SynPAR for Empirical Membership Inference Auditing of Synthetic Radiomic and Embedding Data

To develop and evaluate SynPAR, a framework for reproducible membership-inference auditing and reporting of synthetic radiomic and embedding data. We evaluated feature releases from eight imaging cohorts using distance, density-ratio, and shadow-calibrated attacks. Reference-overlap experiments tested evaluation contamination. Graded exact-copy controls, with separate threshold-calibration and test records, measured discrimination and sensitivity at false-positive-rate (FPR) targets of 1%, 5%, and 10%. Findings informed a seven-item reporting checklist.  Reference overlap increased mean density-ratio area under the receiver-operating-characteristic curve (AUC) from 0.502 to 0.809 across five cohorts. Distance-shadow versus shadow-standardized density AUCs were 0.630 versus 0.523 for the conditional variational autoencoder and 0.584 versus 0.521 for the Wasserstein generative adversarial network; both diffusion implementations were near chance under distance-shadow testing. Across 380 controlled release conditions, sensitivity depended on copying fraction and operating point. With complete copying in independent normal-feature experiments, shadow-standardized density achieved AUC 0.967 but sensitivity 0.480 at rank-calibrated target 1% FPR (achieved 1.04%); distance achieved 1.000 for both, with achieved FPR 1.00%. SynPAR links reference checks and measured sensitivity to practical reporting recommendations. The findings support assessing discrimination and sensitivity to known exposure jointly at explicitly evaluated operating points when interpreting attack results.