Fig. 4: Average privacy risk (n=3) of the synthetic versions of the UW (a–d) and VUMC (e-h) datasets.
From: A Multifaceted benchmarking of synthetic electronic health record generation models
a, e Attribute inference risk. b, f Membership inference risk. c, g Meaningful identity disclosure risk. d, h Nearest neighbor adversarial accuracy (NNAA) risk. The risks associated with the real data are shown in the bottom red bars. The 95% confidence intervals are marked as thin horizontal black lines.
