Biometric Presentation Attack Detection for Mobile Devices Using Gaze Information

Facial recognition systems are among the most widely deployed in biometric applications. However, such systems are vulnerable to presentation attacks (spoofing), where a person tries to disguise as someone else by mimicking their biometric data and thereby gaining access to the system. Significant research attention has been directed toward developing robust strategies for detecting such attacks and thus assuring the security of these systems in real-world applications. This thesis is focused on presentation attack detection for face recognition systems using a gaze tracking approach. The proposed challenge-response presentation attack detection system assesses the gaze of the user in response to a randomly moving stimulus on the screen. The user is required to track the moving stimulus with their gaze with natural head/eye movements. If the response is adequately similar to the challenge, the access attempt is seen as genuine. The attack scenarios considered in this work included the use of hand held displayed photos, 2D masks, and 3D masks. Due to the nature of the proposed challenge-response approaches for presentation attack detection, none of the existing public databases were appropriate and a new database has been collected. The Kent Gaze Dynamics Database (KGDD) consists of 2,400 sets of genuine and attack-based presentation attempts collected from 80 participants. The use of a mobile device were simulated on a desktop PC for two possible geometries corresponding to mobile phone and tablet devices. Three different types of challenge trajectories were used in this data collection exercise. A number of novel gaze-based features were explored to develop the presentation attack detection algorithm. Initial experiments using the KGDD provided an encouraging indication of the potential of the proposed system for attack detection. In order to explore the feasibility of the scheme on a real hand held device, another database, the Mobile KGDD (MKGDD), was collected from 30 participants using a single mobile device (Google Nexus 6), to test the proposed features. Comprehensive experimental analysis has been performed on the two collected databases for each of the proposed features. Performance evaluation results indicate that the proposed gaze-based features are effective in discriminating between genuine and presentation attack attempts