A Robust Method with DropBlock for Face Anti-Spoofing

Abstract

Face anti-spoofing is an important component for reliable face recognition. Previous deep learning methods usually exploit a large network structure and are prone to be overfitting when the face anti-spoofing database is a small scale. And in some occasions, multi-frame information or other auxiliary information is not available. In order to address these issues, we propose an end-to-end deep learning approach with DropBlock layer to distinguish between a fake face and a genuine face, which makes use of frame-only RGB image. In this paper, we evaluate the proposed approach with different settings on three popular databases, i.e., REPLAY-ATTACK, CASIA-FASD, and CASIA-SURF. The proposed approach can achieve competitive results on these databases. The experimental results show that our model is more robust and can learn more generalized features for face anti-spoofing.