SAFARI: Versatile and Efficient Evaluations for Robustness of Interpretability

AI explanations are vulnerable to tiny, imperceptible changes: The paper demonstrates that current explainable AI (XAI) methods often produce different explanations when inputs are slightly modified in ways humans cannot detect, undermining trust in these systems for high-stakes decisions.

Existing tools for testing explanation reliability are inadequate: Current methods for evaluating whether AI explanations are robust lack comprehensiveness and cannot effectively handle the diversity of different XAI techniques, meaning organizations may have false confidence in their interpretability measures.

New evaluation methods can identify which explanation techniques are most trustworthy: The SAFARI framework uses advanced optimization techniques to systematically test XAI methods and rank them by robustness, enabling practitioners to select more reliable explanation approaches for deployment.

Training methods matter for both accuracy and trustworthy explanations: The research shows that how AI models are trained affects not only their predictive accuracy but also the stability and reliability of their explanations, suggesting responsible AI requires attention to training practices.

Robustness testing can be made practical and efficient: By employing genetic algorithms and specialized statistical methods, the paper demonstrates that comprehensive robustness evaluation of explanations can be performed at scale, making it feasible for organizations to audit their AI systems before deployment.