Face Anti-Spoofing Based on Hyperbollic Crossdomain Gradient Alignment and Fine-Grained Linear Interpolation

To enhance the robustness of Face Anti-Spoofing System (FAS) in real-world applications, an increasing amount of research is being conducted on cross-domain generalization in this field. However, under complex network conditions, when faced with limited bandwidth, we may have to opt for low-resolution, lossy images for detection. In such scenarios, many cross-domain methods may not remain applicable; features in Euclidean space may become even less separable due to information loss, and the complex cross-domain modules can significantly increase the cost of training and optimization. Therefore, this paper proposes the combination of Hyperbolic Cross-Domain Flattening Alignment (HYP-GAS) and Patch Fine-Grained Linear Interpolation (PATCH-LISA). The former achieves its goal by mapping features from Euclidean space to hyperbolic space and separately calculating the Supcon loss in Euclidean space and the BCE loss in hyperbolic space, guiding gradient flattening with losses from both spaces to achieve better convergence in both. Moreover, to further improve the model's cross-domain capability without altering the network structure, which could lead to potential gradient vanishing issues, PATCH-LISA is introduced. This method involves dividing images into patches at a fine-grained level based on pixel position and performing patch-level linear interpolation mixing according to the different regions' pixel positions. The aim is to distinguish the importance of live feature positions at different pixel locations and better extract hidden information. Experimental results show that this method outperforms commonly used domain generalization algorithms.