Multi-domain mixup for scenario-universal face anti-spoofing

Recently, disentangled representation learning has been commonly used in face anti-spoofing (FAS). However, such method has limited generalization ability due to the lack of data domains in the training phase. To overcome this issue, we devise a novel disentangling framework, which contains Progressive Refinement Disentangling (PRD) module and Multi-Domain Mixup (MDM) module. Concretely, face images are well disentangled into liveness features and domain features via the PRD module. The MDM module aims to produce more diverse domain features to generate faces of brand-new domains. The generated faces could improve the generalization ability of model in a data augmentation manner. Moreover, our disentangling framework is capable of tapping the potential of unlabeled data so it is universal in semi-supervised, domain generalization and adaption scenarios. Extensive experiments demonstrate the effectiveness of our method on public datasets.