Layer-Wise Prompt-Guided Interaction with Prototypical Contrastive Learning for Face Anti-Spoofing Generalization

Recently, Vision-Language Models (VLMs) have demonstrated remarkable success in Face Anti-Spoofing (FAS). However, most of these VLM methods adopt a CLIP-like in-teraction strategy, where image and text features interact only once at the final layer of their respective encoders. This strategy tends to emphasize global features while largely overlooking local information, which is crucial for FAS tasks, thereby compromising subsequent classification accuracy. To address this issue, we propose layer-wise prompts that facilitate cross-modal interaction at three layers, fully leveraging the potential of textual guidance. Furthermore, to enhance the generalization capability of the FAS model, we introduce a prototype-based contrastive loss to encourage a clear separation between the two classes during training. Experimental results on several datasets demonstrate that our method outperforms current state-of-the-art approaches, confirming its effectiveness and superiority.