Foundation models show promise for face attack detection but struggle with cross-dataset transfer

Researchers have investigated the effectiveness of foundation models for face presentation attack detection (PAD) by evaluating 24 frozen encoders using a unified linear-probing protocol. The study found that while these models contain PAD-relevant information accessible with minimal training, this information does not reliably transfer across different datasets due to domain shift. Model scale showed benefits within certain families, but performance was also influenced by architecture and pretraining methods. InternViT-6B demonstrated the lowest intra-dataset error, while CLIP ViT-B/32 offered the best cross-dataset transfer-compute trade-off among the tested probes. AI

IMPACT Investigating foundation models for face attack detection highlights the need for explicit adaptation to overcome domain shift challenges in real-world applications.

RANK_REASON The cluster is based on an academic paper detailing research findings on foundation models for a specific AI task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CASIA-FASD
• CLIP ViT-B/32
• Face Presentation Attack Detection Using Deep Background Subtraction
• foundation model
• Hugging Face
• InternViT-6B
• MCIO benchmark
• MSU-MFSD
• OULU-NPU
• replay attack

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →