CHSPR Seminar | AI-Generated Synthetic Data Does More Harm than Good for Patient Privacy

Speaker: Bret Nestor, UBC Centre for Health Services and Policy Research Synthetic medical data is increasingly being adopted to protect patients' privacy while increasing data accessibility for researchers. This talk explores the "no free lunch" theorem as it pertains to synthetic clinical data. Synthetic data land on a spectrum between useful and non-private, or non-useful and private. We audit for private patient information by conducting membership inference attacks against a variety of privacy preservation defenses. We find that synthetic data harm the most vulnerable patients more often than when abstaining from any privacy protections. Simple privacy audits can be misleading. Occasionally, we find an audit task that says synthetic data protects patient privacy. Concerningly, for any set of synthetically generated clinical data, we find at least one audit task that reveals more patient information than algorithms which use unprotected patient data. We conclude that synthetic data are not a viable option for protecting patients' right to privacy. Differential privacy continues to be the best way to protect patient privacy. Bret Nestor is a computer scientist whose research focuses on machine learning for clinical time-series. He is keenly interested in the questions "What is the shelf life of an algorithm?" and "For whom does this algorithm work?". He completed his PhD at the University of Toronto and Vector Institute under the supervision of Dr. Marzyeh Ghassemi (MIT) and Dr. Anna Goldenberg. He completed two postdoctoral positions at the University of Washington, and Harvard University before joining the Regulatory Science Lab at the UBC School of Population and Public Health.