Face Anti-spoofing Method Based on Deep Supervision

Abstract

Although face recognition technology is extensively used, it is vulnerable to various face spoofing attacks, such as photo and video attacks. Face anti-spoofing is a crucial step in the face recognition process and is particularly important for the security of identity verification. However, most of today's face anti-spoofing algorithms regard this task as an image binary classification problem, which is easy to over-fit. Therefore, this paper builds the basic deep supervised network as the baseline model and designs the central gradient convolution to extract the pixel difference information within the local region. To reduce the redundancy of gradient features, the central gradient convolution is decoupled to replace the vanilla convolution in the baseline model to form two cross-central gradient networks. A cross-feature interaction module is then built to effectively fuse the networks. And a depth uncertainty module is built for the problem that most face datasets are noisy and it is difficult for the model to extract fuzzy region features. Compared with existing methods, the proposed method performs well on the OULU-NPU, CASIA-FASD, and Replay-Attack datasets.