Researchers identify privacy risks in medical AI models

Researchers have identified significant privacy vulnerabilities in medical AI models. A study involving the Technical University of Munich, Imperial College London, and the Hasso Plattner Institute demonstrates that Membership Inference Attacks (MIAs) can be used to identify specific patients within certain datasets with nearly 100% probability.

While previous research suggested that medical AI models were generally safe from such attacks by focusing on average risks, this new investigation highlights that individual patient data remains highly susceptible to exploitation through relatively simple methods.