Get documents about "Integration Information
Document Sample
Chapter 11
Integration Information
Instructor: Prof. G. Bebis
Represented by Reza
Fall 2005
Outlines
• Introduction
• Integration methods
• Decision level integration
– Boolean combination
– Binning and filtering
– Dynamic authentication protocols
• Score level integration
– Normal distributions
– Degenerate cases
– From threshold to boundaries
• Alternative
Introduction
• For many application there are non-biometric sources of
information that can be used in person authentication.
• Using a single biometric is not sufficiently secure or does
not provide sufficient coverage of the user population.
• Question: How do we integrate multiple biometric
sources of information to make the application more
accurate and more secure?
Integration methods
• There are many different methods that can be used to expand a
biometric system:
– Multiple biometrics: Face image and voiceprint
– Multiple location: Left and right iris
– Multiple sensing: Three tries of index finger
– Multiple sensors: Ultrasonic and optical sensing
– Multiple matchers: Minutia and correlation fingerprint matcher
– Multiple token: Adding possession and/or knowledge
• An example using multiple biometrics:
Integration methods
• Regardless of the methods, there are tow basic approaches to combination
information from different sources:
– Tightly coupled integration:
• A strong interaction among the input measurements and integration schemes.
– Loosely coupled integration:
• There is no interaction among inputs and integration occurs at the output.
A tightly coupled system A loosely coupled system
Integration methods
• Loosely coupled integration systems
advantages:
– Simpler to implement
– More feasible in commonly confronted integration
scenario
• One could try to integrate the biometrics at tow
levels:
– Decision level
– Score level
Decision level integration
• Decision level integration (fusion) is typically concerned
with multiple matchers method.
• Methods:
– Boolean combination
– Binning and filtering
– Dynamic authentication protocols
• Note that in this context, the matchers are considered as
block box and each output is simply a “Yes/No”.
Boolean combination
• The AND rule and the OR rule:
– the most prevalent rules for multiple biometrics combination in
practical systems.
– Used in both Identification and Verification protocol.
Boolean combination
• OR
– Improve convenience (lower the FRR)
• AND
– Improve security (lower FAR)
Binning and Filtering
• The capability of a biometric 1:many search of a
database is a prerequisite.
• Performing N biometric 1:1 matches :
– Advantage:
• Simple way.
– Disadvantage: (When N, the size of the database, becomes
large)
• High computational cost
• High False Positive Rate and large candidate list (FPR N)
Filtering
• Constraining the search with parametric (non-
biometric) data is called “filtering”.
– Filtering down the search of N by, say, a subject’s
surname
– Filtering is an authentication protocol
(Possession, Biometric)=(P, B)=(name, B)
• Search the database N of enrolled subjects by surname
• Search the subjects with matching surname P by matching
input sample biometric B
Binning
• Constraining the search with additional biometric
data is referred to as “binning”.
– The best known instance is first classify the type of
fingerprint and then match the minutia of fingerprint
– Authentication protocol:
B, B
• Select those subjects in database N whose biometric
template matches biometric B’
• Match the input biometric template B with the templates of
those remaining subjects to find those subjects in N with both
matching B’ and B
Binning and Filtering
• Penetration rate or filter rate:
E (# of times B is matched)
Ppr
N
• Binning error rate Pbe is the percentage of subjects in the
data base that are missed classify
• Possession token can be added to an existing
authentication protocol:
– Negative identification
• Decrease the chance of False Positive
• Increase the probability of False Negative dramatically
– Positive identification
• Does not decrease the biometric verification error rates
Dynamic authentication protocols
• One dynamic protocol for speaker verification is the idea
of conversational biometrics
• Conversational biometrics does a biometric match
between the speech sample B and the voiceprint Bm and
a knowledge match between the collected responses
through speech recognition and the knowledge K m
Score level integration
• Loosely coupled integration at score level integration
• Assumption:
– Scores are normalized sa , sb 0,1
– B is the measurements of biometric a and b
Score level integration
• Principle cases:
– The scores ( sa , sb )are monotonically related to the likelihood P(dm B)
– The scores ( sa , sb )are related to the likelihood P(dm B) in more
complex fashion
• Given ground truth marked data, it is possible to
determine a function which relates ( sa , sb ) to P(dm B)
– Using ground truth data to estimate joint probability density
function P(sa , sb Ho ) and P(sa , sb H a )
– Example: Prabhakar and Jain estimate the conditional densities
using non-parametric estimation method
P( sa , sb H o )
P(d m B)
P( sa , sb H a )
Normal distribution
• Approximation of the curve G
with a linear function
G ( sa , sb ) sa (1 ) sb T
• This approximation is correct
only when both and s a are s b
normally distributed.
– If we assume that s aand s b are
independent, the above
equation becomes:
sa sb
G ( s a , sb )
2
a b2
Normal distribution
• Problems:
– The covariance matrix is assumed to be diagonal.
This is good if disparate biometrics are used.
– Modeling match scores with Gaussian is not realistic
• Solution: Use of a Gaussian model for the probability
distribution of the distance between two biometric templates
– Simple example: Dist ( B, Bm ) 1 s( B, Bm )
2
1 d Em
1 2 m
P(d ) e
2
Em E ( Dist ( B, Bm )) m ( Dist ( B, Bm ))
Degenerate cases
• If a b then
sa
G ( sa ,0) T
2
a
• If b a then
sb
G (0, sb ) T
2
b
• It means biometric a does
not contribute much to
class separation.
From thresholds to boundaries
• Question: Is there any way to
estimate some decision
boundary?
• Answer: Drive estimates of the
match and mismatch score
cumulative distributions from
training data and determine
operating point T that satisfy
the design criteria.
• Assume that we have the
cumulative match score
distribution F ( sa , sb ) FAR( sa , sb )
G( sa , sb )
1 FAR( sa , sb ) 1 T
and the cumulative mismatch
score distribution G ( sa , sb )
From thresholds to boundaries
• FR estimates are now the
value of
F ( sa , sb ) along the
curve G(sa , sb ) T
• The FRR are given by
k k
F ( sa , sb ) , k 1,..., K
• A multi biometric system
should not just be
associated with one FAR
and one FRR but with one
FAR and a sequence of FRR k
Alternative
• The OR and AND rules for combining two decisions are
the only way
• When more than two decision are need to be combined
the choice of fusion is not merely limited to applying an
overall OR and AND rules to all decision
• Another method is voting
• Kittler and Alkoot have shown that score combination
strategy are superior when the individual matcher scores
follow Gaussian distribution and voting is superior when
scores distribution are heavy tailed
