Minimal Information Disclosure Framework for Privacy-Preserving AI Verification

A recent paper published on arXiv presents a framework called minimal information disclosure (MID), aimed at assessing and reducing the information content of evidence presented to verifiers in AI verification. Identified as arXiv:2608.02774, the study tackles the trust boundary where verifiers need sufficient information to validate a claim without exposing sensitive aspects of the model, workload, or hardware. MID employs conditional mutual information to evaluate collateral leakage, indicating what is disclosed about protected attributes once the authorized outcome is known. This versatile framework supports various verification objectives, protected attributes, evidence sources, and deployment limitations. The authors test MID through four physical measurements and six verification tasks, utilizing three design variables: evidence channel, collection policy, and release transformation. The paper is classified as a cross-type announcement and can be accessed at https://arxiv.org/abs/2608.02774.