New Framework Audits Explainable AI for Robustness and Fidelity

A new formal methodological framework has been proposed for auditing Explainable AI (XAI) methods, specifically focusing on their robustness and fidelity. The motivation stems from observed variability in explanations generated by tools like SHAP and LIME when inputs are slightly perturbed, raising concerns about their trustworthiness, particularly in critical applications. The auditing protocol measures two key properties: robustness, which assesses the stability of an explanation under minor input changes, and fidelity, which verifies if the features highlighted as important by the explainer genuinely drive the model's prediction. These measures are combined into a single "Trust Score." Initial application of this protocol to a multi-sectoral dataset revealed concerning results: models with high predictive accuracy (AUC > 0.99) often yielded numerically degenerate or uninformative explanations, and fidelity scores lost their discriminative power when models were overfitted. This strongly suggests that auditing XAI outputs is not optional but essential, especially when these explanations inform decisions in sensitive areas.