Peg-tab improves privacy by fixing copied records in synthetic tables
PEG-Tab: Sampling-Time Record Repair and Release Control for Tabular Synthesis
Machine Learning
Summary
Sometimes computer programs that create fake but realistic tables copy real data too closely, which can risk privacy. The authors developed PEG-Tab, a tool that checks and repairs these copied or very similar records after the program has finished training. PEG-Tab picks safer versions of each generated row without retraining the entire program, lowering privacy risks while keeping the table useful. They tested this approach on different generators and datasets and found it reliably reduces duplicate or copied rows with little harm to data quality.
What this means in practice
- •For data privacy teams: Improve synthetic data release safety by reducing records copied too closely from real data without retraining generators.
- •For machine learning practitioners: Control the quality and privacy of synthetic tabular datasets by applying PEG-Tab as a post-processing step across various generator types.
Authors
Pengfei Li, QinYi Liu, Mohammad Khalil
Abstract
Pretrained tabular generators can reproduce training records even when aggregate utility remains high. When retraining is unavailable or too costly, sampling and release are the remaining intervention points. We present PEG-Tab (Post-Training Energy Guidance for Tabular Synthesis), a post-training repair and release-control framework for frozen tabular generators. For each generated row, a generator-native operator creates two alternatives. A shared calibrated score compares the three candidates, favours lower-risk records, and applies a final release check. We instantiate this interface for GReaT, CTGAN, TVAE, and TabDDPM without updating their parameters. Across five datasets and four generator families, PEG-Tab reduces mean Near Copy from $0.078$ to $0.027$ and lowers aggregate Exact Copy to zero. Relative to a $3\times$ post hoc filter, it retains higher utility in 12 of 16 transfer settings and Pareto-dominates the filter in eight. Gains are concentrated in copy and proximity-related risks.