Adversarial Domain Generalization for Surveillance Face Anti-Spoofing

In traditional scenes (short-distance applications), the current Face Anti-Spoofing (FAS) methods have achieved satisfactory performance. However, in surveillance scenes (long-distance applications), those methods cannot be generalized well due to the deviation in image quality. Some methods attempt to recover lost details from low-quality images through image reconstruction, but unknown image degradation results in suboptimal performance. In this paper, we regard image quality degradation as a domain generalization problem. Specifically, we propose an end-to-end Adversarial Domain Generalization Network (ADGN) to improve the generalization of FAS. We first divide the accessible training data into multiple sub-source domains based on image quality scores. Then, a feature extractor and a domain discriminator are trained to make the extracted features from different sub-source domains undistinguishable (i.e., quality-invariant features), thus forming an adversarial learning procedure. At the same time, we have introduced the transfer learning strategy to address the problem of insufficient training data. Our method won second place in "Track Surveillance Face Anti-spoofing" of the 4th Face Anti-spoofing Challenge@CVPR2023. Our final submission obtains 9.21% APCER, 1.90% BPCER, and 5.56% ACER, respectively.