Reproduce “Harnessing Chain-of-Thought Reasoning in Multimodal Large Language Models for Face Anti-Spoofing” on OpenResearch

About this paper

Honglu Zhang, Zhiqin Fang, Ningning Zhao, Saihui Hou, Long Ma, Renwang Pei, Zhaofeng He

A new large-scale multimodal dataset, FaceCoT, featuring Chain-of-Thought (CoT) annotations, is introduced alongside a CoT-Enhanced Progressive Learning (CEPL) strategy, enabling Multimodal Large Language Models (MLLMs) for robust and interpretable Face Anti-Spoofing (FAS). This approach yielded an average AUC increase of 4.06% and HTER reduction of 5.00% across 11 benchmark datasets, demonstrating enhanced cross-domain generalization and explicit decision rationales.

Key insights

• Chain-of-Thought annotations provide richer intermediate information, significantly enhancing task understanding, generalization, and robustness of MLLMs for FAS tasks, enabling systematic reasoning analysis.
• The two-stage CoT-Enhanced Progressive Learning (CEPL) strategy effectively manages task interference between reasoning and classification, fostering synergistic development of fine-grained visual feature extraction and high-level linguistic reasoning.
• Integrating visual and linguistic co-inference through MLLMs, supported by CoT data, substantially improves cross-domain generalization capabilities against diverse and previously unseen spoofing attack types.

Summary by alphaXiv.