Foundation Models for Face Presentation Attack Detection: A Unified Linear-Probing Benchmark

TL;DR - A unified linear-probing benchmark of 24 frozen vision encoders finds that foundation models encode useful face presentation-attack signals, but their performance transfers inconsistently across datasets due to domain shift.

• Frozen encoders paired with only a linear classifier achieved strong intra-dataset detection performance.
• InternViT-6B had the lowest mean intra-dataset error.
• CLIP ViT-B/32 offered the best cross-dataset transfer–compute trade-off among the evaluated probes.
• Model scale helped within some families, but architecture and pretraining mattered more; explicit adaptation remains necessary for robust transfer.