Grounding-Aware Judge Reward Reinforcement Learning for Out-of-Domain Face Anti-Spoofing
Organizations: School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan, Hubei 430074, China · Mashang Consumer Finance Co., Ltd., Chongqing 400000, China
Abstract
Face Anti-Spoofing (FAS) must distinguish genuine faces from presentation attacks. A unified Multimodal Large Language Model (MLLM) is an attractive solution: it detects physical attacks (print, replay, mask) in one model and produces interpretable forensic evidence via Chain-of-Thought (CoT). A central obstacle to post-training such a model with reinforcement learning is that FAS reasoning has no rule-verifiable ground truth: only the binary verdict is checkable, while the reasoning chain cannot be verified by any hand-crafted rule. Optimizing the verdict alone is therefore a sparse, hackable signal. The model can reach the correct answer through a shortcut such as a false-rejection bias, which does not generalize. We address this with a grounding-aware judge reward. A 32B vision-language judge scores each reasoning chain against the image on evidence anchoring, logical coherence, and domain knowledge, turning the image itself into the verification signal. The reward is applied in a Group Relative Policy Optimization (GRPO) stage after CoT supervised fine-tuning. It is the judge, rather than the RL objective, that supplies the process supervision a verdict-only reward cannot. Our study starts from a model that already performs well In-Domain (ID) after CoT-based supervised fine-tuning and asks how to improve its Out-of-Domain (OOD) generalization. We find that grounding-aware rewards are an effective means to that end. We validate the method on OOD generalization, where the reward reduces ACER on two physical benchmarks (LCC-FASD 0.288 to 0.273, MSU-MFSD 0.431 to 0.409) while preserving ID accuracy (CelebA-Spoof 0.033 to 0.035). Against ResNet-50/101 and CLIP-ViT-B/16 baselines, the MLLM dominates ID and matches or surpasses the strongest baseline OOD. To our knowledge this is the first use of LLM-as-judge process supervision for FAS.
Figures & tables
| Model | CelebA-Spoof | LCC-FASD | MSU-MFSD |
|---|---|---|---|
| (ID) | (OOD) | (OOD) | |
| ResNet-50 | 0.134 | 0.322 | 0.486 |
| ResNet-101 | 0.132 | 0.334 | 0.487 |
| CLIP-ViT-B/16 [ 6 ] | 0.098 | 0.282 | 0.410 |
| MLLM CoT SFT (baseline) | 0.033 | 0.288 | 0.431 |
| MLLM judge reward RL (ours) | 0.035 | 0.273 | 0.409 |
| Variant | CelebA-Spoof | LCC-FASD | MSU-MFSD |
|---|---|---|---|
| (ID) | (OOD) | (OOD) | |
| CoT SFT (baseline) | 0.033 | 0.288 | 0.431 |
| no-judge (fmt acc) | 0.033 | 0.270 | 0.429 |
| judge | 0.029 | 0.281 | 0.417 |
| judge | 0.036 | 0.280 | 0.435 |
| judge , lr | 0.035 | 0.273 | 0.409 |
| Variant | CelebA-Spoof | LCC-FASD | MSU-MFSD |
|---|---|---|---|
| (ID) | (OOD) | (OOD) | |
| lr | 0.034 | 0.276 | 0.423 |
| lr , 2000 steps | 0.033 | 0.272 | 0.411 |
| 0.035 | 0.280 | 0.432 | |
| 0.030 | 0.285 | 0.428 | |
| best (lr , ) | 0.035 | 0.273 | 0.409 |
| Seed | CelebA-Spoof | LCC-FASD | MSU-MFSD |
|---|---|---|---|
| (ID) | (OOD) | (OOD) | |
| 42 | 0.035 | 0.273 | 0.409 |
| 123 | 0.036 | 0.277 | 0.404 |
| 2024 | 0.031 | 0.282 | 0.427 |
| mean std |
Face Presentation Attack Detection
Face Presentation Attack Detection identifies spoofing attempts, such as printed, replayed, masked, or digitally altered faces, targeting face recognition systems.
See all 33 papers