Medical AI Is Most Likely to Leak Your Data If You're Underrepresented
Researchers managed to determine whether a specific individual was part of a diagnostic model's training data, with near-certainty for some patients. The risk climbs sharply for underrepresented groups, whether by disease type or ethnic background. A model built to help everyone ends up offering the weakest privacy protection to the very people it knows least about.
practitioners › Study published in Nature, June 24, 2026, membership inference attacks tested on real clinical datasets.
www.nature.com →