Self-supervised Face Anti-spoofing via Anti-contrastive Learning

Jiajiong Cao, Yufan Liu, Jingting Ding, Liang Li

Abstract: Face Anti-Spoofing (FAS) protects face recognition systems from Presentation Attacks (PA). Though supervised FAS methods have achieved great progress recent years, popular deep learning based methods require a large amount of labeled data. On the other hand, Self-Supervised Learning (SSL) methods have achieved competing performance on various tasks including ImageNet classification and COCO object detection. Unfortunately, existing SSL frameworks are designed for content-aware tasks, which may fail on other tasks such as FAS. To deal with this problem, we propose a new SSL framework called Anti-Contrastive Learning Face Anti-Spoofing (ACL-FAS). ACL-FAS contains two key components, namely, one PAtch-wise vIew GEnerator (PAIGE) and one Disentangled Anti-contrastiVe lEarning (DAVE) framework. With the help of two components, ACL-FAS shows its superiority on four different FAS datasets compared with more than 10 supervised methods and 5 SSL methods.