CONTEXT-AWARE FUSION FOR MULTI-MODAL BIOMETRICS: WHOM DO I LISTEN TO AND WHEN?

Abstract

Continuous Authentication using biometrics is receiving renewed attention owing to recent advances in mobile technology. The context in which biometric inputs are acquired can affect the quality of information available for authentication. However, existing fusion methods do not take contextual information into account while combining the decisions of individual classifiers. The fundamental research question in this thesis is to effectively learn to combine decisions of multiple experts by utilizing contextual information to improve the accuracy of the authentication system. Two methods have been proposed to utilize contextual information available during the acquisition of biometric inputs and can operate at both decision and score levels of fusion. The theoretical bounds on the proposed methods are presented along with experiments on real and synthetic data. The experimental findings validate the key idea that context is essential to the fusion process, and show that the proposed methods outperform commonly used fusion methods. Moreover, the results show that the proposed methods outperform score level fusion methods even at the decision level, showcasing the power of contextual learning. Finally, the second method proposed is shown to allow finer control for balancing accuracy and the number samples required for convergence.