A Dual-Branch Network with Supcon for Face Anti-spoofing

In recent years, face anti-spoof (FAS) has become an important research direction for ensuring the security of face recognition systems. Previous works mostly utilized auxiliary pixel-level supervision to train networks, attempting to introduce additional models for deep pixel-level supervision, or only used local feature self-supervision to address deception issues. For the task of FAS, the pixel information in images is adequate for spoof classification. However, the simplicity of binary classification and the subtlety of inter-class differences can lead to overfitting on the training data, thereby resulting in weaker generalization on the test set. In this work, we achieve better generalization ability by sharing convolutional layers and adding global and local dual-branch after the convolutional layers. By supervising the dual branches using supervised contrastive (Supcon), and conducting feature selection and fusion, we are able to effectively perform spoof classification, forcing the network to learn differences.