Abstract
Presentation attacks pose a significant threat to face recognition systems, making face anti-spoofing (FAS) a critical security measure. However, many existing approaches suffer from inadequate exploitation of physical modality cues and rely on overly complex architectures, which hinder their deployment in practical applications. This paper introduces DAH-FAS, a Dynamically-Aware Heterogeneous Face Anti-Spoofing Network designed to mitigate the limitations of existing face anti-spoofing methods. To reinforce the RGB branch’s capacity for detailed feature extraction, we have designed a Variance-Adaptive Multi-Scale Residual Block (VA-MSRB). To improve the model’s perception of bio-thermal diffusion patterns, the BioThermal Enhancer (BTE) is integrated into the GhostNet backbone of the IR branch. On this basis, a Bidirectional Group Cross-Modal Attention (BGC-MA) mechanism is constructed between the IR and depth branches during the feature extraction stage, enabling cross-modal geometric feature alignment and enhancing the complementarity among features. We evaluate our method on the CASIA-SURF, CASIA-SURF CeFA, and WMCA datasets, and results demonstrate that the proposed approach achieves significant advantages in differentiating real from fake faces.