Face anti-spoofing (FAS) is crucial for multimedia AI security, yet existing methods often fail to disentangle domain-specific from domain-invariant features in cross-domain scenarios, undermining robustness and generalization—especially against unseen domains and sophisticated spoofing attacks. To address this, we propose CFB-FAS, a novel domain generalization framework that enforces both feature and decision boundary consistency, where stable cross-domain live-sample clustering supports reliable boundary learning. Specifically, we introduce: (1) Dynamic Soft K-Means Loss (DSK-Loss) to encourage compact clustering of positive samples across domains; (2) Boundary-aware Dynamic Center Loss (BDC-Loss), which enhances intra-class contraction and applies adversarial boundary regularization to better separate ambiguous real-spoof cases; and (3) Attention-Guided Feature Adaptive Normalization (AFAN), which couples style perturbation with attention-guided focusing to suppress domain-specific styles and emphasize content cues, enhancing robustness against cross-domain style variations and challenging presentation attacks such as unseen 3D mask spoofing. Experiments show CFB-FAS enhances robustness and security, outperforming state-of-the-art methods across five benchmarks, particularly under unseen domains and 3D mask attacks, while maintaining stable convergence.