FAS-MBCA: Face Anti-Spoofing Using Multi-Path Bi-Directional Cross-Attention Fusion

Face Anti-Spoofing (FAS) is a critical component for securing face recognition systems against diverse presentation attacks. multi-modal FAS approaches leverage complementary cues from multiple sensors; however, existing methods often adopt single-path fusion or unidirectional interaction schemes, which may introduce modality hierarchy and limit comprehensive cross-modal feature refinement. In this paper, we propose FAS-MBCA, a structured multi-path bi-directional cross-attention framework for tri-modal face anti-spoofing. FAS-MBCA explicitly constructs a complete pairwise interaction topology among RGB, depth, and infrared modalities, enabling reciprocal feature-map refinement across all modality pairs. By modeling six-direction cross-modal interactions within a lightweight CNN-based architecture, the proposed framework enhances multi-modal synergy while maintaining computational efficiency. We evaluate FAS-MBCA on multiple widely used benchmarks, including CASIA-SURF (MmFA), CASIA-CeFA, and WMCA, under both intra-dataset and cross-dataset protocols. Extensive experiments demonstrate that FAS-MBCA consistently outperforms single-modal baselines and naïve fusion strategies, and achieves competitive or superior performance compared to recent multi-modal methods. Efficiency analysis in terms of model parameters and GFLOPs further indicates a favorable accuracy–efficiency trade-off. These results validate that explicitly modeling complete bi-directional cross-modal interaction at the feature-map level provides an effective and practical solution for multi-modal face anti-spoofing. The source code of FAS-MBCA is publicly available at https://github.com/nascafas2025/FAS-MBCA