Biometrics and Robust Face Recognition

The use of conformal prediction in general, and transduction in particular, is discussed in terms of scope, challenges, and vulnerabilities for biometrics with robust face recognition as application domain of interest. Robustness refers to the ability to cope with uncontrolled settings characteristic of incomplete and/or corrupt (adversarial) biometric information on one side, and varying image quality on the other side. The motivation for the conformal prediction approach comes from the use of discriminative methods, such as likelihood ratios, to link biometrics and forensics. The methods and algorithms proposed are realized using transductive inference (transduction for brevity). They leverage nonconfidence measures (NCM), make use of both labeled (annotated) and unlabeled biometric data, address multilayer categorization, and provide measures of reliability in the predictions made, such as credibility and confidence. Toward that end we describe a novel Transduction Confidence Machine for Detection and Recognition (TCM-DR) that expands on the traditional Transduction Confidence Machine (TCM). The two machines, TCM and TCM-DR, are suitable for closed and open-set recognition, respectively, with TCM-DR also suitable for verification. Basic concepts, architectures, and empirical results are presented for open set face recognition and watch list/surveillance using TCM-DR. Recognition-by-parts using transduction and boosting is the adversarial learning solution that addresses vulnerabilities due to occlusion and disguise. Future venues for biometric research are discussed including reidentification using sensitivity analysis and revision for the purpose of metaprediction in general, and interoperability and identity management in particular.