Face Anti-Spoofing (FAS) is crucial for securing face recognition systems against presentation attacks. However, existing multi-modal FAS approaches face challenges in generalizing to unseen attacks and deployment conditions due to modality unreliability and imbalance. To address these issues, the authors propose the Uncertainty-Guided Cross-Adapter (U-Adapter) and the Rebalanced Modality Gradient Modulation (ReGrad) strategy. The U-Adapter recognizes and suppresses unreliable detected regions within each modality, while the ReGrad strategy rebalances the convergence speed of all modalities by adaptively adjusting their gradients. The authors also establish the first large-scale benchmark for evaluating multi-modal FAS performance under domain generalization scenarios. Extensive experiments demonstrate that their method outperforms state-of-the-art methods.Face Anti-Spoofing (FAS) is crucial for securing face recognition systems against presentation attacks. However, existing multi-modal FAS approaches face challenges in generalizing to unseen attacks and deployment conditions due to modality unreliability and imbalance. To address these issues, the authors propose the Uncertainty-Guided Cross-Adapter (U-Adapter) and the Rebalanced Modality Gradient Modulation (ReGrad) strategy. The U-Adapter recognizes and suppresses unreliable detected regions within each modality, while the ReGrad strategy rebalances the convergence speed of all modalities by adaptively adjusting their gradients. The authors also establish the first large-scale benchmark for evaluating multi-modal FAS performance under domain generalization scenarios. Extensive experiments demonstrate that their method outperforms state-of-the-art methods.