Focused LRP: Explainable AI for Face Morphing Attack Detection

Explainability is critical for high-stakes AI systems: The paper demonstrates that AI systems used in security applications like automated border control need to show how they make decisions, not just what decision they make. This is essential when AI assists human operators who must understand and potentially override the system's judgment.

Current AI detection methods lack transparency: Deep neural networks can effectively detect morphed face images (a security threat where two faces are blended together), but they operate as "black boxes" that don't reveal which visual features trigger their alerts—a significant gap for responsible deployment in security contexts.

Pixel-level explanations can be objectively measured: The researchers developed a method called Focused Layer-wise Relevance Propagation (FLRP) that pinpoints exactly which parts of an image the AI uses to make decisions, and they created a rigorous testing framework to verify these explanations actually matter by removing highlighted features and measuring their impact on the AI's output.

Explainability methods perform better on uncertain or wrong decisions: FLRP was particularly effective at highlighting visible artifacts when the AI was uncertain or made mistakes—precisely the scenarios where human operators most need clear explanations to catch errors and maintain appropriate skepticism of automated systems.

Responsible AI requires both accuracy and interpretability: This work shows that building trustworthy AI for security applications requires combining detection capability with the ability to explain decisions in human-understandable terms, not treating these as separate concerns.