Learning facial details for high-resolution face anti-spoofing

With face recognition playing a crucial role in biometric identification technology, Face Anti-Spoofing (FAS) has powerful effects on finding out whether a presented face is live or spoof. As the most common attacks such as photo attacks, print attacks, and video replay attacks can be effectively resolved, high-resolution attacks are easy to occur but still challenging for effective face spoofing because of the rich local facial details. In this paper, a Diagonal-Fusion Transformer network (DFT) which adds self-attention from the vision transformer is proposed. It is designed to learn the facial context information and relation between the local features of the face, and thus enhance the discriminative features of the real face and the fake face to improve the classification efficiency. Furthermore, a Spoofing Region Detection network (SRD) parallel with the DFT network is proposed for fine- grained spoof detection through the enlargement of local facial details. Through comprehensive experiments, the model achieves state-of-the-art results on public benchmark datasets such as OULU and CelebA-Spoof.