Similarity-aware contrastive learning for face anti-spoofing via frequency enhancement and reconstruction.

Abstract

With the widespread application of face recognition systems, face anti-spoofing (FAS) has emerged as a critical research topic. Recent studies increasingly focus on the frequency domain instead of the spatial domain, as subtle spoofing patterns can be more effectively captured there. However, how to more comprehensively mine spoofing clues in different frequency bands remains a challenging issue. In this paper, we propose a similarity-aware contrastive learning method in the frequency domain for FAS. Specifically, we design a frequency domain adaptive enhancement module that dynamically integrates multiple frequency bands to learn domain-invariant features. Furthermore, we introduce a high-frequency information reconstruction strategy to preserve structural consistency while highlighting subtle spoofing clues by mixing the original frequency spectrum with its corresponding high-frequency components. Finally, we construct a similarity matrix to quantify relationships among different types of face samples and formulate a contrastive learning objective to enforce intra-class compactness and inter-class separability. Extensive experiments show that our approach outperforms existing state-of-the-art methods on multiple public datasets and evaluation protocols.