Detecting Distribution-Based Attribute Inference Risk in Synthetic EHRs

Synthetic electronic health records promise to accelerate biomedical research by enabling privacy-preserving data sharing. Realizing this potential requires ensuring that synthetic data do not expose patients to privacy risks. A key concern is attribute inference attacks, in which partial knowledge about a patient is used to infer additional attributes. We present StatAIT, a statistically grounded attribute inference test (AIT) that provides principled, patient-level risk assessments. StatAIT estimates confidence intervals for the expected risk over random subsets of known attributes and models attacks by filtering the synthetic data to exploit conditional distributions. StatAIT also includes a post-processing step that reduces risk by removing synthetic records. In experiments on MIMIC-III and simulated Medicare claims data, StatAIT detects replicated real patients and uncovers more subtle patterns that existing AITs miss, showing that larger synthetic datasets substantially increase privacy risk. In contrast, baseline AITs show little variation across synthetic dataset sizes and indicate substantially lower privacy risk overall. StatAIT's post-processing is already effective when removing less than 5% of the synthetic records. By offering greater sensitivity, StatAIT raises the bar for evaluating attribute inference risk in synthetic data, while its post-processing facilitates practical use.