Auditing Saliency Faithfulness Across Fitzpatrick Skin Types in Dermatology Foundation Models

Trustworthy deployment of foundation models in dermatology requires that both predictions and their explanations behave reliably across the populations the system will serve. Predictive-accuracy disparities across Fitzpatrick Skin Types (FST) are well documented; state-of-the-art classifiers perform substantially worse on FST V–VI (darker skin tone) than on I–II (lighter skin tone), yet it remains unknown whether the post-hoc saliency maps used to mediate clinician trust remain faithful across demographic subgroups. We audit explanation fidelity across Fitzpatrick groups for two dermatology foundation models, comparing a class-dependent method (Grad-CAM) against a class-agnostic one (attention rollout) on an external test set. Our audit incorporates a random-saliency control that separates the quality of a saliency ranking from an image's intrinsic sensitivity to perturbation, and we report the resulting fidelity per group together with the uncontrolled AUCs, so that the effect of the correction is visible. We find that under this experimental setting, insertion AUC has no measurement power in this domain, and no fidelity disparity across skin tone.