Analyzing the effect of reasoning-based supervision on face anti-spoofing

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics Security

Background:

• Face anti-spoofing (FAS) is vital for secure face recognition against attacks like photos and masks.
• Current FAS methods often lack interpretability, functioning as black-box models.
• Explainable AI (XAI) is emerging to provide insights into model decisions.

Purpose of the Study:

• To investigate explainable face anti-spoofing (FAS) using vision-language models (VLMs).
• To analyze the impact of natural language explanations on FAS model behavior and generalization.
• To develop and evaluate an explanation-augmented benchmark for controlled studies.

Main Methods:

• Constructed an explanation-augmented benchmark by adding captions to four standard FAS datasets (MSU-MFSD, CASIA-FASD, Replay-Attack, OULU-NPU).
• Utilized GPT-4o API for generating vanilla and reasoning-structured captions.
• Employed a dual-objective training strategy combining spoof classification and explanation generation losses.

Main Results:

• Reasoning-style captions enhanced FAS detection performance and domain generalization in several scenarios.
• Explanation-based supervision, particularly with reasoning captions, can positively influence model behavior.
• Inductive biases from explanations may degrade performance if cues misalign with unseen attacks.

Conclusions:

• Explanations in FAS serve as interpretable outputs and controllable training signals.
• Natural language explanations can shape the generalization capabilities of FAS systems.
• Further research is needed to balance explanation benefits with potential biases for robust FAS.