Reliability-aware vision-language face anti-spoofing via progressive semantic reorganization

Abstract

Face Anti-Spoofing (FAS) is a crucial task for securing face recognition systems, yet its cross-domain generalization remains challenging. Recently, vision-language methods built upon pretrained models such as CLIP have shown promising performance in addressing these cross-domain scenarios. Nevertheless, existing approaches are still constrained by two primary limitations. First, fragmented spoofing cues at the patch level are typically incorporated into vision-language interactions without explicit organization, which hinders the effective exploitation of local discriminative evidence. Second, the training process generally overlooks reliability variations among samples caused by image degradation and the varying strengths of spoofing cues, allowing samples with low reliability to adversely affect the optimization process. To address these challenges, we propose Reliability-Aware Progressive Semantic Reorganization for Face Anti-Spoofing (RPSR-FAS). The framework comprises two core components: Progressive Semantic Reorganization (PSR), which reorganizes scattered local cues into more discriminative representations under the guidance of task-specific, fine-grained semantic descriptions; and Discriminative Reliability Learning (DRL), which estimates sample-level discriminative reliability via a two-stage training strategy and adaptively adjusts the contribution of different samples to the classification loss. Under a unified cross-domain evaluation protocol, RPSR-FAS is trained solely on CelebA-Spoof and directly evaluated on four unseen target domains, achieving superior average performance across these domains. Experimental results show that RPSR-FAS achieves an average Half Total Error Rate (HTER) of 4.62%, increases the average Area Under the ROC Curve (AUC) to 98.56%, and achieves an average True Positive Rate at a 1% False Positive Rate (TPR@FPR=1%) of 89.98%.