ConceptSMILE audits trustworthiness of concept-based explainable AI

Researchers introduce ConceptSMILE, a model-agnostic perturbation-based auditing framework designed to evaluate the reliability of concept-based explanations in explainable AI. The system extends perturbation logic from feature-level attribution to human-understandable concepts by measuring response shifts and fitting an XGBoost surrogate.

• Reliability is assessed through attribution accuracy, surrogate fidelity, faithfulness, stability, and consistency.
• Evaluation on retinal fundus images compares MedSAM-derived visual concepts with VLM-based semantic concepts.
• MedSAM achieves stronger spatial attribution and the highest surrogate fidelity ($R^2 = 0.8503$, $R_w^2 = 0.8465$).
• The VLM pathway demonstrates stronger vessel faithfulness and stability under selected artefact conditions.

ConceptSMILE provides an independent audit layer for evaluating the trustworthiness of concept-based XAI.