Spoof Trace Discovery for Deep Learning Based Explainable Face Anti-Spoofing

Explainability is critical for user trust: Current face anti-spoofing systems can accurately detect fake faces but fail to explain why they reject someone, which "undermines trustworthiness and causes user confusion." This highlights that high accuracy alone is insufficient for responsible AI deployment in security-sensitive applications.

A new benchmark for evaluating explanations was needed: The paper introduces the first X-FAS (eXplainable Face Anti-Spoofing) benchmark with expert-annotated spoof traces, addressing the lack of standardized methods to assess whether AI explanations are actually reliable and meaningful rather than just plausible-sounding.

The proposed SPTD method discovers interpretable "spoof concepts": Rather than providing opaque decision rules, SPTD identifies specific visual patterns that indicate spoofing attempts (termed "spoof traces"), allowing the system to show users concrete evidence for its decisions—for example, pointing to specific artifacts that reveal a presentation attack.

Explainability methods require rigorous validation: The paper demonstrates that previous XAI methods applied to face anti-spoofing needed quantitative and qualitative comparison against expert-annotated ground truth, suggesting that explanations cannot be assumed reliable without systematic evaluation against human judgment.

Responsible AI requires domain-specific solutions: Rather than applying generic explainability techniques, this work shows that security-critical applications like face anti-spoofing need purpose-built approaches that discover domain-relevant concepts (spoof indicators) to generate trustworthy explanations.