2008 Projects

An Acquisition Platform for Non‐cooperative, Long Range Ocular Biometrics
Reza Derakhshani (UMKC), Plamen Doynov (UMKC) and Besma Abidi (UTK)

Automatic High Resolution Palmprint Matching System
Anil K. Jain (MSU) and Arun Ross (WVU)

Culturally Specific Credibility Classification Using Layered Voice Analysis
Judee Burgoon, Aaron Elkins, Douglas Derrick, Josh Hottenstein (U of A) and Dale Tunnell (Forensitec)

Establishing Chain of Evidence in Biometric Systems
Bojan Cukic, Arun Ross, Nathan Kalka and Nick Bartlow (WVU)

Improving the Identification of Fraud by Adding Word Sense Disambiguation to Linguistic Credibility Assessment/Enhancing Fraud Detection by Building Lexicons and through Collocation Techniques
Sean Humphreys, Kevin Moffitt, Judee Burgoon and Jay Nunamaker (U of A)

Improving Quality Enhanced Biometric Fusion Schemes
Bojan Cukic, Afzel Noore, Nick Bartlow, Nathan Kalka, M. Vasta and R. Singh (WVU)

Matching and Retrieving of Face Images Based on Facial Marks
Arun Ross (WVU) and Anil K. Jain (MSU)

Phase 0 ‐ Participation in Multibiometric Grand Challenge
Stephanie Shuckers (Clarkson), Natalia Schmid (WVU), Besma Abidi and Uma Kandaswamy (UTK)

Psychophysiological Biometrics
Judee Burgoon (U of A), Reza Derakshani (UMKC), Arun Ross (WVU) and Diane Filion (UMKC)

Rapid Assessment Using Kiosk‐based Interviews
Mark Patton, Judee Burgoon and Jay Nunamaker (U of A)

Sequential Biometric Fusion Involving Incomplete or Missing Data
Arun Ross (WVU) and Anil K. Jain (MSU)

Iris Recognition Beyond 1000nm: A Preliminary Study
Arun Ross, Lawrence Hornak and Xin Li (WVU)

Autonomous Interrogation through Synchronous Computer‐Mediated Communication
Matthew Jensen, Douglas Derrick and Judee Burgoon (U of A)

Hybrid Expert System for Credibility Assement
Matthew Jensen, Jay Nunamaker and Judee Burgoon (U of A)

Kinesic Credibility Assessment of Criminal Interviews
Matthew Jensen, Judee Burgoon (U of A), Amy Franklin (Rice‐Linguistics) and Pete Blair (Texas State U‐Criminal Justice)

Summaries:

An Acquisition Platform for Non‐cooperative, Long Range Ocular Biometrics

Reza Derakhshani (UMKC), Plamen Doynov (UMKC) and Besma Abidi (UTK)

The objective of this proposal is to mitigate this problem by a COTS‐based capturing platform that will (a) scan a crowd using multiple, pan‐tilt systems, (b) locate subjects’ eyes in near‐infrared (NIR) (c) tag each locked‐on individual with a NIR pattern for concurrent, multi‐subject recognition (d) using burst (lucky) imaging, pick the best high‐zoom images of the eye regions by fast quality analysis algorithms for long‐range (up to 10m) and simultaneous ocular biometric recognition of unconstrained, freely moving crowds.

Automatic High Resolution Palmprint Matching System

Anil K. Jain (MSU) and Arun Ross (WVU)

Palmprints contain three different levels of information: Level 1 (principle lines, wrinkles and major creases), Level 2 (minutiae points) and Level 3 (ridge contour, pores). While Level 1 features can be extracted from fairly low resolution images (~100 ppi), higher resolution images are needed for extracting Level 2 (~500 ppi) and Level 3 (~1000 ppi) features [4]. Commercial palmprint acquisition systems do have the capability of capturing dual‐resolution images (500 and 1000 ppi), but not much work has been done to date in extracting Level 3 features from 1000 ppi images (ANSI NIST document on Extended Feature Set for fingerprint and palmprints [5]). Matching latent (partial) palmprints to full prints is particularly important in forensics application where Level 3 features will play a major role. In this project, we propose to design and evaluate an automated palmprint matching system that uses multiple levels of features.

Culturally Specific Credibility Classification Using Layered Voice Analysis

Judee Burgoon, Aaron Elkins, Douglas Derrick, Josh Hottenstein (U of A) and Dale Tunnell (Forensitec)

These newly available vocal measurements will provide a rich platform for examining the moderating effect of culture on deceptive leakage cues, particularly vocalics. Culture has been virtually unexplored in this context despite its potential to explain some of the individual variability in verbal and nonverbal behavior expressed during deceptive communication. The findings from this research have the potential to improve existing classification models utilizing other behavioral cues to detect deception. The credibility classification model for LVA will be developed using statistical analysis and machine learning algorithms on training data from a recent deception study focused on culture. Additionally, we will provide an empirical evaluation of the new generation of commercial LVA currently in use by law enforcement agencies.

Establishing Chain of Evidence in Biometric Systems

Bojan Cukic, Arun Ross, Nathan Kalka and Nick Bartlow (WVU)

The process of creating such a chain entails at least three types of validation to provide assurance that the collected biometric evidence has not been fabricated, altered or unintentionally mislabeled. These are validation of (a) evidence transmission, (b) content integrity, and (c) source of origin. While modern cryptography can adequately handle the validation of evidence transmission, validation of content integrity and source of origin must rely on other techniques. Biometric watermarking is the process of clandestinely embedding data into biometric images which can be used to assure validity of content. Digital hardware fingerprinting allows for the identification of source hardware from which an image originated. Through analysis of various types of sensor pattern noise and other artifacts, one can determine the technology, brand, model or specific sensor used to capture a biometric image. By utilizing cryptography, watermarking in conjunction with digital hardware fingerprinting and cryptography, a chain of evidence can be created, providing verification of image origination, authenticity, integrity, ownership, and non‐ repudiation of origin.

Improving the Identification of Fraud by Adding Word Sense Disambiguation to Linguistic Credibility Assessment/Enhancing Fraud Detection by Building Lexicons and through Collocation Techniques

Sean Humphreys, Kevin Moffitt, Judee Burgoon and Jay Nunamaker (U of A)

Automated techniques have been developed to assist in detecting deceit and fraud. However, existing models (e.g. Zhou et al. 2004) rely on proper identification of part‐of‐ speech tags (nouns, modal verbs, adjectives, etc) and can be harmed by ambiguous words. Deceivers are thought to use more ambiguity and hedging language. Computes are not nearly as good as humans at disambiguating words and their parts‐of‐speech. Errors in these automated tools can cause deception detection models to misclassify statements and documents. Using field data collected from the fraudulent SEC filings and insurance fraud, a linguistic analysis that includes a word sense disambiguator (WSD) will be undertaken to assess credibility and identify distinguishing linguistically cues related to credibility in text‐based claims.

Improving Quality Enhanced Biometric Fusion Schemes

Bojan Cukic, Afzel Noore, Nick Bartlow, Nathan Kalka, M. Vasta and R. Singh (WVU)

To design a hybrid fusion algorithm which combines the three approaches along with the uncertainties and precision of individual classifiers to improve the performance in cases of conflicting or missing information. Our hybrid classifier will decrease the computational complexity of resolving cases with conflicting unimodal evidences. Additionally, we will augment existing algorithms through the incorporation of a probabilistic measure of decision dependability which affords the opportunity to perform decision rectification or “reversing” classification decisions likely to be inaccurate. Through careful application of decision rectification, performance of quality enhanced biometric fusion algorithms may be improved. In both the proposed hybrid fusion algorithm and the existing algorithms augmented with decision rectification we expect to observe increases in performance when low quality samples are injected into systems previously trained on high quality subsets. Performance evaluation will be conducted using standard statistical measures (ROC curve) and through cost curves.

 

Matching and Retrieving of Face Images Based on Facial Marks

Arun Ross (WVU) and Anil K. Jain (MSU)

Face recognition systems typically encode the human face by utilizing either local or global texture features. Local (or part‐based) techniques first detect the individual components of the human face, prior to encoding the textural content of each of these components. Global techniques, on the other hand, consider the entire face as a single entity during encoding. However, both these techniques do not explicitly extract wrinkles, scars, moles, and other distinguishing marks that my present in the 2D image of the face. Many of these features are temporally invariant and can be useful for face recognition and indexing. The ability to automatically extract these marks or artifacts from facial images in large digital libraries can assist law enforcement agencies to rapidly locate human faces possessing specific marks. This proposal will design and implement techniques to (a) extract distinguishing marks present of the surface of the face and analyze their distributions, (b) efficiently retrieve face images from a digital database based on these marks, and (c) combine these distinguishing marks with a commercial texture‐based face matcher in order to enhance matching accuracy.

Phase 0 ‐ Participation in Multibiometric Grand Challenge

Stephanie Shuckers (Clarkson), Natalia Schmid (WVU), Besma Abidi and Uma Kandaswamy (UTK)

Fusion of face and iris data at a distance for biometric recognition could be extremely beneficial in places, like airport, port of entry, etc. The MultiBiometric Grand Challenge goal is to provide various recognition challenges for face and iris based on still and video imagery. We propose to participate in the MultiBiometric Grand Challenge (MBGC). MBGC has three stages. (1) Challenge 1 data is made available in May 2008. Results are to be presented in Dec 2008 at a workshop. We are processing the data now. (2) Challenge 2 dataset with results are presented in Spring 2009. (3) The last stage is the MultiBiometric Evaluation in Summer 2009. Our approach will be to fuse biometric information from both face and iris extracted over multiple frames. Quality information will be a critical component to select individual frames and to weigh information at the feature/pixel level. Fusion will be considered at the match score level and feature level where PDE‐texton maps or other features can be used to jointly encode to obtain robust representation of face and iris.

Psychophysiological Biometrics

Judee Burgoon (U of A), Reza Derakshani (UMKC), Arun Ross (WVU) and Diane Filion (UMKC)

We will utilize stimulus‐response psychophysiological reactions of individuals that can be captured by the typical biometric platforms in order to enhance their performance and security via imposter detection. We will focus on the well known pupillometry, blink, gaze, and other deception‐induced gestures of an individual such as hand and head gestures while interacting with the biometric system or an interrogator. In addition to the detection of imposters, the above psychophysiological traits should provide strong anti‐ spoofing measures. This project would serve as a synergistic project between CITeR‐ WVU and the new CITeR‐UA.

Rapid Assessment Using Kiosk‐based Interviews

Mark Patton, Judee Burgoon and Jay Nunamaker (U of A)

This is a first effort to utilize an automated rapid assessment kiosk to evaluate subjects for truth or deception. Subjects will be run through an international airline travel scenario where they will pack bags and proceed through a screening kiosk. Subject may or may not have items which are contraband or illegal for this type of travel. They will be asked a series of screening questions intended to elicit if they have anything to declare or if they are carrying any proscribed items. Their actions during the question and answer interacts will be video tapes, and their audio responses captured, for subsequent analysis. The experiment is intended to determine if standing subjects reveal stress or deception through body movement, if Voice Stress technologies can successfully flag either stress or deception in this setting, or if a fusion of these technologies can reveal stress and/or deception. It is also intended to reveal if there is any material difference based on method of questioning, either a speaking avatar or text on a screen which is read aloud to the subjects.

Sequential Biometric Fusion Involving Incomplete or Missing Data

Arun Ross (WVU) and Anil K. Jain (MSU)

This proposal seeks to address this problem in the context of an identification system by raising two pertinent questions. (a) Can a fusion algorithm be designed such that biometric information is consolidated in a sequential manner (as the traits become available) in order to determine an individual’s identity? (b) Can a score‐level or rank‐ level fusion scheme be designed to work in the presence of partial or incomplete biometric information? We will extend our current approach based on the Likelihood Ratio technique to devise methods that can effectively answer these questions thereby advancing the state‐of‐the‐art in biometric fusion for identification systems.

Iris Recognition Beyond 1000nm: A Preliminary Study

Arun Ross, Lawrence Hornak and Xin Li (WVU)

Most commercial iris recognition systems utilize information available in the 700 –900nm spectral band (near‐IR). Here, we investigate the possibility of performing iris recognition at higher wavelengths (1000 – 1500 nm, i.e., extended near‐ IR) to advance the science and technology of multispectral iris analysis in the biometric domain. The goal of this project is (a) to understand the composition of the iris at multiple resolutions using spectral information beyond 1000 nm (much like the Level I, II and III details in fingerprints); (b) to determine if iris segmentation can be successfully accomplished at these wavelengths; (c) to develop anti‐spoofing methods by studying the information revealed by various components of the eye at these wavelengths; and (d) to report recognition performance using multispectral information associated with these spectral bands.

Autonomous Interrogation through Synchronous Computer‐Mediated Communication

Matthew Jensen, Douglas Derrick and Judee Burgoon (U of A)

We will utilize artificially intelligent “chatbot” technology and advanced text processing algorithms to create a prototype autonomous interrogation system. The artificial integrator will have the ability to interact with a subject via synchronous communication (i.e., chat). In order to conduct the interrogation, the computer‐based agent will use a series of internal scripts and a complicated decision tree. The agent will ask questions of the subject, and process the responses in realtime in two ways. First, it will analyze the communication for potential deception using GATE / WEKA libraries and a text based deception model. Second, it will weigh the deception measurement and original message content against its decision tree and then formulate its response or next question.

Hybrid Expert System for Credibility Assement

Matthew Jensen, Jay Nunamaker and Judee Burgoon (U of A)

Among the most discriminating cues to deception are perceptual measures such as observed uncertainty, cognitive load, and non‐immediacy. We will incorporate such perceptual measures in an existing prototype that uses linguistic and kinesic analysis for credibility assessment. Such a hybrid expert system would include more of the unique capabilities that are necessary for unobtrusively monitoring interactions for indications of deceit and should improve credibility assessment performance.

Kinesic Credibility Assessment of Criminal Interviews

Matthew Jensen, Judee Burgoon (U of A), Amy Franklin (Rice‐Linguistics) and Pete Blair (Texas State U‐Criminal Justice)

Kinesic analysis has been successfully used to discriminate truth from deception in numerous experimental settings. However, these experimental settings do not provide representative levels of deceiver motivation and jeopardy that are present during high‐ stakes deception. For this project, we will have access to a new dataset that contains interviews captured during actual criminal investigations. Ground truth for all interviews has been established by confession or conviction of the suspect. Using this dataset, we will be able to further probe the capabilities of kinesic credibility assessment by using field data.