Research Of Face Anti-Spoofing Based On Representation Learning And Representation Disentanglement

Face recognition systems are vulnerable to face spoofing attacks.Specifically,an attacker can trick a face recognition system into making a wrong judgment by presenting a spoof face instead of a live face to the face recognition system.Typical spoof faces include printed photos,replayed videos,and 3D masks.Therefore,face anti-spoofing(FAS)is one of the key problems that face recognition systems have to solve.Extraction of distinguishable representations is the current mainstream FAS method,the basic idea is to first obtain the representations to distinguish spoof faces from live once via deep learning,and then use the learned representations to complete the live/spoof face classification.Therefore,the quality of the representations plays an important role in the classification ability of FAS.Most existing FAS methods focus on finding some kind of spoof cues(e.g.,heart rate signal,facial depth,etc.)and learning their corresponding representations from the specified spoof cues for classification.However,few studies have been carried out to improve the quality of representation.To address the above issues,we propose two major methods for obtaining highquality representations and using them in FAS from the perspectives of representation learning and representation disentanglement,respectively,i.e.,a FAS method based on representation learning and a FAS method based on representation disentanglement.(1)The key idea of the FAS method based on representation learning is that:improving the quality of representations for FAS by enhancing the representation precision learned from the network.Specifically,the method improves representation precision in three ways: enhancing the representation learning capability of the network,adaptively selecting different types of representations,and reducing the similarity between different types of representations.(2)The key idea of the FAS method based on representation disentanglement is that:observing that the learned representations are often an entanglement of two major representations: critical representations that are favourable to FAS,and irrelevant representations that are unfavourable to FAS.Thus,even if the representation precision is improved by representation learning methods,there may still be irrelevant representations that are detrimental to the FAS.If the critical and irrelevant representations can be disentangled into two mutually independent representation spaces,and the critical representations can be used for FAS,the classification capability of FAS may be enhanced.Experiments on four public datasets(i.e.,OULU-NPU,Si W,Replay-Attack and CASIA-MFSD)show that the proposed two algorithm achieve comparable or even more prominent performance than the state-of-the-art FAS method on representation learning and disentanglement,respectively.