Foundation models struggle with cross-dataset face attack detection

A new study published on arXiv investigates the effectiveness of foundation models (FMs) for face presentation attack detection (PAD). The research found that while LoRA-based adaptation can achieve low intra-dataset error rates, it significantly struggles with cross-dataset generalization. This suggests that the pretrained representations and the adaptation dataset itself are more critical for generalization than lightweight adaptation strategies like LoRA. The study evaluated 32 different FMs, noting that zero-shot prompting performed poorly across various model families. AI

IMPACT Highlights limitations in current foundation model adaptation techniques for real-world generalization in security applications.

RANK_REASON Research paper detailing limitations of foundation models for a specific task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• ACER
• CASIA-FASD
• Face Presentation Attack Detection Using Deep Background Subtraction
• foundation model
• LoRA
• MCIO benchmarks
• MSU-MFSD
• OULU-NPU
• replay attack

• paper
• model release

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