Research On Face Anti-Spoofing

Face anti-spoofing(FAS)is an important biometric technology.Specifically,input images need to make authentication between living and spoofing before entering face recognition systems.Typical presentation attacks include print attacks,replay attacks,and 3D mask attacks.Therefore,a superior FAS method needs to not only recall as many attack images as possible,but also reduce mistakes for real faces.FAS has lately attracted increasing attention from industry and academia due to its vital role in securing face recognition systems.Most previous methods are designed based on CNN structures.However,CNNs are not good at exploiting fine-grained information,which may limit the further improvement of performance.Therefore,we try adopting Transformer structures into single-frame and multi-frame FAS to further improve the performance.In addition,the previous FAS methods have achieved promising performance in intra-domain scenarios,but may encounter dramatic degradation under the cross-domain settings.Therefore,we propose a DG method to realize domain generalization for FAS.Specifically,this paper mainly focuses on the following aspects for FAS:1)We design a single-frame FAS method called TransFAS by using Transformer structures.Comprehensive experiments are conducted to prove the effectiveness of our method.Firstly,we observe the long-range relation difference in the patches between living and spoofing.To take advantage of the above characteristics,we utilize a light-weight Transformer network as the backbone to capture the patch-wise relation awareness in the same layers.Furthermore,to effectively collect multi-scale features,we build up a GCN-based structure to fuse the features from different layers and explore the best structure for hierarchical feature fusion.2)We design a multi-frame FAS method called TTN by using Transfomer structures.Comprehensive experiments are conducted to prove the effectiveness of our method.Compared with the single-frame method,the multi-frame method can utilize temporal information to further develop the algorithms.Thus,we design spatial and temporal transformer parts to explore long-term and shortterm temporal information for binary supervision and depth supervision,respectively.Moreover,we develop our TTN for multi-modal input,which can further enhance the performance of our method.3)We design a DG method called SSAN to realize domain generalization(DG)for FAS.The previous FAS methods have achieved promising performance in intra-domain scenarios,but may encounter dramatic degradation under the cross-domain settings.To address this issue,we design the DG method SSAN.Different from directly using the complete representations for DG,we separate the complete representation into content and style ones in this work.On the other hand,despite the decent performance,there still exists a gap between academia and industry,due to the difference in data quantity and distribution.Thus,a new large-scale benchmark for FAS is built up to further evaluate the performance of algorithms in reality.