Multi-Perspective Features Learning for Face Anti-Spoofing

Abstract

Face anti-spoofing (FAS) is important to securing face recognition. Most of the existing methods regard FAS as a binary classification problem between bona fide (real) and spoof images, training their models from only the perspective of Real vs. Spoof. It is not beneficial for a comprehensive description of real samples and leads to degraded performance after extending attack types. In fact, the spoofing clues in various attacks can be significantly different. Furthermore, some attacks have characteristics similar to the real faces but different from other attacks. For example, both real faces and video attacks have dynamic features, and both mask attacks and real faces have depth features. In this paper, a Multi-Perspective Feature Learning Network (MPFLN) is proposed to extract representative features from the perspectives of Real + Mask vs. Photo + Video and Real + Video vs. Photo + Mask. And using these features, a binary classification network is designed to perform FAS. Experimental results show that the proposed method can effectively alleviate the above issue of the decline in the discrimination of extracted features and achieve comparable performance with state-of-the-art methods.