Contrastive Learning with Global Representation for Face Anti-spoofing

Face anti-spoofing (FAS) technology plays a crucial role in enhancing the face recognition system's safety and reliability. However, existing FAS methods generally rely on a significant volume of annotated data. Labeling the data is time-consuming and expensive. With the aim of solving the problem, PTL (progressive transfer learning) proposed an online learning framework for training FAS models with a few amount of labeled training data, such as only 50 face images. Despite its remarkable achievements, the approach does not explore the spoofing patterns inherent in face anti-spoofing task. To address this issue, we introduce a contrastive learning strategy with global representation (GRCL) into the training framework. By training the model to make it learn the global representation of the training data and create contrast loss for the overlapping regions of real faces and spoofing faces in the feature space, the feature generation ability of the model can be enhanced without destroying the original feature distribution. At the same time, we create a support set to dynamically generate global representative features in the training data for learning feature representations. Our approach can effectively work with the online learning framework to enhance the performance of models by using unlabeled data. Extensive experiments on four datasets with few labeled data validate the efficacy of our method.