Overview
- Researchers at the Technical University of Munich measured per-record vulnerability across seven large clinical datasets and found membership inference attacks could identify some individual patients with near-perfect accuracy.
- The highest risk fell on underrepresented patients, including those with rare diseases, minority ethnicities, unusual imaging protocols, or atypical insurance or sex attributes.
- The team trained roughly 200 model variants per dataset and observed that larger model capacity produced many more highly vulnerable records.
- Membership inference attacks let an outsider test whether a specific record was in a training set without seeing the original database, which can expose sensitive health conditions and enable harms such as discrimination, blackmail, or targeted marketing.
- The authors warn that dataset-average privacy metrics mask these concentrated risks and recommend per-record risk assessments, verifiable mitigation methods, and strict access controls before medical AI models are shared or deployed.