CPG-PAD: Concept-Informed Prompts Guided Presentation Attack Detection

CPG-PAD targets the poor cross-domain generalization of face presentation attack detection under changes in sensors, lighting, and attack materials. It uses a Visual Concept-driven Enhancement module with XAI techniques to discover PAD-relevant concepts and produce localized heatmaps. A Prompt-based Concept Injection mechanism then incorporates these concepts into the prompt space through a Visual-Prompt Decoder and concept-mapping loss. The paper reports consistent state-of-the-art cross-domain performance across nine benchmark datasets under multi-source, limited-source, and single-source settings, although the supplied abstract does not provide quantitative results.