Disparate Privacy Risks from Medical AI: Patient-level Auditing
This presentation investigates the often-overlooked patient-level privacy risks in medical AI deployments, moving beyond aggregate metrics to explore individual vulnerabilities.
A Google TechTalk, 2025-04-30, presented by Moritz Knolle Privacy in ML Seminar. ABSTRACT: Medical AI models promise to improve global access to high-quality diagnostics. However, their deployment also entails privacy risks. Prior research has measured success rates of privacy attacks such as membership inference attacks -in aggregate- across records rather than at the level of individual records or patients. Thus, the risk faced by individual patients, who often contribute multiple records to a training dataset, is poorly understood. This talk will present findings from our recent research on patient-level privacy auditing. Our findings indicate that deploying medical AI models may pose high privacy risks to individual patients and that under-represented patient groups might be disparately at risk.