Hyperbolic face anti-spoofing

• Main Author: Han, Shuangpeng
• Other Authors: Alex Chichung Kot
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/167740
• Institution: Nanyang Technological University

Face anti-spoofing (FAS) plays an important role in face recognition systems, which has attracted the interest of many researchers. Most of the previous models in the field of face anti-spoofing are designed in Euclidean space. They try to learn some discriminative features in Euclidean space to widen the distance between bonafide samples and attack samples. But they may have limited generalization ability for some unseen attacks, and it is challenging to learn hierarchies between and within spoofing attacks. Recent studies have found that hyperbolic space can effectively embed data with latent hierarchical structures and have impressive generalization ability. Therefore, face anti-spoofing in hyperbolic space can be a promising alternative. The proposed hyperbolic framework for face anti-spoofing is to project the features obtained in the Euclidean space to the Poincare ball and complete the classification through the hyperbolic binary logistic regression layer. In addition, a hyperbolic contrastive loss to the hyperbolic space is added to help the model better distinguish between bonafide samples and attack samples. To ensure the stability of model training and avoid the vanishing gradient problem, a simple but effective feature clipping method is designed in hyperbolic space. The proposed hyperbolic framework is implemented on two benchmarks (WMCA and PADISI-Face) with a variety of attacks. Experiments demonstrate that, for unseen attack detection, the proposed hyperbolic framework can surpass the performance of Euclidean baselines.