Disparate privacy risks from medical AI.

Patient-level privacy risks from medical AI are significantly higher than aggregate metrics suggest, with near-perfect attack success rates for individual patients.

• Why it matters: Protecting patient privacy is critical as AI models often contain sensitive health data, yet current assessments overlook individual risk, especially for patients contributing multiple records.
• What they did: The study conducted one of the first patient-level privacy audits using membership inference attacks across various medical datasets, analyzing how individual data can be exposed.
• The result: Findings reveal that attack success increases with model capacity and disproportionately threatens underrepresented groups, highlighting the need for tailored privacy risk mitigation strategies.