xCos: An Explainable Cosine Metric for Face Verification Task

Face verification systems need transparency for real-world deployment: The paper addresses a critical gap in AI trustworthiness—while deep learning models achieve high accuracy in face verification (used in security, surveillance, and phone unlocking), users and regulators cannot understand why the system made a particular match or rejection decision, which is essential for high-stakes applications.

The xCos method reveals which facial features drive verification decisions: Rather than treating face verification as a "black box," the proposed explainable cosine metric shows which specific parts of two faces the model considers similar and how much weight each region receives in the final decision, enabling auditing of potential biases.

Explainability doesn't require sacrificing accuracy: The research demonstrates that adding interpretability through the xCos module maintains competitive performance on standard benchmarks (LFW and others), suggesting that transparency and accuracy are not inherently at odds—a key finding for responsible AI implementation.

Attention mechanisms can be made visible and understandable: By making the model's attention weights explicit and learnable, the method allows stakeholders to see where the system "looks" when comparing faces, which is crucial for identifying whether the model relies on problematic features or demographic characteristics rather than legitimate facial identity markers.