Benchmarking Vision Foundation Models for Domain-Generalizable Face Anti-Spoofing
Researchers have developed a new vision-only baseline for Face Anti-Spoofing (FAS) that demonstrates superior performance and efficiency compared to existing multimodal approaches. The study systematically benchmarks 15 pre-trained vision models, finding that self-supervised models like DINOv2 with Registers are particularly effective at capturing subtle spoofing cues. When combined with specific data augmentation techniques and loss functions, this vision-only approach achieves state-of-the-art results on challenging cross-domain FAS protocols, even under data-constrained conditions. AI
IMPACT Optimizes vision-only models for security applications, potentially reducing computational costs for facial recognition systems.
