SAMCNet: A Multi-channel Face Anti-spoofing Network Combined with Hyperspectral Images via Self-attention Mechanism

In recent years, the importance of countering face spoofing has increased due to the widespread use of facial recognition in biometrics. Current strategies rely on RGB images, but they often struggle to detect presentation attacks. In this study, we propose a novel network for face anti-spoofing that integrates RGB, depth, infrared, and hyperspectral channels. By transforming RGB into hyperspectral channels using sparsity and spectral self-similarity, we extract more information from facial images. We introduce a self-attention mechanism to capture correlations among features and use global mean pooling for classification. Experiments on Replay-Attack and CASIA-SURF datasets show significant improvement, achieving a 0.25% HTER on Replay-Attack and a 99.3% TPR at a 10e-4 FPR on CASIA-SURF.