FAS-R1: A Unified Multi-Task MLLM for Reasoning Face Anti-Spoofing
Published in arXiv, 2026
FAS-R1 combines long-chain-of-thought supervised fine-tuning with face-anti-spoofing-specific GRPO post-training to unify authenticity classification, attack-type recognition, and spoof-region localization in a single multimodal reasoning framework.
Recommended citation: Wang, H., Shi, Y., Li, H., Huo, Y., Feng, J., & Yu, Z. (2026). FAS-R1: A Unified Multi-Task MLLM for Reasoning Face Anti-Spoofing. arXiv, 2607.26432. https://arxiv.org/abs/2607.26432