Learning Optimal Transport Mapping of Joint Distribution for Cross-scenario Face Anti-spoofing

Face anti-spoofing (FAS) under different scenarios is a challenging and indispensable task for a real face recognition system. In this paper, we propose a novel cross-scenario FAS method by learning the optimal transport mapping of joint distributions under the unsupervised domain adaption framework, namely OTJD-FAS. In particular, given the training and testing real or fake face samples from different scenarios (i.e., source and target domains), their deep CNN features are firstly extracted and the labels of the test samples are firstly predicted by an initial binary classifier. Then, the gap of joint distributions (i.e., in the product space of deep features and their corresponding labels) between training and testing sets is measured by the Wasserstein distance and their optimal transport mapping is learned. Finally, an adaptive cross-entropy loss for classification and cross-entropy loss of testing labels are employed for the final FAS. Extensive experimental results demonstrated on the MSU-MFSD, CASIA-FASD and Idiap REPLAY-ATTACK databases under cross-scenario setting show that aligning joint distributions is more effective than the widely used only aligning marginal distributions based methods and the proposed method can achieve competitive performance for cross-scenario FAS.