TL;DR
This paper introduces a grounding-aware judge reward to improve the out-of-domain (OOD) generalization of Multimodal Large Language Models (MLLMs) for face anti-spoofing. By using a 32B vision-language model as a "process supervisor" during reinforcement learning, the authors ensure that the model's reasoning is actually anchored in image evidence rather than relying on "hackable" shortcuts or hallucinations.
What Problem Does This Paper Address?
Face Anti-Spoofing (FAS)—distinguishing real faces from physical attacks like printed photos or video replays—is notoriously difficult to generalize. A model trained on one dataset (In-Domain) often fails miserably when deployed in a new environment (Out-of-Domain) because it latches onto spurious correlations, like specific lighting or background textures, rather than the actual physical artifacts of a spoof.
While MLLMs using Chain-of-Thought (CoT) reasoning are promising because they provide interpretable evidence, they suffer from a "verification gap." In math or coding, RL is easy because the answer is objectively right or wrong. In FAS, the reasoning chain (e.g., "I see a moiré pattern") has no rule-verifiable ground truth. If you only reward the final binary verdict (Real vs. Spoof), the model often learns to "hallucinate" evidence to justify a guess, leading to poor generalization.
Methods
The authors propose a two-stage training pipeline:
- CoT Supervised Fine-Tuning (SFT): A Qwen2.5-VL-7B model is fine-tuned on high-quality, human-verified reasoning chains that describe physical artifacts.
- Grounding-Aware RL (GRPO): The model undergoes Group Relative Policy Optimization. Instead of just rewarding the correct final answer, they introduce a 32B vision-language judge to score the reasoning process.
The judge evaluates three specific dimensions on a 1–5 scale:
- D1: Evidence Anchoring: Does the cited evidence (e.g., "screen frame") actually exist in the image?
- D2: Logical Coherence: Does the path from observation to verdict make sense, or is there a logical leap?
- D3: Domain Knowledge: Is the FAS terminology and judgment rule applied correctly?
Crucially, the judge is "anchored" to the ground-truth label. It doesn't ask "Is this a spoof?" but rather "Given that this is a spoof, does this reasoning honestly and professionally support that conclusion?"
Datasets
The models were trained and evaluated on three major physical attack benchmarks:
- CelebA-Spoof: Used for In-Domain (ID) training and evaluation.
- LCC-FASD: A large crowd-collected dataset used for OOD evaluation.
- MSU-MFSD: A mobile-based dataset used for OOD evaluation, representing a significant domain shift.
- MS-MFAD: The source of the 14K high-quality CoT annotations used for the initial SFT stage.
Key Insights and Ideas
- The Image is the Label: Since there is no text-based ground truth for reasoning, the image itself must serve as the verification signal. A powerful enough vision-language model can act as a proxy for a human expert to check this grounding.
- Verdict-Only Rewards are Hackable: Rewarding only the final answer encourages "false-rejection bias," where the model becomes overly sensitive to any noise, calling it a "spoof" just to be safe. This works ID but fails OOD.
- Process vs. Outcome: The judge provides "process supervision." By penalizing hallucinations (D1) and logical gaps (D2), the RL stage forces the model to move away from memorized shortcuts and toward generalizable forensic reasoning.
- Sequence-Level Importance Sampling: The authors found that standard token-level RL (like PPO) was too noisy for long reasoning chains. They adopted a sequence-level approach (GSPO) to stabilize training.
Key Results
The grounding-aware RL model achieved the best balance between ID accuracy and OOD generalization:
- OOD Performance: On the LCC-FASD benchmark, the method reduced the Average Classification Error Rate (ACER) from 0.288 to 0.273. On the harder MSU-MFSD benchmark, it dropped from 0.431 to 0.409.
- ID Preservation: The model maintained a very low ACER on CelebA-Spoof (0.035), significantly outperforming traditional baselines like ResNet-50 (0.134) and CLIP-ViT-B/16 (0.098).
- Ablation Study: Removing the judge reward and using only "accuracy + format" rewards failed to improve performance on the harder MSU-MFSD benchmark, proving that the judge's grounding signal is the "active ingredient" for generalization.
Limitations and Future Work
- Modest Gains: The authors acknowledge that the OOD improvements, while consistent, are limited. They attribute this to a "grounding-capability bottleneck" in the 32B judge itself.
- Judge Reliability: The performance of the system is strictly bounded by how well the judge can perceive fine-grained artifacts. As vision-language models become more powerful (e.g., 70B+ models or specialized forensic models), the authors expect even larger gains.
- Computational Cost: Training with a 32B judge in the loop requires significant GPU resources (the study used multiple NVIDIA H100s).
Summary updated