Abstract: Face recognition remains vulnerable to presentation attacks, calling for robust Face Anti-Spoofing (FAS) solutions. Recent MLLM-based FAS methods reformulate the binary classification task as the generation of brief textual descriptions to improve cross-domain generalization. However, their generalizability is still limited, as such descriptions mainly capture intuitive semantic cues while struggling to perceive fine-grained visual patterns. To address this limitation, we incorporate external visual tools into MLLMs to encourage deeper investigation of subtle spoof clues. Specifically, we propose the Tool-Augmented Reasoning FAS (TAR-FAS) framework, which reformulates the FAS task as a Chain-of-Thought with Visual Tools (CoT-VT) paradigm, allowing MLLMs to begin with intuitive observations and adaptively invoke external visual tools for fine-grained investigation. To this end, we design a tool-augmented data annotation pipeline and construct the ToolFAS-16K dataset, which contains multi-turn tool-use reasoning trajectories. Furthermore, we introduce a tool-aware FAS training pipeline, where Diverse-Tool Group Relative Policy Optimization (DT-GRPO) enables the model to autonomously learn efficient tool use. Extensive experiments under a challenging one-to-eleven cross-domain protocol demonstrate that TAR-FAS achieves SOTA performance while providing fine-grained visual investigation for trustworthy spoof detection.
The Generalization Gap in Face Anti-Spoofing
Face recognition systems are under constant threat from presentation attacks, ranging from simple printed photos and high-definition video replays to sophisticated 3D masks. While many Face Anti-Spoofing (FAS) solutions exist, they frequently suffer from a "generalization gap." A model trained on paper photos in a controlled laboratory environment often fails when encountering 3D masks in real-world lighting conditions.
Recent attempts to bridge this gap have utilized Multimodal Large Language Models (MLLMs) to generate descriptive text, moving beyond simple binary classification. However, MLLMs are often "blind" to low-level visual artifacts. While they excel at identifying high-level semantic cues (e.g., "this person is wearing a mask"), they struggle to perceive the subtle moiré patterns of a screen or the microscopic texture of printer paper. The TAR-FAS framework addresses this by providing MLLMs with specialized "magnifying glasses" to investigate these fine-grained clues.
The TAR-FAS Framework
The core of the proposed framework is the Chain-of-Thought with Visual Tools (CoT-VT) paradigm. This shifts the model's role from an intuitive guesser to a digital forensic investigator. Instead of a one-shot prediction, the model follows a multi-turn process: it analyzes the image, forms a hypothesis ("Think"), calls a specific tool ("Act"), and observes the output to refine its judgment.
The framework utilizes a diverse toolbox of external visual operators:
- Zoom-In: For inspecting local details.
- Texture-based (LBP): To identify unnatural skin patterns.
- Frequency-based (FFT/Wavelet): To detect periodic artifacts like moiré patterns from screens.
- Structure-based (Edge Detection/HOG): To find mask boundaries or rigid surfaces.
Training and Data: ToolFAS-16K
To support this reasoning-heavy approach, the authors developed the ToolFAS-16K dataset. Comprising 16,172 images, each entry includes multi-turn reasoning trajectories that serve as a blueprint for effective tool use. The dataset was created using an expert-guided annotation pipeline where lightweight models, specialized in tool outputs, guided a larger MLLM to generate high-quality reasoning traces.
To ensure the model learns to use these tools efficiently without becoming "lazy" (e.g., over-relying on a single tool), the authors introduced Diverse-Tool Group Relative Policy Optimization (DT-GRPO). This reinforcement learning strategy rewards the model for correct judgments while encouraging the use of diverse, relevant tools.
Results and Interpretability
TAR-FAS sets a new benchmark for cross-domain FAS performance. Evaluated under a rigorous "one-to-eleven" protocol—where a model is trained on one dataset and tested on 11 others—it achieved an average Half Total Error Rate (HTER) of 7.54%. This is a significant improvement over the previous state-of-the-art MLLM method, I-FAS, which recorded 11.30%.
Beyond performance, TAR-FAS offers unprecedented interpretability. Unlike "black-box" neural networks, it provides human-readable reasoning for its decisions. A user can follow the model's logic: "I used the FFT tool, observed periodic patterns indicative of a screen, and therefore concluded this is a replay attack."
Conclusion and Future Work
While TAR-FAS represents a major step forward in trustworthy spoof detection, it faces challenges regarding inference latency due to its multi-turn reasoning process. Currently, it is best suited for high-security forensic applications rather than instant mobile unlocking. Future research will focus on enabling MLLMs to develop or fine-tune their own visual operators and exploring larger base models to further enhance generalization.
Summary updated