ConceptSMILE: A Model-Agnostic Framework for Auditing Concept-Based Explainable AI

Summary

ConceptSMILE is a model-agnostic, perturbation-based auditing framework that evaluates the trustworthiness of concept-based explanations in explainable AI (XAI). While concept-based explanations make model reasoning more human-interpretable, their concept-level outputs are not automatically reliable. The method perturbs input regions, measures changes in concept responses, applies locally weighted adjustments, and fits an XGBoost surrogate model to quantify explanation fidelity. The authors—Mohadeseh Mollapour, Koorosh Aslansefat, Zeinab Dehghani, Bhupesh Kumar Mishra, Tejal Shah, and Zhibao Mian—evaluate the framework on retinal fundus images. Results show that MedSAM achieves stronger spatial attribution and the highest surrogate fidelity (R² = 0.8503), while a vision-language model (VLM) path demonstrates greater vessel faithfulness and stability under selected artifact conditions. The paper (arXiv:2607.09649) offers a practical approach for auditing whether concept-level AI explanations can be trusted, particularly in medical imaging applications.