ConceptSMILE Audits Trustworthiness of Concept-Based Explainable AI

This research introduces ConceptSMILE, a novel auditing framework designed to assess the trustworthiness of concept-based Explainable AI (XAI) explanations. While concept-based XAI aims to make AI reasoning more human-understandable, the reliability of these concept-level outputs is not guaranteed. ConceptSMILE addresses this by extending the perturbation-based logic of the existing SMILE framework, adapting it to audit human-understandable concept explanations rather than just feature or region attributions. The framework operates by perturbing input regions, measuring the resulting shifts in concept responses, applying locality weighting, and fitting an XGBoost surrogate model to approximate local concept behavior. It evaluates reliability across several dimensions, including attribution accuracy, surrogate fidelity, faithfulness, stability, and consistency. A practical evaluation on retinal fundus images, comparing MedSAM-derived visual concepts with VLM-based semantic concepts, demonstrated that reliability varies significantly across concepts and pathways. This highlights ConceptSMILE's utility as an independent audit layer for ensuring the trustworthiness of concept-based XAI.