Documents
Resources
Learning Center
Upload
Plans & pricing Sign in
Sign Out

Sample Resume Head

VIEWS: 7 PAGES: 32

Sample Resume Head document sample

More Info
  • pg 1
									Automated Detection of
 Deception and Intent
          Judee Burgoon, Ed.D.
Center for the Management of Information
           University of Arizona



                 19MAR04
   Collaborative Partners
DETECTING
DECEPTION IN
THE MILITARY
INFOSPHERE




               Funded by Department of Defense
                  Center for the Management of Information,
                  University of Arizona
AUTOMATED
INTENT
DETECTION         Center for Computational Bioengineering,
                  Imaging and Modeling, Rutgers University

                  Funded by Department of Homeland
                  Security
Deception and Intent Defined
    Deception is a message knowingly transmitted with
     the intent to foster false beliefs or conclusions.
    Hostile intent refers to plans to conduct criminal or
     terrorist activity
    Intent is inferred from:
         suspicious behavior
         overt hostility
         deception
Relationship of Deception to
           Intent
Many Ways To Deceive
   Lies              Deflections
   Fabrications      Evasions
   Concealments      Equivocation
   Omissions         Exaggerations
   Misdirection      Camouflage
   Bluffs            Strategic ambiguity
   Fakery            Hoaxes
   Mimicry           Charades
   Tall tales        Imposters
   White lies
America under Attack!
Statement of the Problem
 Humans have very poor ability to detect
deceit and hostile intent.
    True of experts as well as untrained
   individuals
    Accuracy rates of 40-60%--about the
   same as flipping a coin




 Reliance on new communication
technologies--text, audio, video--may make us
more vulnerable to deceit.
In the Headlines
           Questions
 Are there reliable indicators of:
    deceit?

    intent to engage in hostile actions?

 Can detection be automated to
augment human abilities?
 Does mode of communication make a
difference?
        The Objective:
Reducing False Alarms & Misses
    Sample Deception Indicators
   Arousal
      Higher pitch, faster tempo

   Emotion
      Absence of emotional language, false smiles

   Cognitive effort
      Delays in responding, nonfluent speech

   Memory
      Fewer details, briefer messages

   Strategic communication
      Controlled movement, increasing involvement
                Our Experiments
No.   Experiment                                Subjects
1.    Desert Survival 1                         60         16
2.    Desert Survival 2                         52         exper-
3.    Mock Theft Pilot Study                    30
4.    Mock Theft Main Study                     207
                                                           iments,
5.    Mock Theft Observer Study                 51         2136
6.    Bunker Buster                             50         subjects,
7.    Scud Hunt                                 60
                                                           in 2.5
8.    StrikeCom                                 622
9.    ARI Experiment                            60
                                                           years
10.   Resume Pilot Study                        25
11.   Resume Main Study                         156
12.   Multiple Receiver Study                   60
13.   Agent99 Trainer Pilot Studies (1 and 2)   120
14.   Agent99 Trainer AFB Main Study 1          190
15.   Agent99 Trainer AFB Main Study 2          160
16.   Air Force ROTC StrikeCom Air Operations   200
           Typical Experiment:
               Mock Theft
   Task
      half of participants steal wallet from
      classroom, other half are innocents
      all are interviewed by trained and/or untrained
      interviewers
   Mode of interaction
      face-to-face, text, audio, video

   Outcomes
      accuracy in detecting truth and deception

      judged credibility

      coding of verbal and nonverbal behavior
          Sample Results
 Deceivers create                  Message Length (# of Words) by

 longer messages              240
                                    Deception and Modality

 under text than              220

                              200
 FtF.                         180


                 Mean Words   160


                              140                                         DECEPTION

                              120                                           thief

                              100                                           innocent
                                         face-to-face              text


                                                        MODALITY
           Implications

   Text-based deception allows for
    planning, rehearsal, editing.
   Deceivers can use text messages to
    their advantage.
        Questions
 Are there reliable text-based
 indicators of deceit or hostile
 intent?
 Can these be automated to
 overcome deceivers’ advantages?
Sample Text-based Cues We Analyzed
 Category Cues                  Category      Cues

 Quantity    # of words         Diversity &   Lexical
             # of verbs                       diversity
             # of sentences     Specificity   Spatio-temporal
                                              details
 Complex-    Avg word length    Affect        Imagery
             Avg sentence                     Activation
 ity         length                           Pleasantness
             Punctuation
 Uncertain   Modal verbs        Expressive-   Emotiveness
             Modifiers                        index
 -ty                            ness
 Verbal      Passive voice      Informality   Typographical
             Group and you                    errors
 Nonim-      references
 mediacy     Self- references
          Sample Results from
          Automated Analysis
   Deceivers use different language than truth tellers.
     Deceivers—more

        quantity

        uncertainty

        references to others
        informality

     Truthtellers—more

        diversity
        complexity

        positive affect

        references to self
        Automating Analysis:
          Agent99 Parser
   Find cues in text




   Submit to data mining tool
              Decision Tree Analysis
                           Modifier Quantity: 51

                                              Temporal Immediacy: 0.0

  Sensory Ratio: 0.0325


                                        Verb_Quantity: 63




                                             Modifier Quantity: 0.0325
Modal Verb Ratio: 0.2698




                                                         True: it is
                                                         deceptive
Accuracy in Detecting Deceit

                  Judged Truth Judged Deceit
Actual
Truth                  61-91%               9-29%

Actual
Deceit                 12-35%              61-88%

 Note: Preliminary findings from Mock Theft, from
 transcribed face-to-face sessions
           Implications
    Linguistic and content features
    together can reliably identify
    deceptive or suspicious
    messages.
    Text analysis can be
    successfully automated.
          Questions

 Can hostile intent be mapped to
behavior?
 Are there reliable video-based
indicators of deceit and intent?
 Are the indicators open to
automation?
The Mapping Problem
        Approach to Analysis
    Four data sets:
     Pre-polygraph interviews from actual
       investigations
     Mock theft experiment
       Two states: innocent (truthful),
         deceptive (guilty)
     Actors in airport/screening location
       scenarios
       Three states: relaxed, agitated
         (nervous), overcontrolled
     Actors showing normal behavior to train
       neural networks
Intent Recognition from Video
   Track and estimate human movement
    including:
      Head
      Facial & Head Features
      Hands

      Body

      Legs
   Tracking techniques:
      Physics-based tracking of face and hands

      Statistical model-based motion
       estimation
Skin Color Tracker: Face & Hands
Sample Results from Human Coders

 “Thieves” use                                Counts of Head Movements,

 fewer head                               Adaptors and Illustrator Gestures
                                         20


 movements and                           18



 gestures, more                          16
              Estimated Marginal Means



                                         14


 self-touching                           12                                VARIABLE


 than “innocents.”                       10


                                         8
                                                                              Head Movements

                                                                              Adaptor Gestures

                                         6                                    Illustrator Gestures
                                                      guilty    innocent


                                              GUILT
                           Sample Patterns: Actors
                                                                                    Head pos. L. hand pos.     R. hand pos.
Head pos. L. hand pos.     R. hand pos.
                                          Head pos. L. hand pos.     R. hand pos.




                                                                                    Head vel.   L. hand vel.   R. hand vel.
Head vel.   L. hand vel.   R. hand vel.
                                          Head vel.   L. hand vel.   R. hand vel.




            nervous                                   controlled                                  relaxed
 Sample Patterns: Mock Thieves
Head pos.   L. hand pos.   R. hand pos.    Head pos.      L. hand pos.   R. hand pos.




Head vel.   L. hand vel.    R. hand vel.
                                           Head vel.     L. hand vel.    R. hand vel.




       Nervous (lying)                                 Relaxed (not lying)
Sample Results: Scores differ
among relaxed, agitated, and
overcontrolled suspects
                   Summary
   Humans are fallible in detecting deception and
    hostile intent
   Automated detection tools to augment human
    judgment can greatly increase detection accuracy
   Verbal and nonverbal behaviors have been
    identified that:
      Can be automated

      Together significantly improve detection
       accuracy
   More research under a variety of contexts will
    determine which indicators and systems are the
    most reliable

								
To top