Binary - Feature Based Recognition and Cryptographic Key Generation from Face Biometrics

Identifying an individual from surveillance video is a difficult, time consuming andlabour intensive process. This is due to surveillance footage being of poor resolutionlimited by processing, storage and cost constraints. In addition, the process is prone tohuman error caused by fatigue as operators often need to scan through hours of videojust to find a few seconds of useful footage.The proposed system [3] aims to streamline this process by filtering out unwantedscenes and enhancing an individual’s face through super-resolution. An automatic facerecognition system is then used to identify the subject or present the human operatorwith likely matches from a database. A person tracker is used to speed up the subjectdetection and super-resolution process by tracking moving subjects and cropping aregion of interest around the subject’s face to reduce the number and size of the imageframes to be super-resolved respectively.As the super-resolution reconstruction process is ill-posed, visual artifacts are oftengenerated. These artifacts can be distracting to humans and/or affect machinerecognition. While it is intuitive that higher resolution should lead to improvedrecognition accuracy, the effects of super-resolution and such artifacts on facerecognition performance have not been systematically studied as done here. It is shownthat super-resolution allows more accurate identification of individuals from low-resolution surveillance footage.The proposed robust optical flow-based super-resolution method used in the systemis benchmarked against Baker et al.’s [1] hallucination and Schultz et al.’s [5] MAPsuper-resolution techniques on images from the Terrascope [2]and XM2VTS [4]databases. Ground truth and interpolated images were also tested to provide a baselinefor comparison. Face identification performance were tested on an Eigenface [6] andElastic Bunch Graph Matching (EBGM) [7] system using the XM2VTS database.The experiments show that the MAP method is limited in performance by theassumption of rigid objects in the scene. The hallucination and optical flow-basedmethods performed comparably. The optical flow-based method produces less artifactsthat are visually distracting for human recognition. On average, the proposed opticalflow-based method outperformed hallucination by 1.68% and bilinear interpolation by5.82% in recognition rate. References[1] S. Baker and T. Kanade. Limits on Super-Resolution and How to Break Them. 24(9):1167–1183,September 2002.[2] C. Jaynes, A. Kale, N. Sanders, and E. Grossmann. The Terrascope dataset: scripted multi-camera indoorvideosurveillance with ground-truth. In Proc. Visual Surveillance and Performance Evaluation of Tracking andSurveillance,pages 309–316, October 2005.[3] F. Lin, S. Denman, V. Chandran, and S. Sridharan. Automatic Tracking and Super-Resolution of HumanFaces fromSurveillance Video. In Proc. MVA 2007, May 2007.[4] K. Messer, J. Matas, J. Kittler, J. Luettin, and G. Maitre. XM2VTS: The Extended M2VTS Database. In Proc.AVBPA-1999, pages 72–76, 1999.[5] R. Schultz and R. Stevenson. Extraction of High-Resolution Frames from Video Sequences. IEEETransactions onImage Processing, 5(6):996–1011, June 1996.[6] M. Turk and A. Pentland. Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3(1):71–86, March1991.[7] L. Wiskott, J. Fellous, N. Kr ̈uger, and C. Malsburg. Face recognition by elastic bunch graph matching. InProc. CAIP’97, number 1296, pages 456–463, 1997. Biometrics Institute Poster Exhibition, 7 8 June 2007_______________________________________________________________________________ © Biometrics Institute 20073Binary-Feature Based Recognition and Cryptographic Key Generation from FaceBiometrics Brenden ChenImage and Video Research LaboratoryGPO Box 2434, Brisbane QLD 4001, Australia Queensland University of Technology AbstractCryptography has proven to be the principal method of securing digital data fromattack. The use of RSA based public/private key architecture is widespread amongstinternet applications. Although RSA is strong against brute force attacks it still suffersfrom weaknesses due to insecure key protection by poor user selected passwords. Aproposed solution to this is to generate keys directly from user biometrics.The primary obstacle of using biometric based keys is that biometric data is almostnever constant, differences exist in acquisitions of the same user due to misalignment,lighting change and pose or expression changes. This makes it virtually impossible toextract a deterministic key directly from the raw biometric data itself. But despite thesechanges recognition can still be performed indicating that consistent information exists inthe biometric data. It is the challenge of biometric cryptography to isolate this constantinformation while discarding the random.A method proposed in [1] discusses an iterative one-way chaotic bi-spectraltransform that can be applied to face images extracted from video to recognize usersand generate random numbers (symmetric keys). The output of this transform can beconverted into binary form allowing the use of simple binary analysis techniques, namelybit probabilities and entropy to detect the desirable bits for cryptographic keygeneration, these bits can be considered as ‘binary features’ and can also be used forrecognition. The process allows for keys of varying lengths to be generated as well asrevoked.The process is empirically tested using the FRGC[2] and ORL[3] databases forrecognition performance and the Talking Face Video project [4] and Caltech facedatabase[5] for key generation. Results produced indicate keys of up to 256 bits can begenerated with reasonable consistency and can be compared to other biometriccryptographic methods proposed by Monrose [6] and Hao [7,8].Cryptography has proven to be the principal method of securing digital data fromattack. The use of RSA based public/private key architecture is widespread amongstinternet applications. Although RSA is strong against brute force attacks it still suffersfrom weaknesses due to insecure key protection by poor user selected passwords. Aproposed solution to this is to generate keys directly from user biometrics.The primary obstacle of using biometric based keys is that biometric data is almostnever constant, differences exist in acquisitions of the same user due to misalignment,lighting change and pose or expression changes. This makes it virtually impossible toextract a deterministic key directly from the raw biometric data itself. But despite thesechanges recognition can still be performed indicating that consistent information exists inthe biometric data. It is the challenge of biometric cryptography to isolate this constantinformation while discarding the random.A method proposed in [1] discusses an iterative one-way chaotic bi-spectraltransform that can be applied to face images extracted from video to recognize usersand generate random numbers (symmetric keys). The output of this transform can beconverted into binary form allowing the use of simple binary analysis techniques, namelybit probabilities and entropy to detect the desirable bits for cryptographic keygeneration, these bits can be considered as ‘binary features’ and can also be used forrecognition. The process allows for keys of varying lengths to be generated as well asrevoked.The process is empirically tested using the FRGC[2] and ORL[3] databases forrecognition performance and the Talking Face Video project [4] and Caltech facedatabase[5] for key generation. Results produced indicate keys of up to 256 bits can begenerated with reasonable consistency and can be compared to other biometriccryptographic methods proposed by Monrose [6] and Hao [7,8]. References[1] V. Chandran and B. Chen, “Simultaneous Biometric Verification and Random Number Generation,” The 5Workshop on the Internet, Telecommunications and Signal Processing, 2006[2] P. J. Phillips, P. J. Flynn, T. Scruggs, Kevin W. Bowyer, J.Chang, K. Hoffman, J. Marques, J. Min, and W.Worek,“Overview of the face recognition grand challenge,” Proceedings of IEEE Comp. Society Conf. onComputer Vision and Pattern Recognition (CVPR’05),Vol. 1,Washington, DC, USA, pp. 947–954, 2005.[3] F. Samaria, “ORL Face Database”, AT&T; Laboratories Cambridge, Cambridge University, 1994, Available:http://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html[4] T. Cootes, “Talking Face Video”, Division of Imaging Science and Biomedical Engineering, University ofManchester, Available: http://www.isbe.man.ac.uk/~bim/data/talking_face/talking_face.html[5] M. Weber, “Frontal Face Dataset”, Computational Vision Group, California Institute of Technology, 1999,Available: http://www.vision.caltech.edu/html-files/archive.html[6] F. Monrose, M.K.. Reiter, Q. Li and S. Wetzel, “Cryptographic Key Generation from Voice,” Proceedings ofthe IEEE conference on Security and Privacy, May 2001.[7] F. Hao and C.W. Chan, “Private Key Generation from On-Line Handwritten Signatures,” InformationManagement and Computer Security, Vol. 10, no. 2, pp. 159-164, 2002.[8] F. Hao, R. Anderson and J. Daugman, “Combining Crypto with Biometrics Effectively,” IEEE Transactionson Computers, Vol. 55, no. 9, pp. 1081-1088, September 2006. Biometrics Institute Poster Exhibition, 7 8 June 2007_______________________________________________________________________________ © Biometrics Institute 20074Face Recogintion Robust to Large Pose Angle from One Gallery Image Ting Shan, Brian Lovell and Shaokang Chen, Queensland [email protected] National ICT Australia Confidential. Copyright © NICTA, 2007Face Recognition Robust to LargePose Angle from One Gallery Image Facial Feature Interpretation by AAMs 0.00%10.00%20.00%30.00%40.00%50.00%60.00%70.00%80.00%90.00% 25 L15 L15 R25 RPCA onoriginalimagesAPCA onoriginalimagesPCA onsynthesizedimagesAPCA onsynthesizedimagesRealistic Frontal View Synthesis via Correlation Model Significant Improvement in RecognitionPerformance by Pose CompensationDifferent Views Synthesized from One Frontal Face ImageTing Shan, Brian Lovell, and Shaokang Chen, Queensland Lab Acknowledgement: This project is supported by a grant from theAustralian Government Department of the Prime Minister and Cabinet.NICTA is funded by the Australian Government's initiative, in partthrough the Australian Research Council. Biometrics Institute Poster Exhibition, 7 8 June 2007_______________________________________________________________________________ © Biometrics Institute 20075Continuous Verification Using Multimodal Biometrics Terence Sim,Assistant Professor School of Computing,[email protected]/~tsim National University of Singapore Abstract:Conventional verification systems, such as those controlling access to a secure room,do not usually require the user to re-authenticate himself for continued access tothe protected resource.Conventional verification systems, such as those controlling access to a secure room,do not usually require the user to re-authenticate himself for continued access tothe protected resource. This may not be sufficient for high security environments in which the protectedresource needs to be continuously monitored for unauthorized use. In such cases, continuous verification is needed. In this poster, we presentthe theory, architecture, implementation, and performance of a multimodal biometricsverification system that continuously verifies the presence of a logged-in user. Twomodalities are currently used — face and fingerprint — but our theorycan be readily extended to include more modalities. We show that continuous verificationimposes additional requirements on multimodal fusion when compared to conventionalverification systems. We also argue that the usual performance metrics of false accept and false rejectrates are insufficient yardsticks for continuous verification, and propose new metricsagainst which we benchmark our system. Biometrics Institute Poster Exhibition, 7 8 June 2007_______________________________________________________________________________ © Biometrics Institute 20076The Psychology of Facial Identification Systems: The Importance of ConsideringHuman Performance when Designing Biometrics Systems Richard Kemp & Mark Howard, School of Psychology University of New South Wales, Sydney 2052 In this poster we present a summary of some of our recent research into theperformance of human operators who undertake the task of determining whether twophotographs are of the same unfamiliar person. Research which we and others haveundertaken (e.g. Kemp et al 1997; Burton et al, 1999) has demonstrated that this is asurprisingly difficult task and that even expert operators make both false positive andfalse negative errors (see Figure 1a). In some instances the performance of automaticsystems is better than human systems in this task. Our more recent research suggestsseveral avenues to improve performance. We argue that it is critical for the designers of facial biometrics systems, such as thosecurrently being introduced at some international airports, to understand the limitedability of human operators working in this role. We note that many of the systems beingplanned are designed to generate a relatively high number of false negative errors whicha human operator will be expected to resolve. Our concern is that no consideration isbeing given to the ability of the operators to undertake that task. In an ideal system the performance of both the machine and the human will beconsidered as components of an integrated system. In such a system the machineelements will be designed to complement the performance of the human operators sothat, for example, decisions which are challenging for the machine are easy for thehumans and vice versa. Further, in an ideal system the task will be presented to thehuman operator in such a way as to maximise their performance. In some recent research (Kemp et al 2007) we have demonstrated that changes to theway in which the images are presented to the human operator can improve their abilityto detect non-matching photographs. In a series of experiments we have shown thatmasking the external features of the face including the hair and face outline can bothincrease hits and reduce false alarms (see figure 1b). We offer this as an example of howknowledge of human face perception can be incorporated into the design of wholesystem. Only with a thorough understanding of the role and performance of human operator canwe maximise the accuracy and efficiency of the systems we design. We invite thedesigners of biometrics systems to collaborate with us to achieve this goal. Figure 1. An example of a different-matched trial. When shown whole images,(part a), many participations incorrectly conclude that these two photographs areof the same person, but when shown the same images with external featuresremoved (part b), significantly more participants are able to correctly determinethat the two images are of different individuals. Correspondence: Dr Richard Kemp, School of Psychology,University of New South Wales, Sydney 2052. Australia. Ph: +61 (2) 9385 1401;email: [email protected] (a)Part (b)