auditing-synthetic-data-leakage

What problem does it solve?

This Skill helps you decide whether a synthetic tabular dataset is safe to release when the source data contains sensitive PHI, PII, financial, or customer records. It focuses on re-identification risk, not just utility scores, so you do not mistake plausible-looking synthetic data for private synthetic data.

Core Features & Use Cases

• Checks for exact and near-duplicate rows between synthetic and real data to catch direct copying and noise-tolerant memorization.
• Uses DCR and NNDR to measure whether synthetic rows are closer to the training set than to held-out real data.
• Runs shadow-model membership inference attacks with confidence intervals and per-attribute disclosure analysis.
• Produces a release verdict of publish, publish-with-DP, restrict-distribution, or withhold, with clear remediation guidance.
• Use it when releasing CTGAN, DP-CTGAN, Synthpop, or similar synthetic tabular outputs for external research or partner sharing.

Quick Start

Ask the assistant to audit your synthetic tabular dataset against real_train and real_holdout, then return leakage findings, a release verdict, and remediation steps.