Gradient-Reversed Domain-Generalizable Multi-Modal Cross-Attention Network for Robust Face Anti-Spoofing

Face recognition systems are increasingly vulnerable to presentation attacks, wherein adversaries employ artifacts such as printed photographs, replayed videos, or 3D masks, etc, to impersonate genuine users. Despite significant progress in face anti-spoofing, most existing methods exhibit poor generalization when confronted with unseen attack types or domain shifts such as changes in sensors, environments, or acquisition protocols, thereby limiting their robustness in real-world applications. We propose the Gradient-Reversed Domain-Generalizable Multi-Modal Cross-Attention Network (GR-DXNet) that enhances resilience against previously unseen attacks and enables seamless adaptation across diverse domains. GR-DXNet employs dual-modality learning by fusing RGB frames, which capture fine-grained texture, with depth maps that reveal 3D structural cues. Temporal Convolutional Networks (TCNs) are integrated to model motion-based inconsistencies, improving the detection of dynamic emerging spoof patterns. To enhance cross-modal representation, a query-key-value-based cross-attention mechanism is introduced, enabling effective alignment and fusion of RGB and depth features. Furthermore, a post-fusion Gradient Reversal Layer (GRL) is employed that adversarially aligns cross-modal embeddings to suppress domain-specific bias without handcrafted augmentations or complex disentanglement, encouraging the model to learn domain-invariant features and strengthens generalization under unseen domains. Extensive evaluations on benchmark datasets across both intra-and cross-dataset protocols, offering a reliable solution for real-world face spoof detection.