Cross-domain face presentation attack detection techniques with attention to genuine faces

• Main Author: Li, Zhi
• Other Authors: Alex Chichung Kot
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/165704
• Institution: Nanyang Technological University

Face recognition as a convenient approach for automatic identity verification has become increasingly prevailing in recent years. The presentation attack (PA) is a serious threat hindering the application of face recognition systems in security-critical scenarios. Face presentation attack detection (PAD) is an essential anti-spoofing measure to enhance the security of face recognition systems by discriminating presentation attacks from bona fide attempts. Existing methods have achieved good performance in intra-domain testing, where the testing data is from the same distribution as training data. However, when testing the face PAD models in a new target domain, the performance will degrade severely since the testing data could be from unseen distributions which are different from the training data. In this thesis, we explore the cross-domain problems in face PAD and introduce several methods to apply to different application scenarios. In consideration of the intrinsic difference between genuine face and attack samples, such as the feasibility and the expense of data collection in practical scenarios, our methods are devised with more attention to genuine face samples. Considering that the attackers may launch presentation attacks with novel spoofing mediums, we study the unseen attack problem in face PAD in the first work and propose method based on deep metric learning. We learn a discriminative feature space with a hypersphere loss which forces the genuine face samples to maintain intra-class compactness and ensure inter-class separation from the attack samples. Since the decision-making is directly conducted on the learned feature space, there is no need for additional classifiers to be trained. Beyond the threats of unseen attacks, the changes in illumination conditions and camera sensors will also degrade the reliability of the face PAD systems. In the second work, we tackle the generalization problems in face PAD and propose a bi-modality method that better generalizes to unseen attack and illumination variations. We establish the connection between face images of different modalities via asymmetric modality translation. The discrepancy of modality translation between genuine faces and attack samples is used as a compelling clue for discriminating various spoofing faces from genuine faces. Domain adaptation is a typical approach to improving the cross-domain performance of face PAD with the help of target domain data. However, it has always been a non-trivial challenge to collect sufficient data samples in the target domain, especially for attack samples. In the third work, we improve the cross-domain performance of the face PAD by only using a few genuine face samples collected in the target domain. We propose a method by introducing teacher-student learning to address the one-class domain adaptation problem in face PAD. The similarity score between the representations of the teacher and student networks is used to distinguish attacks from genuine ones. To verify the effectiveness of the proposed methods, we devise protocols and conduct extensive experiments on multiple datasets. The experimental results show that our methods outperform prior methods.