be effectively simulated by the input and the output of the site. This is not the same as saying that private information is protected [6].

C. Data Swapping Data swapping is a simple technique to preserve confidentiality of individual values of sensitive data without changing the aggregate statistical information of the data. The basic idea is to transform the database by switching a subset of attributes between selected pairs of records. As a result, the lower order frequency counts or marginals are preserved and data confidentiality is not compromised [15]. D. Privacy Preserving Distributed Data Mining Kantarcioglu and Clifton proposed the privacy preserving distributed mining of association rules on horizontally partitioned data [6]. They considered the individual sites: Si = {S1, S2, …, Sn}. The criterion is that the each site calculates the locally frequent itemsets and these results are securely transmitted to the global site. Then the global site calculates the globally frequent itemsets. Each and every site calculates the support of itemsets. An algorithm for privacy preserving mining of association rules in distributed databases that builds a global hashing table Hi in every iteration, is proposed by Liu [18]. E. Data Distortion The algorithms belonging to this group work by first distorting the data using randomized techniques. The perturbed data is then used to extract the patterns and models for reconstructing the support of items from perturbed data. The perturbation approach results in some amount of information loss but larger perturbations also lead to a greater amount of privacy. Thus there is a natural trade-off between greater accuracy and loss of privacy [1]. Distortion using multiplicative noise to protect confidentiality of the data is also considered as another option in data mining. The first approach is based on generating random noise that follows truncated normal distribution with unit mean and small variance, and multiplying each element of the original data by the noise. The second approach is inspired by additive random perturbation in a logarithmic domain [7]. The data undergoes a logarithmic transformation, and random noise is generated following a multivariate normal distribution with mean zero and constant variance [19]. The non-negative matrix factorization (NMF) with sparseness constraints for data perturbation is proposed to provide the privacy [20]. III. FAKE TRANSACTION RANDOMIZATION METHOD The fake transaction randomization method generates fake transactions randomly in between the real transactions based on the two characteristics: quantity and quality of fake transactions. The goal of the fake transaction method is to preserve the privacy of critical customer data. As long as these fake transactions look like real transactions and the number of fake transactions is similar to or exceeds the number of real transactions, the privacy of the real transactions is preserved. If the distribution of the lengths of the real transactions differs from that of the fake transactions, privacy breaches are likely. Consequently, if the distribution of the lengths of the real transactions is known and used as a parameter for generating

217

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

fake transactions, then the quality of the fake transactions can be markedly improved [2]. Given w and the number of the real transactions ‘N’ the number of fake transactions is wN. X fake transactions are inserted between every two real transactions, where X is a random variable that is uniformly distributed with a mean of w, a minimum of 0, and a maximum of 2w. The steps by which a fake transaction is generated are as follows. First, the length of the fake transaction is determined using a uniformly distributed random variable Y, whose mean, minimum, and maximum are l, 1 and 2l -1, respectively. Then, Y distinct items are randomly selected from I and these Y items are finally used to generate a fake transaction [2]. If both the mean and the variance of the lengths of the real transactions are known, then the length of the fake transactions can be allowed to follow a normal distribution with the same mean and variance, to improve further the fake transactions. Example 1 Suppose the items provided by the supermarket are green apples, red apples, oranges, bananas and grapes. Consider it as the set of items I = {green apples, red apples, oranges, bananas, grapes}. Transactions of the super market are given in Table I.
TABLE I. Transaction T1 T2 T3 T4 T5 T6 T7 T8 TRANSACTIONS OF SUPERMARKET DATA WITH INTEGER NUMBER REPRESENTATION Customer C1 C2 C3 C4 C5 C6 C7 C8 Items bought (transactions) green apples, grapes oranges oranges, grapes red apples, bananas bananas green apples, red apples green apples, oranges oranges, green apples, grapes Positive number representation 1, 5 3 3, 5 2, 4 4 1, 2 1, 3 3, 1, 5

of the real transactions. To calculate the average length of the real transactions, first find out the length of the each real transaction, which is equal to the number of items involved in that particular transaction. Average length of the real transaction = sum of lengths of real transactions / total number of transactions. In Example 1, the average length of fake transaction (l) is 2. Let w is the ratio of number of fake transactions to the number of real transactions. Consider that three fake transactions will be added for every two real transactions. The total number of fake transactions (wN) =12, where N is the total number of real transactions. Therefore, 12 fake transactions are generated and mixed with the N real transactions. Table II shows the mixed transactions in the Example 1.
TABLE II. SUPERMARKET DATA WITH MIXED TRANSACTIONS Items bought (transactions) 1, 5 3 1, 4 2 5, 4 3, 5 2, 4 1 3, 1 5, 2 4 1, 2 3 4, 2 1, 3 3, 1, 5 5, 3 3 4, 5 1, 4

A.

Transaction T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 T13 T14 T15 T16 T17 T18 T19 T20

The new proposed method represents the items in the super market with the positive numbers (1- green apples, 2 red apples, 3 - oranges, 4 – bananas and 5 – grapes). The super market transactions are replaced by the corresponding positive numbers. In the mining process, the transactions are having the integer numbers as well as strings. It is somewhat difficult to deal with different data types, so it is better to consider all items of transactions with the positive numbers. This makes the mining process easier to produce the results. Now, the fake transactions have to be added to the above real transactions to preserve the privacy. To generate the fake transactions some major characteristics have to be considered. The average length of the fake transactions is determined by calculating the average length of the real transactions and assigns it by an almost near value equal to the average length

So, the probability of selecting a real transaction from the mixed transactions is very less. Therefore, the privacy is guaranteed in between the transactions by mixing the fake transactions to the real transactions. B. Limitations of Fake Transaction Randomization Method The limitation of fake transaction randomization method is that it achieves the privacy up to the level of transaction to transaction only but not within the transactions. That means, this method adds the fake transactions in between the real transactions and it would not distort the any of the items involved in the transactions. By distorting the items within the transactions we can achieve an even higher level of privacy than the earlier. So, we proposed a new per-transaction randomization method to achieve higher privacy. IV. PROPOSED PER-TRANSACTION RANDOMIZATION METHOD

This method involves the modification of data items and can be applied on both the real and fake transactions. It adds

218

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

noise directly to the data items involved within the transactions. Specifically, a per-transaction randomization function is proposed which modifies the data items within the transaction. A. Per-Transaction Randomization Function: Let R is the integer number generated by per-transaction randomization function. The data item within the transaction is replaced by R. R = (item + tnoi + i) % tnoi. (3)

TABLE III.

SUPERMARKET DATA AFTER PER-TRANSACTION
RANDOMIZATION METHOD

Where, item is the data item within the transaction, tnoi is the total number of items i.e. |I|, and i is the random number generated from the random number generator and it is fixed during the entire randomization process. The per-transaction randomization function is applied on each and every data item within the transaction to modify each data item. The process of per-transaction randomization is applied after the fake transaction method, so that a real data item within the transaction can not be identified. The main advantage of this function is that it does not affect the support of any itemset in the mixed transactions. By applying these two methods, privacy is guaranteed between the transactions as well as within the transaction. Suppose, the per-transaction method is applied before the fake transaction randomization method, then there is a large variation occurred between the real and the fake transactions. If the per-transaction randomization method is applied first, then it modifies the real transactions by adding some noise to the data items involved in the transaction. After that, fake transaction randomization method applied, and then the fake transactions are mixed with the resultant transactions of the per-transaction randomization method. The problem here is that, actually the fake transaction randomization method uses I (set of items) to generate the fake transactions. That is, in the mixed transactions only some transactions are took the affect of per-transaction randomization method. In the mining process, at the time of finding out the support of any item, it is resulting in support value error. That is the reason the fake transaction randomization method is applied before the pertransaction randomization method. Therefore, a high level privacy of the customer data is achieved. B. Example 1 After the fake transaction randomization method is applied on Example 1 which is shown in section 3, the pertransaction randomization method is applied on the resultant (Example 1) of fake transaction randomization method. In Example 1, total number of items (toi) = 5, and considered that i = 4, which is generated by the random number generator. The per-transaction randomization function is applied on each transaction in Table II and the resultant transactions are given in Table III. Thus after applying the present approach a higher-level of privacy is provided to the customer data.

Transaction T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 T13 T14 T15 T16 T17 T18 T19 T20

Items bought (transactions) 5, 4 2 5, 3 1 4, 3 2, 4 1, 3 5 2, 5 4, 1 3 5, 1 2 3, 1 5, 2 2, 5, 4 4, 2 2 3, 4 5, 3

V.

SUPPORT RECONSTRUCTION

During the reconstruction of itemsets after fake transaction randomization method and per-transaction randomization method, the items in the transactions are modified to preserve the privacy. Hence, the support of frequent itemsets are affected (i.e., support values are modified). So, there is a need to reconstruct the support of frequent itemsets from the mixed transactions. In order to achieve 100% support of itemsets in data the fake transaction and per-transaction reconstruction methods are applied and are discussed in the following sections. Support Reconstruction of Fake Transaction Method At this point, the transactions of any data are the resultant of both the fake transaction randomization method and pertransaction randomization method. In this section, reconstruction procedure of fake transaction randomization method is discussed. For a given k-itemset A, the number of real transactions that support A in mixed transactions is given as follows [2]: The number of real transactions that support A = the number of mixed transactions that support A – the number of fake transactions that support A. Let S` is the support of some k-itemset A in mixed transactions T`, that is suppT`(A) = S`. Then, the number of transactions that support A is S`(1+w)N, where (1+w)N is the number of transactions in T`. Let support of itemset A in T by S, i.e., suppT (A) = S. Therefore, S can be derived from S` as follows: (excerpted from [2]) S = S` (1 + w) A.

w nCk (2l − 1)

2l −1 Y =k

∑ YCk

(4)

219

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

B. Reconstruction of the Per-transaction Randomization Method After reconstructing the support from the fake transaction method, the next problem is to reconstruct the original items from per-transaction randomization method. So, a pertransaction reconstruction function is proposed to solve the above problem. The reconstruction function is applied on the result of fake transaction reconstruction method i.e., the pertransaction reconstruction function is applied on frequent itemsets. The per-transaction reconstruction function is presented as follows: Oitem = (R – i + |I|) % |I| (5)

where, Oitem is the original item after the reconstruction function, R is the frequent item on which per-transaction randomization function is applied, i is the random number which is passed from random number generator and it is fixed entire the process of reconstruction method and |I| is the total number of items. The per-transaction randomization function does not affect the support value of any item in the transactions and it just modifies the original item by adding the noise. The pertransaction reconstruction function reconstructs the original items without any probability of error i.e., the percentage of error for per-transaction reconstruction method is zero. There is no need to apply the per-transaction reconstruction function on the every item in the mixed transactions. That is the pertransaction reconstruction function is applied only on the result of the mining process i.e., frequent items. Therefore, the per-transaction randomization method provides higher accuracy in finding frequent itemsets and guarantees a higher level of privacy to the customer data. C. Example To find the frequent itemsets in Example 1, the data mining process is applied on transactions in Table II which is real transaction database and Table III (mixed transactions) which is the resultant of fake and per-transaction randomization methods. The data mining process is applied to each item in the set I (set of items). As per definition 1, Support of an item = the number of occurrences of item in transactions total number of transactions Considering the Example 1, it has 5 items in the item set I. Let Smin = 0.4, which is the minimum support provided by the user. Suppose if any item is having the support value greater than or equal to the minimum support, then that item is frequent item. First, the data mining process is applied on Table II i.e. real transactions. The supports for 5 items in real transactions are calculated. Support of items in real transactions S1 = 0.5, S2 = 0.25, S3 =0.5, S4 = 0.25 and S5 = 0.385. The support for item 1 (S1) = 0.5 ≥ 0.4 (Smin) and support for item 3 (S3) = 0.5 ≥ 0.4 (Smin). Therefore, items 1 and 3 are frequent 1-itemsets in real transaction database.

To find out the support of items of real transactions in mixed transactions, first the mining process has to find out the support (S’) of items in mixed transactions. Now, the data mining process is applied on Table III that is mixed transactions. The supports for 5 items in mixed transactions are calculated. Support of items in mixed transactions S’1=0.25, S’2=0.4, ’ S 3=0.35, S’4=0.35 and S’5=0.4. The support for item 2 (S’2) = 0.5 ≥ 0.4 (Smin) and support for item 5 (S'5) = 0.5 ≥ 0.4 (Smin). Therefore, items 2 and 5 are frequent 1-itemsets in mixed transaction database. Now, it has to be check that whether the items 2 and 5 are frequent items of real transactions in mixed transactions or not. It is considered that, n = 5, which is number of items (|I|), N = 20, which is number of real transactions, k = 1, that is k-itemset is having only one item, l = 2, average length of fake transactions, w = 3/2, ratio of number of fake transactions to the number of real transactions, and S’ = 0.4 for both the items 2 and 5, using the Eq. 3. For item 2, S = 0.4 ≥ Smin, and for item 5, S = 0.4 ≥ Smin. For both of the items, the support values are greater than equal to the user-defined minimum support (Smin.). As per the Definition 1, both the items are frequent items of real transactions in mixed transactions. Earlier, the per-transaction randomization method is applied on mixed transactions, so the frequent items 2 and 5 are modified items. To get the reconstructed items, the per-transaction reconstruction method has to be applied on these two frequent items. Eq. 3 is used to find out the reconstructed frequent itemsets. In Example 1, it is considered that total number of items |I| = 5 and i = 4 which is passed to the per-transaction reconstruction function. The per-transaction reconstruction function is applied on items 2 and 4, and it results that, items 1 and 3 are original frequent items of real transactions in mixed transactions. In real transaction database also the items 1 and 3 are frequent items. So, it is concluded that both the reconstructed methods produced the same items (1 and 3) as the frequent items. So, in Example 1, the items 1 and 3 correspond to the green apples and oranges. Therefore, the items green apples and oranges are frequent items. VI. TOOL AND EXPERIMENTAL RESULTS In this section, the snapshots of the tool developed are presented to find out the frequent itemsets and association rules from the transactions of any dataset. The snapshots are taken when the mining process is working with the CSC dataset. The CSC dataset results (i.e., frequent itemsets) are also shown in the snapshots of the tool. Figure 1 is the snapshot of the menu screen in the tool. It allows the user to select any particular functionality, like generating fake transactions, applying per-transaction randomization method, generating database, finding frequent itemsets and finding association rules.

220

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Figure 3 is the screen shot of the tool to find the frequent itemsets in the database and this screen shot using three tables to compare the results. Table III shows the frequent itemset and corresponding support of real transactions in mixed transaction database and Table IV shows the support of frequent itemsets in the real transaction database. The user must select the appropriate radio button when working with the different databases (like real transactions database and mixed transactions database).

Figure 1.

Snapshot of the Menu screen of the tool

Figure 4. Finding support of individual itemset

Figure 2.

Generating fake transaction file

Figure 2 shows how to generate the fake transaction file. The snapshot allows the user to select the real transaction file through the ‘Browse’ button and also allows the user to mention the fake transaction file name that has to be generated. The term ‘user’ corresponds to a company’s authorized person, who is appointed by the company to use the tool. On clicking the ‘Generate Fake Transactions’ button, it adds the fake transactions with the real transactions and privacy is provided in between the transactions. The snapshot for applying per-transaction randomization approach is also same as the above snapshot shown in Figure 2.

Figure 4 shows the support of a particular itemset given by user i.e. it allows the user to find the support of any particular itemset. Figure 5 shows the screen shot of the tool to find the association rules. It allows the user to provide the minimum support, and minimum confidence and it produces the association rules with corresponding support and confidence values.

Figure 5. Finding Association Rules

VII. EXPERIMENTAL RESULTS
Figure 3. Finding frequent itemsets in the database

To test and validate the developed tool, experiments were conducted on the CSC and mushroom datasets. The CSC

221

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

dataset have nearly 300 real transactions, 88 data items. The average length of fake transactions (l) is set to the integer which is closer to the average length of real transactions. It is considered that w=2 and Smin=0.4, then the mining process is applied to find out the support of frequent itemsets. The support of frequent itemsets in real transaction database and support of frequent itemsets of real transactions in mixed transaction database is obtained and is shown in Table IV.
TABLE IV. SUPPORT OF FREQUENT ITEMSETS IN CSC DATASET

accurate reconstructed items for corresponding original items (i.e., 10 for 10, 19 for 19 …). Therefore, the proposed pertransaction randomization approach guarantees privacy within the transactions with improved accuracy.

Itemset CS610 CS611 CS612 CS613

Support in Real Transactions Database 0.4463087248322148 0.40604026845637586 0.40939597315436244 0.40857629514354242

Support of Real transactions in Mixed Transactions Database 0.4429530201342282 0.4214765100671141 0.4080536912751678 0.41879194630872485
Figure 7. Comparison of previous and new per-transaction randomization methods

Consider the frequent item CS611, it is having a small support difference of 0.015436 between the real and mixed transaction database. This difference would not make any problem while finding the frequent itemsets because the support values in the real and mixed transactions is greater than the user-defined minimum support (i.e., Smin = 0.4) and is negligible. Figure 6 shows the closeness of supports of different itemsets in real and mixed transactions database (of CSC dataset). It can be concluded that the support values of frequent itemsets are almost equal in real and mixed transaction databases leading to the inference that the probability of error is infinitesimally small and can be treated as negligible. Therefore, accurate mining results are obtained, highlighting the novelty of the proposed approach.
real database
0.5

The mining process is also applied on mushroom dataset of 8124 real transactions and 28 data items. The average length of fake transactions (l) is set to the integer which is closer to the average length of real transactions. It is considered that w = 2 (i.e., 3 fake transactions are added for every 2 real transactions) and Smin = 0.9, then the mining process is applied to find out the support of frequent itemsets. The support of frequent itemsets in real transaction database and real transactions in mixed transaction database is obtained as shown in Table V.
TABLE V. SUPPORT OF FREQUENT ITEMSETS IN MUSHROOM DATASET Support in Real Transactions Database 0.9741506646480 1.0 0.97538158542 0.92171344165 0.97415066469 0.973165928114 0.975381585425 0.8385032003938 Support of Real transactions in Mixed Transactions Database 0.96878388503200 0.9937961595273 0.9730182176267 0.9293943870014 0.918513047759 0.917134416543 0.9218611521418 0.858493353028065

Itemset number 1

Itemset gillattachment=free veil-type=partial veil-color=white ring-number=one gillattachment=free, veil-type=partial gillattachment=free, veil-color=white veil-type=partial, veil-color=white gillspacing=close

mixed database

0.4

support

2 3 4 5

0.3

0.2

6
0.1 CS610 CS611 CS612 CS613

7 8

itemset

Figure 6. Comparison of supports of items in CSC dataset

Figure 7 shows the comparison between the new approach and the approach of Lin and Liu. The arrows a1, a2, a3 and a4 in Figure 7 indicate the faults in reconstructing the items using the previous approach. At a1, the original item is 10 but the reconstructed item is 7, at a2, the original item is 12 but the reconstructed item is 14, and so on, using the pervious method. The per-transaction randomization approach produces

Table V shows the closeness of supports of different itemsets in real and mixed transactions database (of mushroom dataset). It can be concluded that the support values of frequent itemsets are almost equal in real and mixed transaction databases leading to the inference that the probability of error is infinitesimally small and can be treated as negligible.

222

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

VIII. CONCLUSIONS AND FUTURE WORK In this work, a new approach for privacy preserving association rule mining is presented. This approach provides excellent accuracy in reconstructing frequent itemsets with no influence on support of itemsets. At present, the application of this approach is limited to a local environment i.e., used within the organization. This work can be extended further to deal with a distributed environment. REFERENCES
R. Agrawal and R. Srikant, “Privacy-preserving data mining,” In Proc. of ACM SIGMOD Int. Conf. on Management of Data, 2000, pp. 439450. [2] J. Lin and J. Y. Liu, “Privacy preserving itemset mining through fake transactions,” In Proc. of 22nd Annual ACM Symp. on Applied Computing, 2007, pp. 375-379. [3] A. Evfinievski, R. Srikant, R. Agrawal, and J. Gehrke, “Privacy preserving mining of association rules,” In Proc. of 8th ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 2002, pp. 217228. [4] J. Vaidya and C. Clifton, “Privacy preserving association rule mining in vertically partitioned data,” In Proc. of the 8th ACM SIGKDD Int. Conf. on Knowledge discovery and Data Mining (KDD-02), 2002, pp. 639644. [5] S. J. Rizvi and J. R. Haritsa, “Maintaining data privacy in association rule mining,” In Proc. of 28th Conference of Very Large Database, Hong Kong, China, 2002, pp. 682-693. [6] M. Kantarcioglu and C. Clifton, “Privacy-Preserving distributed mining of association rules on horizontally partitioned data,” IEEE Trans. Knowledge and Data Engineering, Vol. 16, Issue 9, 2004, pp. 1026– 1037. [7] D. Agrawal and C. C. Agrawal, “On the design and quantification of privacy preserving data mining algorithms,” In Proc. of the 20th ACM Symposium on Principles of Database Systems, 2001, pp. 247-255. [8] Y. Lindell and B. Pinkas, “Privacy preserving data mining,” In Proc. of Crypto 2000, LNCS, Vol. 1880, Springer-Verlag, 2000, pp. 36–53. [9] R. Srikant, “Privacy preserving data mining: Challenges and opportunities,” Invited Plenary Talk at 6th Pacific-Asia Conf. on Knowledge Discovery and Data Mining, 2002. [10] N. Zheng, S. Wang, and W. Zaho, “A new scheme on privacy preserving association rule mining,” In Proc. of Springer-Verlag Berlin Heidelberg on PKDD, 2004, pp. 484-495. [11] Y. Koike, “References for p3p implementations”, http://www.w3.org/P3P/implementations, 2004. [12] R. Agrawal, J. Kiernan, R. Srikant and Y. Xu, “Implementing p3p using database technology”, 19th Int. Conf. on Data Engineering, Bangalore, India, 2003. [13] R. Agrawal, J. Kiernan, R. Srikant and Y. Xu, “Hippocratic databases”, 28th Int. Conf. on Very Large Data Bases, Hong Kong, China, 2002. [14] A. C. Yao, “How to generate and exchange secrets”, In Proc. of the 27th IEEE Symp. on Foundations of Computer Science, 1986, pp. 162–167. [1]

[15] T. Dalenius and S. P. Reiss. “Data-swapping: A technique for disclosure control”, Journal of Statistical Planning and Inference, 1982, pp. 73–85. [16] R. Agrawal, A. Evfimievski, and R. Srikant. “Information sharing across private databases”, ACM SIGMOD Int. Conf. on Management of Data, San Diego, CA, 2003. [17] C. Clifton, M. Kantarcioglu, J. Vaidya, X. Lin, and M. Zhu, “Tools for privacy preserving distributed data mining”, ACM SIGKDD Explorations, 2003. [18] J. Liu, X. Paio, and S. Huang, “A privacy-preserving mining algorithm rules in distributed databases,” In Proc. of 1st Int. Multi-Sympos. on Computer and Computational Sciences, Vol. 2, 2006, pp. 746-750. [19] M. E. Nergiz, M. Atzori, and C. W. Clifton, “Hiding of presence of individuals from shared databases,” Proc. of ACM SIGMOD, 2007, pp. 665-676. [20] S. M. A. Kabir, A. M. Youssef and A. K. Elhakeem, “On data distortion for privacy preserving data mining”, Proc. of the IEEE Canadian Confer. on Electrical and Computer Engineering, 2007, pp. 308-311.

AUTHORS PROFILE

Rajesh Kumar Boora is an MTech in Software Engineering from the Motilal Nehru National Institute of Technology, Allahabad, India and is currently serving as an Associate Application Developer at CSC, Hyderabad, India. His research interests include data mining and artificial intelligence. He has around 2 years of experience in industry and has published papers in international conferences. Ruchi Shukla is an MTech in Software Engineering and is currently pursuing Ph.D. in Computer Science and Engineering from the Motilal Nehru National Institute of Technology, Allahabad, India. Her research interests include software engineering, data mining, effort estimation and artificial intelligence. She has over 5 years of experience in teaching and research and has published papers in international conferences and journals. A K Misra is a Ph.D. in Computer Science and Engineering from the Motilal Nehru Regional Engineering College, Allahabad, India and is currently the Professor and Head of the Department of Computer Science and Engineering at Motilal Nehru National Institute of Technology, Allahabad, India. His research interests include software engineering, artificial intelligence and data mining. He has over 35 years of experience in teaching, research and administration and has supervised many graduate and doctoral students. To his credit he has over 65 papers in international journals and conferences.

223

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Call Admission Control performance model for Beyond 3G Wireless Networks
Department of Information Science and Engineering, Acharya Institute of Technology1 Department of Computer Science and Engineering, B.M.S. College of Engineering,2 Department of Electronics and Communication Engineering, B.M.S. College of Engineering,3 Bangalore, INDIA

Ramesh Babu H.S.1, Gowrishankar2, Satyanarayana P.S3.

Abstract— The Next Generation Wireless Networks (NGWN) will be heterogeneous in nature where the different Radio Access Technologies (RATs) operate together .The mobile terminals operating in this heterogeneous environment will have different QoS requirements to be handled by the system. These QoS requirements are determined by a set of QoS parameters. The radio resource management is one of the key challenges in NGWN.Call admission control is one of the radio resource management technique plays instrumental role in ensure the desired QoS to the users working on different applications which have diversified QoS requirements from the wireless networks . The call blocking probability is one such QoS parameter for the wireless network. For better QoS it is desirable to reduce the call blocking probability. In this customary scenario it is highly desirable to obtain analytic Performance model. In this paper we propose a higher order Markov chain based performance model for call admission control in a heterogeneous wireless network environment. In the proposed algorithm we have considered three classes of traffic having different QoS requirements and we have considered the heterogeneous network environment which includes the RATs that can effectively handle applications like voice calls, Web browsing and file transfer applications which are with varied QoS parameters. The paper presents the call blocking probabilities for all the three types of traffic both for fixed and varied traffic scenario. Keywords: Radio Access Technologies, Call admission control, Call blocking probability, Markov model and Heterogeneous wireless Networks.

anywhere and anytime in the Complementary access technologies like Wireless Local Area Networks (WLAN),Worldwide Inter operability for Microwave Access (Wi-Max) and Universal Mobile Telecommunication Systems (UMTS) and which may coexist with the satellite networks [1- 3]. The mobile communication networks are evolving into adaptable Internet protocol based networks that can handle multimedia applications. When the multimedia data is supported by wireless networks, the networks should meet the quality of service requirements. One of the key challenges to be addressed in this prevailing scenario is the distribution of the available channel capacity among the set of multiple users; those are operating with different bandwidth requirements ensuring the QoS requirements of the traffic. The existing admission control strategies can handle the resource management in homogeneous wireless networks but are unable to handle the issue in heterogeneous wireless environment. The mobility of the terminals in the mobile communication environment makes the resource allocation a challenging task when the resources are always in scarce. The efficient call admission control policies should be in place which can take care of this contradicting environment to optimize the resource utilization. The design of call admission control algorithm must take into consideration the packet level QoS parameters like minimum delay, jitter as well as session level QoS parameters like call blocking probability (CBP) and call dropping probability (CDP). The CBP is the probability of denial of accepting the new call and CDP the likelihood of dropping the call by a new access network due to decline of the network resources to an unacceptable level in other words the networks is exhausted with the available resources at which it drops the handover calls. In mobile networks the admission control traffic management mechanism is needed to keep the call blocking probability at a minimal level and another RRM strategy vertical handover plays crucial role in reducing the and call dropping probability in an heterogeneous wireless networks. The further sections of the paper are organized as follows. The section II discusses on the motivation and related work. Section III focuses on the proposed system model for the call

1. INTRODUCTION The recent advances in the wireless networks and mobile devices are inclined towards emerging of ubiquitous computing where the user and application running in the mobile terminal (MT) can enjoy seamless roaming. It is well known that the basic problem in the wireless networks is the scarce of the radio resources. The efficient radio resource management is very essential. The admission control is one of the radio resource management technique this plays dominant role in effectively managing the resources. The admission control in the wireless networks will reduce the call blocking probability in the wireless networks by optimizing the utilization of the available radio resources. The mobile communication environment is featured by moving terminals with different QoS requirements in this current scenario the need of guaranteed QoS. The future users of mobile communication look for always best connected (ABC)

224

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

admission control based on multi dimensional Markov chain. The section IV is focused on the traffic model. The simulation results are represented in section V and conclusion and future work is indicated in section VI. 2. RELATED WORKS At present, dissimilar wireless access networks including 2.5G,3G, Bluetooth, WLAN and Wi-MAX coexist in the mobile computing environment, where each of these Radio access technologies offer complementary characteristics and features in terms of its coverage area, data rate, resource utilization and power consumption. With all these there are constant improvements in the existing technologies offering better performance at lesser cost. This is beneficial in both the end users and service provider’s perspective. The idea of benefiting from integrating the different technologies has lead to the concept of beyond International mobile telephony 2000(IMT-2000) wireless networks known as the next generation wireless networks (NGWN). In this heterogeneous environment, the end user is expected to be able to connect to any of the different available access networks. The end user will also be able to roam seamlessly within these access networks through vertical handover mechanisms. The global roaming is supplemented by the existence of IP networks as the backbone which makes the mobile computing environment to grow leaps and bounds and can effectively address the issue with regard to converge limitations is concerned. In this multifaceted wireless radio environment the radio resource management plays major role. The effective utilization of the limited available resources is the challenge. The admission control is one such challenge that a network service provider face to achieve better system utilization and to provide the best QoS to the users of the network. Call admission control schemes can be divided into two Categories, local and collaborative schemes [4]. Local schemes use local information alone (e.g. local cell load) when taking the admission decision [4, 5, 6]. The Collaborative schemes involve more than one cell in the admission process. The cells exchange information about the ongoing sessions and about their capabilities to support these sessions [7, 8]. The fundamental idea behind all collaborative admission control schemes is to consider not only local information but also information from other cells in the network. The local cell, where the new call has been requested, communicates with a set of cells that will participate in the admission process. This set of cells is usually referred to as a cluster. In [9] for example, the cluster is defined as the set of direct neighbours. The main idea is to make the decision of admission control in a decentralized manner. There are good amount of work reported on homogenous wireless networks and single service wireless networks. There are few works in the heterogeneous wireless networks.

The Call admission control in Heterogeneous networks is a real challenge. The varied QoS requirements of multimedia applications and the coexistence of different RATs, facade major challenges in designing CAC algorithms for next generation heterogeneous wireless networks. The challenges are heterogeneous networking, multiple service classes, flexible in bandwidth allocation and cross layer issues based design. Some of the issues the call admission control mechanism should address and point of interest of our research work are as follows. Firstly, The B3G networks should be able to accommodate the applications and user with different QoS requirements, so the CAC algorithms should be designed handle different classes of service meet the QoS needs of all types of applications. Second, there will be diversity in multimedia applications and mobile users QoS requirements in NGWN, The resource utilization and QoS performance can be improved by adaptive bandwidth allocation. This clearly indicates that the CAC should be designed taking into consideration the flexible bandwidth allocation, where, more resources can be allocated when the there is less traffic and the allocated bandwidth can be revoked when there is congestion. The NGWN has different RATs coexisting which are with different capabilities and they should cater the varied QoS requirements of multimedia applications admission control with single criteria mat be too trivial, in this prevailing scenario the admission control decision should be based on Multi criteria such that the optimization user satisfaction and selection of optimal RAT is achieved. The multi criteria decision making system is an optimization technique used to analyse the contradicting decision making parameters. The MCDM based decision making systems are generally used in the fields o reliability, financial analysis, social and political related analysis and environmental impact analysis etc. There are several algorithms proposed on handling the admission control decision making using MCDM in heterogeneous wireless networks. This section discusses one specific admission control algorithm based on multiple criteria on which the further work is planned namely computation-Intelligence-based CAC. The computation-Intelligence-based CAC use evolutionary approaches like Genetic Algorithm (GA), fuzzy logic and Artificial Neural Networks(ANN) [10].The Majority of the computational-intelligence-based CAC algorithms incorporate fuzzy logic[11],fuzzy neural[12] and fuzzy MCDM[13-14] methods. There are very few works reported on the usage of Fuzzy Neural Artificial Neural Networks in CAC. 3. SYSTEM MODEL In this paper we propose a novel admission control mechanism for effectively handling the call blocking probability in multi class traffic heterogeneous network environment there by increasing the resource utilization. This would in turn achieve the objective of guaranteeing the user

225

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

QoS requirements. The proposed model is able to handle three types of the traffic. The traffic considered for the study involves conversation traffic, interactive traffic and background traffic. The representative applications could be voice calls, Web browsing and file transfer applications respectively. The proposed model is developed keeping in mind the WCDMA, Wi-Fi and Wi-Max The CAC mechanism proposed is focused only on the system’s ability to accommodate newly arriving users in terms of the total channel capacity which is needed for all terminals after the inclusion of the new user. In the case when the channel load with the admission of a new call, was precompiled (or computed online) to be higher than the capacity of the channel the new call is rejected, if not, the new call could be admitted. The decision of admitting or rejecting a new call in the network will be made only based on the capacity needed to accommodate the call. Here the heterogeneous network which comprises a set of RATs, Rn, with co-located cells in which radio resources are jointly managed. Cellular networks such as Wireless LAN and Wi-Max can have the same and fully overlapped coverage, which is technically feasible, and may also save installation cost [4]. H is given as H {RAT 1, RAT 2, RAT k} where K is the total number of RATs in the heterogeneous cellular network. The heterogeneous cellular network supports n-classes of calls, and each RAT in set H is optimized to support certain classes of calls. The paper presents the analytical model for Call admission control mechanism in heterogeneous wireless networks is modelled using higher order Markov chain as shown in figure2.The study considers that, whenever a new user enters the network will originate the network request at the rate λi and is assumed to follow a Poisson process. The service time of the different class of traffic and types of calls is µi .The mean service time of all types of users were assumed to follow negative exponential distribution with the mean rate 1/µ. Since Voice traffic is exhibits Erlang distribution, the condition that is considered for simulation is Negative Exponential distribution .The parameters of analytic performance model are also called as Performance model parameters and are number of virtual channels (N), user arrival rate (λ), arrival rate of type1 call (λ1), arrival rate of type2 call (λ2) arrival rate of type3 call (λ3) and service time of the user (μ). The total number of virtual channel in the system are N. When the numbers of available channels are below the specified threshold the system will drop the calls. The threshold limit is determined by three positive integers A1, A2 and A3. When the number of available channels falls below the threshold A3 the proposed system will accept only the voice calls and web browsing. When the number of available channels falls below the threshold A2 the proposed system will accept only the voice calls and when the available number of channels falls below the threshold A1, the proposed system

will not accept any calls, it reaches the stage where there will be no channels available to allocate to the incoming calls and leads to system blocking. The P (0) is probability that there are no allocated channels in the designated system. The equations (1) - (3) are lower boundary equations for the system states P0 , P1 and P2 respectively

λ1 P0 + λ2 P0 + λ P0 - μ 2 P2 - μ1 P - μ3 P3 = 0 1
3

(1) (2) (3)

λ1 P + λ2 P + λ3 P − μ1 P2 − μ2 P3 − μ3 P3 = 0 1 1 1
λ1 P2 + λ2 P2 + λ P2 + μ1 P2 + μ2 P2 − μ1 P3 − μ2 P4 − μ3 P5 = 0
3

The equations (4) - (6) are upper boundary equations for the system states Pn, P n-1 and Pn-2 . They are expressed as
P−3(λ +λ2 +λ3 +μ +μ2 +μ3)−λP−4 −λ2P−5 −λ3P−6 −μP−2 −μ2P−1 −μ3P =0 n 1 1 1 n n n 1 n n n

(4) (5)

P−2(λ +λ2 +μ1 +μ2 +μ3) −λP−3 −λ2P−4 −λ3P−5 −μ1P−1 −μ2P =0 n 1 1 n n n n n

Pn−1 (λ1 + μ1 + μ2 + μ3 ) − λ1Pn−2 − λ2 Pn−3 − λ3 Pn−4 − μ1Pn = 0

(6)

The repeated states are those which are in-between these states i.e. between upper and lower boundaries based on figure1. The repeated states of the system are represented in a generic form as shown in (7).
P4 (λ1 + λ2 + λ3 + μ1 + μ2 + μ3 ) - λ1P3 − λ2 P2 − λ3 P -μ1P6 - μ2 P6 -μ3 P7 = 0 5

(7) The equation that can be presumed as the general equation for call blocking probability for traffic type 1 is represented in (8).

Pn =

λ1 Pn-1 + λ 2 Pn − 2 + λ Pn-3 (μ1 + μ 2 + μ 3 )
3

(8)

Assuming and λ1= λ2= λ3= λ and μ1=μ2 =μ3 = μ, the call blocking probability for type 1 traffic could be expressed as

Pn =

a ( Pn-1 + Pn − 2 + Pn-3 ) 3

(9)

Similarly, the call blocking probability for type 2 traffic is

Pn −1 =

a ( Pn-2 + Pn − 3 + Pn-4 ) 3
a ( Pn -3 + Pn − 4 + Pn -5 ) 3

(10)

And for Type 3 traffic is
Pn − 2 =

(11)

226

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

The call blocking probability for the overall system traffic
λ3
λ2
λ1
μ1

λ3
λ2 λ2
λ1
μ1

λ3
λ2
λ1
μ1

λ3

λ3

λ3

λ3
λ2
λ1

λ3

λ3
λ2
λ1 λ1

λ2
λ1
μ1

λ2
λ1

λ1
μ1

Pn − 3
μ1

Pn − 2
μ1
μ1

Pn −1
μ1

Pn

μ2
μ3

μ2
μ3

μ2
μ3

μ2
μ3

μ2
μ3 μ3

μ2
μ3 μ3

μ2
μ3

μ2

Figure 1: Performance model

The call blocking probability for the overall system traffic Pnb can be expressed as

Pn b =

a ( Pn + Pn −1 + Pn -2 ) 3
4. TRAFFIC MODEL

(12)

If we consider a system which is stationary time series system, the series can be modelled as y(t) = f(xt) , where xt = Where k is the number of factors determines each element of the time series, hence the no stationary time series can be modelled as shown in (13) and (14).

Lognormal process or time varying Poisson process and the session inter arrival is a Bi Pareto, Wei bull, Markov Modulated Poisson Process (MMPP) or Time varying Poisson process [15] [17]. The f3 (t) is the size of each session, f3 (t) is discrete value continuous time process and either Bi Pareto or Lognormal random Process [15]. [y (t-1),y(t-2),…y(t-m)]. However this is not true in case of non stationary time series, the entire series cannot be determined by single function f(.), instead by set of functionf1 , f2,…..fk . 5. SIMULATION RESULTS AND DISCUSSION In this section, we present the numerical results and compare the call blocking probabilities of the different types of traffic. The experiment setup is conducted keeping 2 types of traffic constant varying the other type .The second set of experiment setup is conducted varying all the three types of traffic. The first set of experiments is indicated by the simulation result shown in figure 2.The call blocking probability for a system with N channels which supports three types of traffic is conducted. The experiment considers that, whenever a new user enters the network will originate the network request at the rate λ1 for type1 traffic, and λ2 for type2 traffic and λ3 for type3 traffic and is assumed to follow a Poisson process. The service time of the different types of traffic based calls is considered as µ1 for type1 traffic, µ2 for type3 traffic and µ3 for type3 traffic and is assumed to follow a Lognormal random Process. For the first set of experiments we have considered the arrival rate of all the three types of traffic as λ and service rate of all the three type of calls is same and is equal to µ. The arrival rate of the calls was taken as the varying traffic intensity of Type1 traffic and blocking probability of the type 1, type2, type3 traffic, and overall call blocking probability of the system is plotted. The Figure 2 shows call blocking probability for all three types of traffic when then intensity of the type1 traffic is increased. The horizontal axis shows the number of users with type1 traffic while the vertical axis shows the call blocking probability of all types of traffic. The simulation results show that the call blocking probability of the system and different types of traffic will

y (t ) =

∑
k

k

p it f i ( t )
(13)

i =1

∑

p it = 1

i =1

(14)

Where fi(t) is a random process and hence the non stationary time series can be modelled by a set of random processes[15].Since the traffic at AP or BS is a non stationary time series, this can be determined by number of users and their arrival pattern, numbers of sessions(application) of individual user, session inter arrival and size of each session. All these are random processes [16].
Tr ( t ) =

∑

3

p it f i ( t )

i =1

(15)

Where f1 (t) is random process of number of users and the user arrival pattern, f1 (t) is discrete value continuous time process User distribution either uniform or lognormal process and user arrival pattern is time varying Poisson process [16] [17] [18]. The f2 (t) is random process of number of sessions and session inter arrival pattern. The f2 (t) is a discrete value and continuous time random process. User session is

227

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

increase with the increase in the intensity of type1 traffic. The simulation results with increase the intensity of type2 traffic and simulation results with increase the intensity of type3 also showed the similar kind of results. The second set of experiments conducted will present the numerical results and compare the call blocking probabilities of the different types of traffic. The proposed a performance model for call admission control mechanism in the heterogeneous RATs and analysing the call blocking probability keeping the variation in the number of channels was conducted . The experiment setup is conducted considering the varying the traffic intensity of type1 traffic the blocking probability of type 1 and blocking probability of the type 2, and type3 traffic of the system is plotted. The Figure 3 shows call blocking probability for all three types of traffic when then intensity of the type 1 traffic is increased. The horizontal axis shows the number of users with type 1 traffic while the vertical axis shows the call blocking probability of all types of traffic.

The parameters of analytic performance model are called as Performance model parameters are number of virtual channels (N), user arrival rate (λ), arrival rate of type 1 call (λ1), arrival rate of type 2 call (λ2.) arrival rate of type 3 call (λ3) and service time of the calls is taken as µ1 , µ2 and µ3. The simulation results show that the call blocking probability of the different types of traffic will increase with the increase in the intensity of type1 traffic. The simulation results with increase the intensity of type2 traffic and simulation results with increase the intensity of type3 also showed the similar kind of results. The simulation results indicate that at particular state the call blocking probability of all three types of traffic will be minimal. 6. CONCLUSION AND FUTURE WORK In this paper, we have proposed a performance model for call admission control mechanism in the heterogeneous RATs and analysing the call blocking probability keeping the variation in the number of channels. In order to measure the call blocking probability of the analytical model the simulation study was conducted and following observations were recorded. Firstly, increase in the number of type 1 users will increase the call blocking probability of type2 and type 3 calls and vice versa. Second, Increase in the traffic intensity of one type of traffic will increase the system blocking probability. The concept of minimizing the call blocking probability is an optimization technique to provide fair QoS to the set of users in the wireless network and there is also a need of intelligent call admission control strategy in the admission control mechanism to make the decision of accepting are rejecting a call keeping the blocking probability minimal in a heterogeneous RATs based network working under dynamic network condition. The future work of this research includes applying intelligence for the process of decision making .The future works includes the use of Neural Network (NN) based decision making based on different criteria in making the decision of admitting or rejecting the call. 7. REFERENCES
X.G. Wang, J.E. Mellor, G. Min, A QoS-based bandwidth management scheme in heterogeneous wireless networks international journal of simulation systems, Science and Technology, ISSN:1473-8031,5 (1-2) (2004) 9–17. [2] E. Vanem, S. Svaet, F. Paint, Effects of multiple access alternatives in wireless networks, IEEE Wireless and heterogeneous Networking(2003) 1696–1700. [3] K. Murray, R. Mathur, D. Pesch, Network access and handover control in heterogeneous wireless networks for smart space environments, in: First International Workshop on Management of Ubiquitous Communications and Services, MUCS, Waterford, Ireland, December 11, 2003. [4]. T. Zhang, E.v.d. Berg, J. Chennikara, P. Agrawal, J.-C. Chen, T. Kodama, Local predictive resource reservation for handoff in multimedia wireless IP networks, IEEE Journal on Selected Areas in Communications (JSAC) 19 (10) (2001) 1931–1941. [5] C.-T. Chou, K.G. Shin, Analysis of combined adaptive bandwidth [1 ]

Figure 2.Call blocking probability of the system

Figure 3. call blocking probablity of varying traffic

228

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

allocation and admission control in wireless networks, in: IEEE Conference on Computer Communications (INFOCOM), 2002 pp. 676– 684. [6] C.W.Ahn, R.S. Ramakrishna, QoS provisioning dynamic connection admission control for multimedia wireless networks using a Hopfield neural network, IEEE Transactions on Vehicular Technology 53 (1)(2004) 106–117. [7] D. Hong, S. Rappaport, Traffic modelling and performance analysis for cellular mobile radio telephone systems with prioritized and non prioritized handoff procedures, IEEE Transactions on Vehicular Technology 35 (1986) 77–92. [8] S. Choi, K.G. Shin, Adaptive bandwidth reservation and admission control in QoS-sensitive cellular networks, IEEE Transactions on Parallel and Distributed Systems 13 (9) (2002) 882–897. [9] M. Naghshineh, M. Schwartz, Distributed call admission control in mobile/wireless networks, IEEE Journal on Selected Areas in Communications (JSAC) 14 (4) (1996) 711–717. [10] R.T. Marler and J.S. Arora, “Survey of multi-objective optimization methods for engineering”, Structural and multi disciplinary optimization, Vol.26.No.6,pp 369-395, 2004. [11] P.M.L. Chan, R.E. Sheriff, Y.F. Hu, P. Conforto, C. Tocci, Mobility management incorporating fuzzy logic for a heterogeneous IP environment, IEEE Communications Magazine 39 (12) (2001) 42–51. [12] R. Agusti, O. Sallent, J. Perez-Romero, L. Giupponi, A fuzzy-neural ´ based approach for joint radio resource management in a beyond 3G framework, in: 1st International Conference on Quality of Service in Heterogeneous Wired/Wireless Networks (Qshine’04), Dallas, USA, October, 2004. [13]W. Zhang, Handover decision using fuzzy MADM in heterogeneous networks, in: Proceedings of IEEE WCNC’04, Atlanta, GA, March, 2004. [14] A.L. Wilson, A. Lenaghan, R. Malyan, Optimizing wireless network selection to maintain QoS in heterogeneous wireless environments, in: Proceedings of World Communication Forum, Denmark, September, 2005.

[15] Dahlhaus. R, Fitting Time Series Models to Non stationary Processes, The Annals of Statistics, vol.25, no.1, 1997, 1 -37. [16] Felx.H, Merkouris.K. Maria.P and et.al, Spatio-Temporal Modelling of Traffic Workload in a campus WLAN, First International workshop on Technology and policy for Accessing Spectrum (TAPAS’06),Boston , Massachusetts, USA 2006. [17] George.M, Strasky.H.Y.W, Yuan.Y and Songwu .L, Characterizing flows in Large Wireless Data Network, Mobicom’ 04, Philadelphia, Pennsylvania, USA 2004. [18]B.Anand, M.V.Geoffrey, P. Bahal and P.Venkatragan, “Characterizing the user Behaviour and network performance in public wireless LAN”, ACM Sigmetrics , pp 195-205,June 2002

AUTHORS
Ramesh Babu.H.S is with the Information Science and Engineering Department, Acharya Institute of Technology, Visvesvaraya Technological University, Soladevanahalli, Bangalore-560 090, Karnataka, INDIA (e-mail: rameshbabu@acaharya.ac.in)

Dr.Gowrishankar is with the Computer Science and Engineering Department, B.M.S. College of Engineering, Visvesvaraya Technological University, P.O. Box. 1908, Bull Temple Road, Bangalore-560 019, Karnataka, INDIA (e-mail: gowrishankar.cse@bmsce.ac.in)

Dr.P.S.Satyanarayana was with the Electronics and communication Engineering Department, B.M.S. College of Engineering, Visvesvaraya Technological University, P.O. Box. 1908, Bull Temple Road, Bangalore-560 019, Karnataka, INDIA (e-mail: pss.ece@bmsce.ac.in).

229

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Efficient Candidacy Reduction For Frequent Pattern Mining
M.H Nadimi-Shahraki1, Norwati Mustapha2, Md Nasir B Sulaiman2, Ali B Mamat2
1

Faculty of Computer Engineering, Islamic Azad University, Najafabad branch, Iran, And Ph.D. Candidate of Computer Science, University of Putra Malaysia 2 Faculty of Computer Science and Information Technology, University of Putra Malaysia (UPM), Selangor, Malaysia.

Abstract— Certainly, nowadays knowledge discovery or extracting knowledge from large amount of data is a desirable task in competitive businesses. Data mining is a main step in knowledge discovery process. Meanwhile frequent patterns play central role in data mining tasks such as clustering, classification, and association analysis. Identifying all frequent patterns is the most time consuming process due to a massive number of candidate patterns. For the past decade there have been an increasing number of efficient algorithms to mine the frequent patterns. However reducing the number of candidate patterns and comparisons for support counting are still two problems in this field which have made the frequent pattern mining one of the active research themes in data mining. A reasonable solution is identifying a small candidate pattern set from which can generate all frequent patterns. In this paper, a method is proposed based on a new candidate set called candidate head set or H which forms a small set of candidate patterns. The experimental results verify the accuracy of the proposed method and reduction of the number of candidate patterns and comparisons. Keywords- Data mining; Frequent patterns; Maximal frequent patterns; Candidate pattern

disciplines including statistics, database systems, machine learning, intelligent computing and information technology.

Figure 1. Data mining as a main step in KDD process

I.

INTRODUCTION

The explosive growth of data in all business, government and scientific applications creates enormous hidden knowledge in their databases. Certainly, in this decade knowledge discovery or extracting knowledge from large amount of data is a desirable task in competitive businesses. For example, daily a large amount of purchase data called market basket transactions are collected in the cashier counters of huge markets. The market management systems are interested in analyzing the purchase data to understand more about the behavior of their customers. The association analysis can represent interesting relationships hidden in large data set in the form of association rules. For example, 75% of customers who buy diapers also buy orange juice. These rules can be used to identify new opportunities for cross-selling markets’ products to their customers. The association analysis is also useful in other applications such as web mining scientific applications. Data mining therefore appears to address the need of sifting useful information such as interesting relationships hidden in large databases. As shown in Figure 1, data mining is an essential step in the process of knowledge discovery from data (KDD) to extract data patterns. It is a composite process of multiple

Since the first introducing [2], frequent patterns mining plays an important role in data mining tasks such as clustering, classification, prediction and especially association analysis. Frequent patterns are itemsets or substructures that exist in a data set with frequency no less than a user specified threshold. Identifying all frequent patterns is the most time consuming process due to a massive number of candidate patterns. In general, 2i-1 candidate patterns can be generated from a data set contains i items. Therefore the computational requirements for frequent patterns mining are very expensive. For the past decade there have been an increasing number of efficient algorithms to mine the frequent patterns by satisfying the minimum support threshold. They are almost based on three fundamental frequent patterns mining methodologies: Apriori, FP-tree and Eclat [9]. The Aprioribased algorithms significantly reduce the size of candidate sets using the Apriori principle that says all subsets of an infrequent itemset must be infrequent. But they still suffer from the generate-and-test strategy. They mine frequent patterns by generating candidates and checking their frequency against the transaction database. The FP-tree keeps only frequent items respect to the minimum support threshold by two database scan. Recently some FP-tree-based algorithms have been developed to capture the content of the transaction database only by one database scan which can be very useful for incremental updating of frequent patterns [12]. They usually traverse the tree to mine frequent patterns without candidate generation in same fashion. Only a few of them can fit the

________________________________
1Corresponding Author
230 http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

content of the transaction database in memory to eliminate the database rescanning and the mining model restructuring. Meanwhile reducing the number of candidate patterns and the comparisons for support counting are still two problems in this field which have made the frequent pattern mining one of the active research themes in data mining. A reasonable solution is identifying a small candidate pattern set from which can generate all frequent patterns. In this paper, a method is proposed based on a new candidate set called candidate head set or H which form a small set of candidate patterns. It is an improvement of our previous method presented for maximal frequent pattern mining [13]. The proposed method is based on prime number characteristics for frequent pattern mining including a data transformation technique, an efficient tree structure called Prime-based encoded and Compressed Tree or PC_Tree and mining algorithm PC_Miner. The salient difference is that mining process makes use of the candidate head set and its properties to reduce the number of the candidate sets and comparisons. The PC_Miner algorithm strives to find long promising patterns during of the initial steps based on the candidate head set and its properties. Consequently, it prunes the search space efficiently. The rest of this paper is organized as follows. Section 2 introduces the problem and reviews some efficient related works. The proposed method is described in section 3. The experimental results and evaluation show in section 4. Finally, section 5 contains the conclusions and future works. II. PROBLEM DEFINITION AND RELATED WORK

contains a number of shorter frequent sub-patterns. Given the itemset lattice shown in Figure 2, which presents the list of all possible itemsets for L= {A, B, C, D, E}. A brute-force approach for mining frequent itemsets is to count the support every candidate itemsets in the lattice structure. It needs to compare each itemsets against every transaction. Obviously, this approach can be very expensive and it needs O (NML) comparisons, where N is the number of transactions, M is equal 2i-1 candidate itemsets for i items and L is maximum length of transactions. Therefore there are two main ideas to reduce the computational complexity of frequent itemsets mining. Firstly, reducing the number of candidate itemsets and secondly reducing the number of transactions those must be compared to count the support of the candidate itemsets.

Simply, frequent patterns are itemsets or substructures that exist in a dataset with frequency no less than a user specified threshold. The first definition of frequent itemset was introduced for mining transaction databases (Agrawal et al. 1993). B. Problem Definition Let L= {i1, i2 … in} be a set of items and D be a transaction database where each transaction T is a set of items and |D| be the number of transactions in D. Given X= {ij … ik} be a subset of L (j ≤ k and 1 ≤ j, k ≤ n) is called a pattern. The support of the pattern X or Sup (X) in D is the number of transactions in D that contains X. The pattern X will be called frequent if its support is no less than a user specified support threshold min_sup σ (0 ≤ σ ≤ |D|). The problem of frequent pattern mining is finding all frequent patterns from dataset D with respect to specified min_sup σ. Various kinds of frequent patterns can be mined from different kinds of data sets. In this research, we use itemsets (sets of items) as a data set and the proposed method is for frequent itemset mining, that is, the mining of frequent itemsets from transactional data sets. However, it can be extended for other kinds of frequent patterns. The complexity of frequent patterns mining from a large amount of data is generating a huge number of patterns satisfying the minimum support threshold, especially when min_sup σ is specified low. This is because, all sub-pattern of a frequent pattern are frequent as well. Therefore a long pattern A.

Figure 2. The itemset lattice for item set L= {A, B, C, D, E}

Related work the Agrawal and Srikant [3] introduced an interesting property called Apriori principle, to reduce the number of candidate sets among frequent k-itemsets: if a k-itemset is frequent then all of its sub-itemsets must also be frequent. The Apriori algorithm find the frequent 1-itemsets by first scanning the database , then generating the candidate frequent 2-itemsets by using the frequent 1-itemsets, and check them against the database to find the frequent 2-itemsets. This process is iterates until it can not generate any frequent k-itemsets for some k. The Apriori algorithm has been extended by several extension for improving efficiency and scalability. The most important improvement of the Apriori which introduced new technique or method such as partitioning technique [16], hashing technique [14], sampling approach [17], dynamic itemset counting [5], use the Apriori principle approach as well. In many cases, the Apriori approach showed a good performance to reduce the size of candidate sets. However, in condition with a large number of frequent patterns or low minimum support thresholds, it almost suffers inherently from two problems; multiple database scans that are costly and generating lots of candidates [9].

231

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No.3, 2009

Han et al. [8] proposed frequent pattern tree or FP-Tree as a prefix-based tree structure, and an algorithm called FP-growth. The FP-Tree stores only the frequent items in a frequencydescending order. The highly compact nature of FP-tree enhances the performance of the FP-growth. The FP-Tree construction requires two data scans. The FP-growth unlike the Apriori algorithm mines the complete set of frequent patterns without candidate generation. There have been introduced many extensions based on the FP-Tree approach such as depthfirst generation by Agarwal et al. [1], H-Mine by Pei et al. [15], array-based implementation of prefix-tree-structure by Grahne and Zhu [7] and CanTree by leung et al [10]. The experimental results showed that FP-Tree and almost all its extensions have a high compactness rate for dense data set. However, they need a large amount of memory for sparse data set where probability for sharing common paths is low [10, 12]. The presentation of data which will be mined is an essential consideration in almost all algorithms. The mining algorithms can be classified according to two horizontal and vertical database layouts. Both the Apriori and FP-growth methods use horizontal data format (i.e., {TID: itemset}) to mine frequent patterns. Zaki [18] proposed Eclat algorithm or Equivalence CLASS Transformation by using the vertical data format (i.e., {item: TID_set}). The Eclat uses the lattice theory to represent the database items. The results showed that Eclat outperforms Apriori significantly. However, it needs an additional conversion step. This is because most databases use a horizontal format. Moreover, it uses a Boolean power set lattice that needs to much space to store the labels and tid-lists. Consequently, there have been introduced some efficient algorithms based on vertical layout. The Flex [11] is a lexicographic tree designed in vertical layout to store pattern X and list of transaction identifier where pattern X appears. Its structure is restricted test-and-generation instead of Apriori-like is restricted generation-and-test. Thus nodes generated are certainly frequent. The Flex tree is constructed in depth-first fashion. The experimental results showed the Flex is an efficient algorithm to find long and maximal frequent patterns. However, it needs a large amount of memory especially to store the list of transaction identifier. III. PROPOSED METHOD

technique, an efficient tree structure called Prime-based encoded and Compressed Tree or PC_Tree and mining algorithm PC_Miner. The salient difference is that mining process makes use of the candidate head set and its properties to reduce the number of candidate sets and comparisons. In fact the PC_Miner algorithm prune the search space by using the candidate head set and it finds the most promising candidate set efficiently. This section is followed by reviewing of the data transformation technique and the PC_Tree. Then the candidate head set and its properties are explained to show how they can use in the PC_Miner algorithm to reduce the number of candidate sets. A. Data Transformation Technique As shown in Fig 1 the data transformation is an essential process in data preprocessing step which can reduce the size of database. Obviously, reducing of the size of database can enhance the performance of mining algorithms. Our method uses a prime-based data transformation technique to reduce the size of transaction database. It transforms each transaction into a positive integer called Transaction Value (TV) during of the PC_Tree construction as follows: Given transaction T = (tid, X) where tid is the transaction-id and X = {ij … ik} is the transaction-items or pattern X. While the PC_Tree algorithm scans transaction T, the transformer procedure considers a prime number pr for each item ir in pattern X, and then TVtid is computed by Equation 1 where T= (tid, X), X = {ij … ik} and ir is presented by pr.

TV tid = ∏ pr
j

k

(1)

The data transformation technique utilizes Equation 1 based on simple following definitions: “A positive integer N can be expressed by unique product
m m p1 1 p2 2 K pmr where pi is prime number, r p1 p p2 p L pr and mi is a positive integer, called the multiplicity of pi ” [6].

N

=

Based on related works’ results, reducing the number of candidate sets and comparisons (to count the support) are two effective way to enhance the performance of mining process. They showed using the Apriori principle can reduce the number of candidate sets and well-organized tree structure such as FP-Tree which captures the content of the transaction database reduces the number of comparisons of the support counting. Therefore, in this research, we aim to use both the Apriori principle and well-organized tree structure in our proposed method. We proposed a method by using a simple and effective tree structure for maximal frequent pattern mining which can capture all content of the transaction database [13]. This research proposes an improvement of previous version using a new efficient candidate set called candidate head set based on the Apriori principle to reduce the number of candidate sets. The proposed method is also based on prime number characteristics including a data transformation

For example, N = 1800=23*32*52. Fundamentally, there is no duplicated item in transaction T. Hence we restrict the multiplicity only to mi = 1 without losing any significant information. Therefore N can be produced by P1 P 2 K P r . To facilitate the transformation process used in our method, let’s examine it through an example. Let item set L= {A, B, C, D, E, F} and the transaction database, DB, be the first two columns of Table 1 with eight transactions. The fourth column of Table 1 shows TVtid transformed for all transactions. Generally the average length of items used in the benchmark datasets is smaller than those in the real applications. For example customer purchase transaction Tc= (3, {55123450, 55123452, 55123458}) from a market can be presented by third transaction T= (3, {A, B, E}) in Table 1. Although, the length of items in transaction Tc is bigger than T but both T and Tc can be transformed into the same TV 66 by using our data transformation technique. Hence it is an item-length

232

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

independent transformation technique. The experiments showed that by applying this data transformation technique, the size of real transaction databases can be reduced more than half [13].
TABLE I.
TRANSACTION DATABASE DB AND ITS TRANSACTION VALUES

TV ( n j ) | R j = ( root , n0 , n1 ,..., n j ) where 0 ≤ j ≤ i . It means

any TV stored in nodes of the sub descendant R j = (root , n0 , n1,..., n j ) can be divided by TV stored in node nj.

Definition 3: Sub descendant Ri = ( root , n0 , n1,..., ni ) is a
descendant if ni is a leaf.

TID
1 2 3 4 5 6 7 8

Items
A, B, C, D, E A, B, C, D, F A, B, E A, C, D, E C, D, F A, C, D, F A, C, D C, D, F

Transformed
2, 3, 5, 7, 11 2, 3, 5, 7, 13 2, 3, 11 2, 5, 7, 11 5, 7, 13 2, 5, 7, 13 2, 5, 7 5, 7, 13

TV
2310 2730 66 770 455 910 70 455

Definition 4: In a PC_Tree, If TV (nr) = TV (ns) then r = s . The insertion procedure increases local - count field of node nr
by 1 if the current TV is equal with TV of nr. Based on definition 1-4, the PC_Tree can be constructed by algorithm 1 as follow:

Algorithm 1 (PC_Tree construction). Input: A transaction database DB. Output: The PC_Tree of DB. Method: The PC_Tree is constructed as follows.
1. Create the root of the PC_Tree and label it as “null”. 2. For each transaction T in DB, insert T into the PC_Tree as follow: 2.1. Scan T from input file, transform T into its TV and update the item frequency table. 2.2. Add TV in the tree as follows. 2.2.1. If TV can be an element of existent descendant R then insert TV as follows. If there is a node in R with same TV then increase its local-count and global-count fields and its children’s global-count field by 1; else create a new node, with its local-count field initialized to 1. Link the new node to its parent and children. Increase its children’s global-count field by 1 and set its global-count field by summation value of its local-count field and values of its parent’s global-count field. 2.2.2. Else; the TV cannot be an element of existent descendants. Create a new node, with its local-count and global-count initialized to 1. Link the new node to the root as its parent and to its children. Increase its children’s global-count field by 1. Figure 3 shows the PC_Tree constructed for transaction database shown in Table 1. Each node presents a TV or pattern followed by two numbers after “:” to indicate the local-count and global-count respectively. There are several important properties of PC_Tree that can be derived from the PC_Tree construction procedure.

B. PC_Tree construction There have been introduced several methods to reduce the complexity of frequent pattern mining process using wellorganized tree structure. Thus the tree structures have been considered as a basic structure in previous data mining research [8, 10-13]. Recently, we introduced a novel tree structure called Prime-based encoded and Compressed Tree or PC_Tree [13].It is very simple but still powerful to capture the content of transaction database efficiently. Unlike the previous methods, the PC_Tree is based on prime number characteristics. Moreover PC_Tree has some nice properties which used to prune the search space during of mining process. Let’s review the PC_Tree. A PC_Tree consists of one root labeled as “null” and some nodes that form sub trees as children of the root. The node structure consists of several fields: value, local-count, globalcount, status and link. The value field stores the TV made by the data transformation technique during of insertion procedure in the PC_Tree construction algorithm. In fact, the value field registers which transaction this node represents. The localcount field registers the number of individual transaction represented by its node in the whole of transaction database. It is set to 1 during inserting current TV in a new node and or if there is a node with same TV then the value of its local-count field is increased by 1. Hence there is no duplicated TV in the PC_Tree. The global-count field registers frequency of its TV in its sub tree (descendant) to use in the support computing function of its TV. The status field is to keep tracking of traversing which is changed from 0 to 1 when a node visited in the traversing procedure. The link field is to form sub trees. The PC_Tree construction algorithm forms a PC_Tree by inserting TV(s) based on definitions below: Definition 1: TV of the root is assumed null and can be divided by all TVs. Definition 2: Sub tree Ri = ( root , n0 , n1,..., ni ) is a sub
descendant if only if

Property 1: In the PC_Tree, all nodes are arranged according to TV-descending order. Property 2: Important procedures used in the PC_Tree algorithm are almost done only by two simple mathematic operations product and division. Obviously using mathematic operations will enhance the performance instead of string operations.

233

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

H = {hr | hr is candidate head of Rr , 0 ≤ r ≤ j such that if hr , hs ∈ H and hr ⊆ hs ⇒ hr = hs } is candidate head set of the

PC_Tree. Based on above definitions, the PC_Tree has some interesting properties which will facilitate frequent-pattern mining. Given σ as min_sup and pattern P and Q have been represented by TV (P) and TV (Q) in descendant R respectively.

Property 3: P , Q ∈ R, sup( P ) p sup(Q ) if and only if

TV (Q ) | TV ( P ) (i.e. TV (P) can be divided by TV (Q)).
Property 4: sup (P) ≥ σ and sup (P’s parents in its descendants) p σ if and only if P is a maximal frequent pattern.
Now, this is possible to use the candidate head set to find a smaller candidate set from which the maximal frequent patterns can be derived. Consequently all frequent patterns can be generated by the maximal frequent pattern set. Therefore, based on the candidate head set definition and above properties, we have the following algorithm for frequent pattern mining using PC_Tree.

Figure 3. Step by step PC_Tree construction

C. PC_Miner algorithm using candidate head set In our previous work [13], the PC_Tree was mined by the PC_Miner algorithm in a top-down traversing fashion. It finds the maximal frequent patterns as the smallest representative set from which all frequent patterns can be derived. The PC_Miner makes use from superset and subset pruning to enhance the performance of maximal frequent pattern mining. The weakness of PC_Miner is that it considers all sub trees or descendants with same possibility during of traversing. To solve this weakness, the PC_Miner is improved by an efficient heuristic called candidate head set denote H to reduce the effective branch factor of the PC_Tree as follows. Without considering the root, let Rr = ( n0 , n1 ,..., ni ) be a descendant in the PC_Tree, the node n0 as the first node of the descendant Rr that registers the biggest TV in Rr is called head of the descendant Rr. In other word the immediate children of the root make the head set.

Algorithm 2 (PC_Miner: Mining frequent patterns from the PC_Tree by candidate head set) Input: A transaction database DB, represented by PC_Tree and minimum support threshold σ. Output: The complete set of frequent Patterns of DB. Method: The PC_Miner mines frequent patterns as follows. 1: Make the candidate head set H by using the item frequency table. 2: Let kmax denote the maximum size of itemsets in H. 3: F = F = { f | sup( f ) ≥ min_ sup, | f |= 1}.
1

Definition 6 – candidate head: the positive integer h = p1 K p which made by product prime j numbers p1 K p j is the candidate head of the descendant Rr if
only if

4: k = k

max

5: for k downto 2 do 6: H = { f | f ∈ H , | f |= k }.
k

h

is

the

largest

subset

of

sup( p ) ≥ min_ sup, 1 ≤ k ≤ j. k

n0

and

7:

for each f ∈ H and f ∉ F do
k

According to the data transformation technique used in the PC_Tree construction, TV (n0) is a product of prime numbers p1 K pk which their frequency register in the item frequency table(see Table 2) by step 2.1 of the algorithm1. Using this table and the minimum support threshold, infrequent prime numbers can be removed from TV (n0) as head of descendant Rr to make its candidate head.

8: if Sup ( f ) ≥ min_sup // property 4 and corollary 1 // 9: F=F U all subsets of f. 10: else if k>2 then 11: add all (k-1)-subsets of f to H. 12: end if 13: end for 14: end for The correctness and completeness of the process in the PC_Miner algorithm should be justified. This is accomplished by first introducing important lemma and corollary as follow:

Definition 7 – candidate head set: Let R = ( R0 ,K , R j ) be the
descendant set of the PC_Tree,

234

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
k

Lemma 1: Let H =

UH
i =1

i

consist of all k-candidate head

sets, then the complete set of the frequent itemsets can be generated by H.

Rationale: Let F be the frequent items in DB, n be head of the descendant R and h (n) is the set of frequent items in n, i.e., h (n) = n ∩ F. According to candidate head set definition, h (n) is the candidate head of the descendant R and belongs to H. In other word the candidate head set H used in the PC_Miner algorithm includes all frequent items. The relationship among the candidate head set H, frequent and maximal frequent itemsets are shown in Figure 4.

presented in Table1. The PC_Miner uses the frequency of items computed in step 2.1 of the PC_Tree algorithm to find the candidate head set. Table 2 shows the frequency of items of DB. Consequently, items B and E are infrequent when min_sup=4. Therefore, the PC_Miner starts mining with candidate head set H includes only 4-candidate itemset ACDF and the PC_Miner only examined 6 candidate itemsets {ACDF, ACD, ACF, ADF, CDF, AF}. Above definitions and PC_Miner algorithm show that using the candidate head set can reduce the number of candidate itemsets. Moreover the experimental results support the accuracy and efficiency of our method. IV. EXPERIMENTAL RESULTS

Figure 4. Relationship among candidate head set, frequent and maximal frequent itemsets

Corollary 1: The complete set of the frequent itemsets F can
k

be generated by (

U H )U F
i=2 i

1

where F1 is set of 1-length

In this section, we evaluate the accuracy and performance of our method by several experiments. All experiments were performed in a time-sharing environment in a 2.4 GHz PC with 2 GB memory. We used several sparse and dense datasets which used in previous works as benchmark datasets. The synthetic sparse datasets are generated by the program developed at IBM Almaden Research Center [3] and real dense datasets are download from UC Irvine Machine Learning Repository [4]. The results reported in figures were computed by the average of multiple runs. According to the space limitation and the problem specifications, only the results of experiments by using synthetic sparse dataset T10I5D100K and real dense dataset mushroom which are the most popular benchmark datasets in this field are presented in this paper. The number of transactions, the average transaction length, the number of items and the average frequent pattern length of T10I5D100k are set to 100k, 10, 1000 and 4 respectively. The mushroom dataset consists of the characteristics of various mushroom species. The number of records, the number of items and the average record length are set to 8124, 119 and 23 respectively. The first experiment is to proof the accuracy of the proposed method. However the correctness and completeness of the process in the PC_Miner algorithm was justified in previous section. The number of frequent patterns mined by the PC_Miner algorithm versus support using T10I5D100k and mushroom datasets is shown in Figures 5 and 6 respectively. We compared all frequent patterns mined by the proposed method using the candidate head set with those mined by the Apriori algorithm in several datasets. They were exactly equal. The second experiment evaluates the number of the candidate sets generated by the PC_Miner and Apriori which is one of the most efficient algorithms in term of candidacy reduction. Figure 7 shows the number of candidate sets generated by PC_Miner and Apriori algorithms versus different minimum support thresholds over dataset T10I5D100k. The efficiency of using the candidate head set in the PC_Miner algorithm is verified by third experiment which compares the run time versus the support. Figure 8 and 9 present the efficiency of the PC_Miner algorithm using the candidate head set versus the Apriori algorithm in T10I5D100k and mushroom datasets over different minimum support thresholds respectively.

frequent itemsets which can be derived from item frequency table made in the PC_Tree algorithm. The number of the candidate head sets is very smaller than the frequent and even maximal frequent itemsets. However, Figure 4 shows that the candidate head set is a superset of frequent and maximal itemsets. To illustrate let’s examine the mining process. In the PC_Miner algorithm, the line 3 makes 1-frequent itemsets by using the item frequency table which made during the tree construction process. In outer for-loop between lines 5-14, all k-candidate heads (k>2) are made and investigated respectively. Then during of line 7-13 all frequent itemsets are generated by using maximal frequent itemsets found in line 8 based on property 4. According to the relationship shown in Figure 4 and steps of the PC_Miner algorithm which investigate all candidate heads; therefore there is no missing frequent itemsets. The PC_Miner algorithm make use from superset infrequent pruning based on the candidate head set introduced in this research which is an effective way to eliminate some of the candidate itemsets. In fact the candidate head set definition is based on the Apriori principle which all subsets of a frequent itemset must also be frequent. For instance there are 26 − 1 = 63 possible candidate itemsets in the transaction database DB

235

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009
T10I5D100K
100000

140 120

T10I5D100k

#Frequent Patterns

10000 1000 100 10 1 5 4 3 2 1 0.5 0.1

Run time (sec)

100 80 60 40 20 0 6 5

PC_Miner Apriori

4

3

2

1

0.5

0.1 0.05

Support (%)
Figure 5. #Frequent Patterns Vs. Support

Support(%)
Figure 8. Run time Vs. Support

Mushroom
1000000 #Frequent Patterns 100000 10000 1000 100 10 1 50 40 30 20 10

Mushroom
80 70

Run time (sec)

60 50 40 30 20 10 0 50 45

PC_Miner Apriori

40

35

30

25

20

15

Support (%)
Figure 6. #Frequent Patterns Vs. Support

Support(%)
Figure 9. Run time Vs. Support

V.
T10I4D100k Candidate sets generated by PC_Miner

CONCLUSION AND FUTURE WORKS

Candidate sets generated by Apriori
100000 10000 1000 100 10 1 5 4 3 2 1 0.5 0.1

In this paper we introduced the candidate head set to reduce the number of candidate sets in mining process. Our previous method [13] was improved by using the candidate head sets to propose an efficient method for frequent pattern mining. The experimental results verified the accuracy and efficiency comparing with the Apriori algorithm which is one of the most efficient algorithms in term of candidacy reduction. Particularly, we introduced a new method based on prime number characteristics using candidate head sets to find completed frequent patterns by using maximal frequent patterns. The proposed method can be improved for incremental mining of frequent patterns where database transactions can be inserted, deleted, and/or modified incrementally. Moreover it can be improved for interactive mining of frequent patterns where minimum support threshold can be changed to find new correlation between patterns without rerunning the mining process from scratch.

Support(%)
Figure 7. The number of candidate sets Vs. Support

236

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

REFERENCES
[1]. Agarwal, R.C., C.C. Aggarwal, and V.V.V. Prasad, "A tree projection algorithm for generation of frequent item sets", Journal of Parallel and Distributed Computing, 2001. Vol. (3): p. 350-371. [2]. Agrawal, R., T. Imieli ski, and A. Swami, "Mining association rules between sets of items in large databases", ACM SIGMOD Record, 1993. Vol. (2): p. 207-216. [3]. Agrawal, R. and R. Srikant. "Fast algorithms for mining association rules", Proc. 20th Int. Conf. Very Large Data Bases, VLDB, 1994 [4]. Blake, C. and C. Merz. Uci repository of machine learning databases, university of california – irvine, irvine, ca. 1998. [5]. Brin, S., et al. "Dynamic itemset counting and implication rules for market basket data", proc. ACM-SIGMOD int. conf. management of data (SIGMOD’97), 1997 [6]. Cormen, T.T., C.E. Leiserson, and R.L. Rivest, Introduction to algorithms. 1990: MIT Press Cambridge, MA, USA. [7]. Grahne, G. and J. Zhu. "Efficiently using prefix-trees in mining frequent itemsets", Proc. ICDM’03 int. workshop on frequent itemset mining implementations (FIMI’03) [8]. Han, J., J. Pei, and Y. Yin. "Mining frequent patterns without candidate generation", Proc. ACM-SIGMOD int. conf. management of data (SIGMOD’00), 2000 [9]. Han, J., et al., "Frequent pattern mining: Current status and future directions", Data Mining and Knowledge Discovery, 2007. Vol. (1): p. 55-86. [10]. Leung, C.K.S., Q.I. Khan, and T. Hoque. "Cantree: A tree structure for efficient incremental mining of frequent patterns", Proc. ICDM 2005, 2005 [11]. Mustapha, N., et al., "Fast discovery of long patterns for association rules", International Journal of Computer Mathematics, 2003. Vol. (8): p. 967-976. [12]. Nadimi-Shahraki, M., et al. "Incremental updating of frequent pattern: Basic algorithms", Proceedings of the second International Conference on Information Systems Technology and Management ( ICISTM' 08), 2008 [13]. Nadimi-Shahraki, M., et al., PC_Tree: Prime-based and compressed tree for maximal frequent patterns mining, Advances in electrical engineering and computational science, S.-l. Ao and L. Gelman, Editors, 2009, Springer. ch. 42, p. 495-504. [14]. Park, J.S., M.S. Chen, and P.S. Yu, "An effective hash-based algorithm for mining association rules", int. conf. management of data (SIGMOD’95), 1995. Vol. (2): p. 175-186. [15]. Pei, J., et al. "Prefixspan: Mining sequential patterns efficiently by prefix-projected pattern growth", Proc. int. conf. data engineering (ICDE’01), 2001 [16]. Savasere, A., E. Omiecinski, and S. Navathe. "An efficient algorithm for mining association rules in large databases", int. conf. very large data bases (VLDB’95), 1995 [17]. Toivonen, H. "Sampling large databases for association rules", Proc. int. conf. very large data bases (VLDB’96), 1996 [18]. Zaki, M.J., "Scalable algorithms for association mining", IEEE Transactions on Knowledge and Data Engineering, 2000. Vol. (3): p. 372-390.

237

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009

Application of a Fuzzy Programming Technique to Production Planning in the Textile Industry
*I. Elamvazuthi , T. Ganesan, P. Vasant
Universiti Technologi PETRONAS Tronoh, Malaysia
*

J. F. Webb
Swinburne University of Technology Sarawak Campus, Kuching, Sarawak, Malaysia

Abstract—Many engineering optimization problems can be considered as linear programming problems where all or some of the parameters involved are linguistic in nature. These can only be quantified using fuzzy sets. The aim of this paper is to solve a fuzzy linear programming problem in which the parameters involved are fuzzy quantities with logistic membership functions. To explore the applicability of the method a numerical example is considered to determine the monthly production planning quotas and profit of a home-textile group. Keywords: fuzzy set theory, fuzzy linear programming, logistic membership function, decision making

I.

INTRODUCTION

Many problems in science and engineering have been considered from the point of view optimization. As the environment is much influenced by the disturbance of social and economic factors, the optimization approach is not always the best. This is because, under such turbulent conditions, many problems are ill-defined. Therefore, a degree-of-satisfaction approach may be better than optimization. Here, we discuss how to deal with decision making problems that are described by fuzzy linear programming (FLP) models and formulated with elements of imprecision and uncertainty. More precisely, we will study FLP models in which the parameters are known only partially to some degree of precision. Even though the information is incomplete, the model builder is able to provide realistic intervals for the parameters in these FLP models. We will demonstrate that the modeling complications can be handled with the help of some results which have been developed in fuzzy set theory. The FLP problem which we will be considering in this work is to find ways to handle fuzziness in the parameters. We will develop a FLP model in which the parameters are known with only some degree of precision. We will also show that the model can be parameterized in such a way that a satisfactory solution becomes a function of the membership values. The FLP model derived in this way is flexible and easy to handle computationally [1]. The first and most meaningful impetus towards the mathematical formalization of fuzziness was pioneered by Zadeh [2]. Its further development is in progress, with numerous attempts being made to explore the ability of fuzzy set theory to become a useful tool for adequate mathematical

analysis of real-world problems [3]. The period of development of fuzzy theory from 1965 to 1977, is often referred to as the academic phase. The outcome was a rather small number of publications of a predominantly theoretical nature by a few contributors, mainly from the academic community. At this time, not much work in the area of fuzzy decision making was reported. The period from 1978 to 1988, has been called the transformation phase during which significant advances in fuzzy set theory were made and some real-life problems were solved. In this period, some important principles in fuzzy set theory and its applications were established. However, work on fuzzy decision making was not very active, in the area of engineering applications. Some earlier work on fuzzy decision making can be found in [4] and [5]. From 1989 to the present work on fuzzy techniques has boomed . In this period, many problems concerning applications in industry and business have been tackled successfully. In the early 1990s, fuzzy techniques were used to aid the solution of some soft computing problems. The aim of soft computing is to exploit, whenever possible, the tolerance for imprecision and uncertainty in order to achieve computational tractability, robustness, and low cost, by methods that produce approximate but acceptable solutions to complex problems which often have no precise solution. Currently, fuzzy techniques are often applied in the field of decision making. Fuzzy methods have been developed in virtually all branches of decision making, including multiobjective, multi-person, and multi-stage decision making [6]. Apart from this, other research work connected to fuzzy decision making includes applications of fuzzy theory in management, business and operational research [7]. Some representative publications can be found in [8], [9], [10], [11] and [12]. Decision making is an important and much studied application of mathematical methods in various fields of human activity. In real-world situations, decisions are nearly always made on the basis of information which, at least in part, is fuzzy in nature. In some cases fuzzy information is used as an approximation to more precise information. This form of approximation can be convenient and sufficient for making good enough decisions in some situations. In other cases, fuzzy information is the only form of information available.

238

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

The first step in mathematically tackling a practical decision-making problem consists of formulating a suitable mathematical model of a system or situation. If we intend to make reasonably adequate mathematical models of situations that help practicing decision makers in searching for rational decisions, we should be able to introduce fuzziness into our models and to suggest means of processing fuzzy information. In this paper a methodology to solve an FLP problem by using a logistic membership function is considered. The rest of the paper is organized as follows. In section 2, the basic fuzzy model is defined and this is followed by a numerical example in section 3. Section 4 provides the results and discussion, and finally, concluding remarks are made in section 5. II. THE MODEL

μaij %

⎧ ⎪ ⎪1 ⎪ B ⎪ =⎨ a ⎛ aij − aij ⎞ α⎜ b a ⎟ ⎪ ⎜ aij − aij ⎟ ⎠ ⎪ 1 + Ce ⎝ ⎪ ⎪0 ⎩
III.

a if aij ≤ aij a b if aij ≤ aij ≤ aij

(4)

b if aij ≥ aij

NUMERICAL EXAMPLE

A conventional linear programming problem is defined by
Maximize Cx Subject to Ax ≤ b, x ≥ 0.

(1)

in which the components of a 1×n vector C, an m×n matrix A and an n×1 vector b are all crisp parameters and x is an ndimensional decision variable vector. The system (1) may be redefined in a fuzzy environment with the following more elaborate structure: Maximize Subject to

In this example the profit for a unit of sheet sales is around 1.05 Euro; a unit of pillow case sales is around 0.3 Euro and a unit of quilt sales is around 1.8 Euro. The firm concerned would like to sell approximately 25.000 sheet units, 40.000 pillow case units and 10.000 units quilt units. The monthly working capacity and required process time for the production of sheets, pillow cases and quilts are given in Table 1 [14]. In view of this, let us determine monthly production planning details and profit for a home-textile group. X1 presents the quantity of sheets that will be produced, X2 presents the quantity of pillow cases and X3 presents the quantity of quilts. The profit figures with logistic membership functions as given in Table I.
TABLE I. REQUIRED PROCESS TIME FOR SHEET, PILLOW CASE AND OF A QUILT [14]

∑cjxj
~
j =1

n

Departments

Required unit time(hour) Sheet Pillow case Quilt

Working hours per month

∑ aij x j ≤ bi , ~
~
j =1

n

i = 1,2 L m

(2)

Cutting Sewing Pleating Packaging

0.0033 0.056 0.0067 0.01

0.001 0.025 0.004 0.01

0.0033 0.1 0.017 0.01

208 4368 520 780

~ a b ~ a b ~ ~ All fuzzy data c j ≡ S (c j , c j ) and aij ≡ S (aij , aij ) are fuzzy variables with the following logistic membership functions [13],
⎧ ⎪ ⎪1 ⎪ B ⎪ μc j = ⎨ % ⎛ c j −ca ⎞ j ⎪ α⎜ b a ⎟ ⎜ c j −c j ⎟ ⎪ ⎝ ⎠ ⎪ 1+ Ce ⎪0 ⎩

% If we consider, around 1.05 ≡ S (1.02,1.08) , around 0.3 ≡ % % S (0.2, 0.4) , and around 1.8 ≡ S (1.7, 2.0) , then, the

mathematical model of the above problem with fuzzy objective coefficients can be described as follows.
if c j ≤ ca j if ca ≤ c j ≤ cb j j

(3)

Maximize ~ ~ ~ S (1.02,1.08) x1+ S (0.2,0.4) x 2 + S (1.7, 2.0) x 3 subject to
0 . 033 x1 + 0 . 01 x 2 + 0 .0033 x 3 ≤ 208 ; 0 . 056 x1 + 0 .25 x 2 + 0 .1 x 3 ≤ 4368 ; 0 . 0067 x1 + 0 .04 x 2 + 0 .17 x 3 ≤ 520 ; 0 . 1 x1 + 0 . 1 x 2 + 0 .01 x 3 ≤ 780 ; x1 ≥ 25000 ; x 2 ≥ 40000 ; x 3 ≥ 10000 ;

if c j ≥ cb j

(5)

239

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

and we set

B = 1, C = .001, ε = 0.2 and d = 13.8 [15].

The aspiration of the objective function is calculated by solving the following:
Maximize subject to .0033x1 + .001x2 + .0033 x3 ≤ 208; .056 x1 + .025 x2 + .1x3 ≤ 4368; .0067 x1 + .004 x2 + .017 x3 ≤ 520; .01x1 + .01x2 + .01x3 ≤ 780; x1 ≥ 25000; x2 ≥ 40000; x3 ≥ 10000;
Figure 1. 3D plot for iterations M=748.

1.08 x 1 +0.4 x2 + 2.0 x3

(6)

which gives the optimal value of the objective function as 67203.88 for x1 = 29126.21, x2 =35000.00 and x3 =10873.79 [15]. With the help of the program LINGO version 10.0 we obtain the following results [15]:
λ = 0.5323011, x1 = 27766.99, x2 = 40000.00, x3 = 10233.01, η = 0.4911863

Fig. 2 shows the 3D outcome for M = 749 iterations and various alpha values with respect to G. The optimum values for the objective function as per this figure are 86,691.8 (maximum) and 86,639.5 (minimum).

Therefore, to achieve maximum profit the home-textile group should plan for a monthly production of 27766.99 sheet units, 40000 pillow case units and 102333.01 quilt units. This plan gives an overall satisfaction of 0.5323011. The decision making method may be improved further by adopting a recursive iteration methodology. IV. RESULTS AND DISCUSSION

Figure 2. 3D plot for iterations M = 749.

The numerical example is solved by using a recursive method for various iterations. This was carried out using the C++ programming language on a personal computer with a dual core processor running at 2 GHz [16]–[17]. Fig. 1 shows the 3D outcome of the iterations with M = 748 for various alpha values with respect to the objective function G. The values of α 1 and α 2 vary from 0 to 1. The optimum values for the objective function as per Fig. 1 are 86,807.7 (maximum) and 86,755.4 (minimum).

Fig. 3 shows the 3D outcome for M = 750 iterations and various alpha values with respect to G. The optimum values for the objective function as per this figure are 86,576.2 (maximum) and 86,524.0 (minimum).

Figure 3. 3D plot for iterations M=750.

240

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

Fig. 4 shows the 3D outcome for M = 751 iterations and various alpha values with respect to G. The optimum value for the objective function as per this figure are 86,440.7 (maximum) and 86,408.0 (minimum).

Figure 6. Decision variable, X1 versus M iterations.

Figure 4. 3D plot for iterations M=751.

Fig. 5 shows the linear approximation for G with respect to iterations 748 to 751. It can be seen that as the iterations are increased, the values of the objective function decrease. The percentage error is minimum at iteration, M = 748; however, after that it increases until it peaks at M = 750; thereafter, the percentage error decreases again to a level lower than that at M = 748. This shows that the maximum number of iterations that can be used for similar cases in the future can be limited to M = 750.
Figure 7. Decision variable, X2 versus M iterations.

Figure 8. Decision variable, X2 versus M iterations. Figure 5. Objective Function (G) versus iterations

Figs. 6, 7 and 8 show the linear approximation for the decision variables x1, x2 and x3 with respect to the number of iterations. It can be observed that x1, x2 and x3 decrease as the iterations are increased from M = 748 to M = 751.

Table II presents results that involve

α1 , α 2

and

α3

with M

= 748 for G, x1, x2 and x3. Other results for M = 749 to 751 are given in the appendix.

241

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

TABLE II ALPHA, OBJECTIVE FUNCTION AND DECISION VARIABLES FOR M=748
α1
1 1 0.5 0.5 0.3333 0.3333 0.25 0.25 0.2 0.2 0.1667 0.1667 0.1429 0.1429 0.125 0.125 0.1111 0.1111

α2
1 0.5 1 0.5 1 0.5 1 0.5 1 0.5 1 0.5 1 0.5 1 0.5 1 0.5

α3
*all all all all all all all all all all all all all all all all all all

G
86755.4 86780.3 86767.5 86792.4 86770.9 86795.9 86772.5 86797.5 86773.5 86798.4 86774.1 86799.0 86774.5 86799.4 86774.8 86799.8 86775.1 86800.0

x1
33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5 33422.5

x2
53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9 53475.9

x3
13369 13369 13369 13369 13369 13369 13369 13369 13369 13369 13369 13369 13369 13369 13369 13369 13369 13369

From Table IV, the optimum value for the objective function using the proposed method outweighs the results obtained in [14] and [15]. It can be deduced that the recursive iteration method proposed here is an efficient and effective way to solve our example fuzzy problem of production planning in the textile industry. V. CONCLUSION

This paper has discussed the use of fuzzy linear programming for solving a production planning problem in the textile industry. It can be concluded that the recursive method introduced is a promising method for solving such problems. The modified s-curve membership function provides various uncertainty levels which are very useful in the decision making process. In this paper, only a single s-curve membership function was considered. In the future, various other membership functions will be considered. Apart from providing an optimum solution for the objective functions, the proposed method ensures high productivity. In this regard, there is a good opportunity for developing an interactive selforganized decision making method by using hybrid soft computing techniques. ACKNOWLEDGMENT The authors would like to thank Universiti Teknologi PETRONAS and Swinburne University of Technology Sarawak Campus for supporting this work. REFERENCES
[1] [2] [3] Delgado, M., Verdegay, J.L. and Vila, M. A. 1989. A General Model For Fuzzy Linear Programming. Fuzzy Sets and Systems 29 : 21-29. Zadeh, L. A. 1965. Fuzzy sets. Information and Control 8 : 338-353. Orlovsky, S. A. 1980. On Formalization Of A General Fuzzy Mathematical Programming Problem. Fuzzy Sets and Systems 3 : 311321. Kickert, W.J. 1978. Fuzzy Theories on Decision-Making: Frontiers in Systems Research. Leiden, The Netherlands : Martinus Nijhoff . Zimmermann, H.J. 1987. Fuzzy Sets, Decision Making, and Experts Systems. Boston: Kluwer. Tamiz, M. 1996. Multi-objective programming and goal programming: theories and applications. Germany: Springer-Verlag. Zimmermann, H.J. 1991. Fuzzy Set Theory-and Its Applications, (2nd rev. ed.). Boston: Kluwer. Ross, T. J. 1995. Fuzzy Logic with Engineering Applications, New York: McGraw- Hill. Klir, G. J., and Yuan, B. 1995. Fuzzy Sets and Fuzz