PA-FAS: Towards Interpretable and Generalizable Multimodal Face Anti-Spoofing via Path-Augmented Reinforcement Learning

Conference: AAAI 2026 arXiv: 2511.17927 Code: None Area: Reinforcement Learning Keywords: Face Anti-Spoofing, Multimodal Fusion, Domain Generalization, Reinforcement Learning, Interpretability

TL;DR

This paper proposes PA-FAS, a framework that addresses two critical bottlenecks of the SFT+RL paradigm in multimodal FAS — insufficient reasoning path diversity and reasoning shortcut — via a Reasoning Path Augmentation strategy and an answer shuffling mechanism, achieving the first unified solution for multimodal fusion, domain generalization, and interpretability simultaneously.

Background & Motivation

Problem Definition

Face Anti-Spoofing (FAS) aims to distinguish genuine faces from spoofing presentations (printed photos, replay videos, 3D masks, etc.), serving as a critical safeguard for the security of face recognition systems. Multimodal FAS leverages RGB, depth, and infrared modalities to improve detection accuracy and robustness.

Three Research Gaps in Existing Methods

Insufficient domain generalization: Most existing methods are designed for single-modality settings with limited cross-domain generalization capability.

Lack of interpretability in multimodal methods: Despite superior performance, multimodal methods lack explicit interpretability mechanisms to identify spoofing cues in depth and infrared modalities.

Limitations of MLLMs in domain generalization scenarios: Existing MLLM-based FAS methods (e.g., FaceCoT, SHIELD) demonstrate strong reasoning ability but overlook generalization and cannot handle cross-modal cue integration.

Failure Analysis of the SFT+RL Paradigm (Core Motivation)

The paper conducts an in-depth analysis of the failure mechanism when applying the SFT+RL paradigm to multimodal FAS:

Problem 1: Insufficient reasoning path diversity - Multimodal FAS datasets typically provide only simple binary labels, lacking language-level annotations of key visual cues. - The SFT stage suffers from task homogeneity and limited data, causing the model to overfit rigid patterns. - The RL stage lacks effective feedback and exploration space — most samples receive extreme rewards (all 1s or all 0s), providing no informative intermediate signals. - Empirical evidence: Models fine-tuned on the original data exhibit a nearly linear growth in accumulated invalid samples.

Problem 2: Reasoning shortcut - Even when reasoning chains are augmented via conventional data augmentation, the model bypasses the chain-of-thought (CoT) by directly predicting answers from images. - Evidence: Randomly replacing the reasoning text in the SFT stage with CoT from other samples yields nearly identical performance, demonstrating that the model relies on visual prediction while ignoring the CoT. - This leads to overconfidence and significantly narrows the exploration space in the RL stage.

Method

Overall Architecture

PA-FAS follows a four-step pipeline: (a) low-level data annotation → (b) high-level data annotation to generate CoT → (c) positive-negative random path sampling to expand reasoning paths → (d) two-stage training with SFT (with answer shuffling) + RL (GRPO).

Key Designs

1. Reasoning Path Augmentation

The core innovation, which systematically expands reasoning paths by constructing a structured reasoning tree:

• Reasoning tree construction: Based on a fine-grained hierarchical taxonomy of FAS tasks (sunburst diagram in Figure 4), a formal reasoning tree \(\mathcal{T} = (\mathcal{V}, \mathcal{E})\) is established:

  • \(\mathcal{V}\): set of reasoning nodes (each node represents a semantic category or logical decision unit)
  • \(\mathcal{E}\): set of directed edges
  • Path \(\mathcal{P} = (v_1, v_2, \ldots, v_n)\): a complete reasoning chain from the root to the target leaf node

• Positive-Negative Random Path Sampling (PNRPS): The core mechanism of Algorithm 1:

  • Single-node operation constraint: Each node permits at most one forward exploration step \((+, v)\) and one backward reflection step \((-, v)\) to avoid redundant traversal.
  • Path length constraint: \(L_{\max} = \alpha(D-1)\), where \(D\) is the maximum classification depth and \(\alpha > 1\) is a tunable scaling factor.
  • Semantic consistency: Each node is associated with predefined CoT clause templates, which are concatenated along the path to generate the final reasoning text.
  • Structured sampling: Rule-based depth-first traversal randomly samples \(N\) valid paths utilizing arbitrary combinations of RGB, IR, and DEPTH modalities.

• Data augmentation effect: $\(\{(x_i, \ell_i)\}_{i=1}^{M} \rightarrow \bigcup_{i=1}^{M}\{(x_i, \text{CoT}(\mathcal{P}_i^{(j)})) \mid j=1,\ldots,N\}\)$ Starting from 800 annotated samples, \(N=50\) reasoning paths are generated per sample, yielding approximately 40,000 structurally diverse and semantically consistent augmented samples.

2. Answer Shuffling

A key design to address the reasoning shortcut problem:

• Motivation: The coupling of a single-task objective with rich reasoning paths causes the model to directly predict answers from visual input, bypassing the reasoning chain.
• Method: During the SFT stage, the final answer in each CoT is randomly replaced with the answer from another sample.
• Effect: This forces the model to focus on learning diverse reasoning paths rather than memorizing answers, thereby preserving sufficient exploration space for the RL stage.

3. RL Stage (GRPO)

Group Relative Policy Optimization is adopted for policy optimization:

• For each question-answer pair \((q, a)\), \(G\) responses are sampled from the old policy.
• Reward definition: \(\mathcal{R} = \mathcal{R}_{\text{format}} + \mathcal{R}_{\text{classification}}\)
• Relative advantage estimation: \(\hat{A}_{i,t} = \frac{\mathcal{R}_i - \text{mean}(\{\mathcal{R}_i\})}{\text{std}(\{\mathcal{R}_i\})}\)
• The KL divergence term \(D_{KL}(\pi_\theta \| \pi_{\text{ref}})\) is removed under data-scarce conditions to avoid suppressing exploration.

Loss & Training

• Backbone model: Qwen2.5VL-3B
• SFT and RL stages: Both trained for 500 steps with a constant learning rate of 1e-6.
• Data: Only approximately 800 high-quality structured reasoning path samples are required.
• Evaluation datasets: WMCA, CASIA-SURF, CASIA-CeFA, PADISI.

Key Experimental Results

Main Results (Protocol 1: Full-modality Cross-dataset Testing)

Method	Category	Avg. HTER (%)↓	Avg. AUC (%)↑
FLIP	Unimodal DG	16.11	90.83
MMDG	Multimodal DG	22.93	84.19
DADM	Multimodal DG	13.63	92.96
Qwen2.5-VL-3B (zero-shot)	Interpretable Multimodal DG	33.46	69.36
Qwen2.5-VL-3B-SFT	Interpretable Multimodal DG	23.25	79.32
Qwen2.5-VL-3B-SFT+GRPO	Interpretable Multimodal DG	34.37	68.68
PA-FAS (Ours)	Interpretable Multimodal DG	15.21	89.13

Key observation: Naïve SFT+GRPO performs even worse than SFT alone (HTER increases from 23.25% to 34.37%), validating the authors' failure analysis. PA-FAS reduces HTER to 15.21%, significantly outperforming all interpretable models.

Ablation Study (Missing Modality Scenario, Protocol 2)

Method	Missing D HTER↓	Missing I HTER↓	Missing D&I HTER↓	Avg. HTER↓
DADM	21.56	20.82	22.61	21.66
Qwen2.5-VL-3B-SFT	23.25	23.25	23.25	23.25
PA-FAS	15.68	17.32	14.67	15.85

PA-FAS maintains superior performance under missing modality conditions, demonstrating that the reasoning path augmentation strategy effectively exploits complementary multimodal information.

Key Findings

1. SFT+RL collapse validated: On datasets with only binary labels, SFT+GRPO degrades performance (HTER: 23.25% → 34.37%).
2. Reasoning shortcut confirmed: Replacing reasoning text leaves model performance unchanged, proving the model completely ignores the CoT reasoning chain.
3. Importance of path diversity: Datasets with diverse reasoning paths significantly outperform those with single-path data under both SFT and SFT+RL settings.
4. Data efficiency: Only 800 samples with path augmentation suffice to surpass methods trained on 35,000 raw data samples.
5. Limited source domain scenario (Protocol 3): PA-FAS achieves an average HTER of 9.22% (as low as 0.15% on CW→PS), demonstrating strong domain generalization capability.

Highlights & Insights

1. In-depth diagnosis of SFT+RL failure: Rather than simply proposing a method, the paper first rigorously diagnoses the failure causes through experiments (cumulative invalid sample analysis and reasoning text replacement experiments).
2. Elegant design of reasoning path augmentation: A formal reasoning tree is constructed based on a hierarchical taxonomy, and positive-negative random path sampling expands the reasoning space at minimal cost.
3. Counter-intuitive answer shuffling design: Deliberately introducing incorrect answers during the SFT stage forces the model to learn the reasoning process — a counter-intuitive yet highly effective design.
4. First unification of three objectives: An integrated solution for multimodal fusion, domain generalization, and interpretability.
5. Exceptional data efficiency: SOTA-level performance is achieved with only 800 samples.

Limitations & Future Work

1. Backbone model constraint: Only Qwen2.5VL-3B (3B parameters) is used; larger models may yield further improvements.
2. Manual reasoning tree construction: The hierarchical taxonomy requires domain expert knowledge; automated construction remains a future direction.
3. Annotation cost: Although only 800 samples are needed, their CoT annotation still requires non-trivial human effort.
4. Evaluation scope: Validation is conducted primarily on four datasets; more diverse attack types and scenarios remain to be explored.
5. Inference latency: MLLM-based methods may not meet the real-time deployment requirements.

Related Work & Insights

• "SFT Memorizes, RL Generalizes" theory: The paper builds upon Chu et al.'s SFT-memorization + RL-generalization theory, while noting that this theory requires new adaptations for data-scarce multimodal scenarios.
• GRPO applied to FAS: This work represents the first application of GRPO to the FAS task, demonstrating the feasibility of rule-based group advantage strategies under low training cost conditions.
• Deeper understanding of reasoning shortcuts: The answer shuffling mechanism provides an important reference for other visual understanding tasks combining SFT and RL.

Rating

• Novelty: ⭐⭐⭐⭐⭐ — The combination of reasoning path augmentation and answer shuffling is highly original, with rigorous analysis of the failure mechanism.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Three protocols, ablation studies, and diverse baselines constitute a rigorous experimental design.
• Writing Quality: ⭐⭐⭐⭐ — Logic is clear and motivation is well-justified, though some notation is slightly complex.
• Value: ⭐⭐⭐⭐⭐ — The failure analysis of the SFT+RL paradigm and the proposed solutions have broad reference value.

Related Papers

• [AAAI 2026] Distilling Deep Reinforcement Learning into Interpretable Fuzzy Rules: An Explainable AI Framework
• [AAAI 2026] Think, Speak, Decide: Language-Augmented Multi-Agent Reinforcement Learning for Economic Decision-Making
• [ACL 2026] Language-Coupled Reinforcement Learning for Multilingual Retrieval-Augmented Generation
• [AAAI 2026] TAdaRAG: Task Adaptive Retrieval-Augmented Generation via On-the-Fly Knowledge Graph Construction
• [ICLR 2026] Spotlight on Token Perception for Multimodal Reinforcement Learning