Reproduce “Generating High-quality Privacy-preserving Synthetic Data” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

David Yavo, Richard Khoury, Christophe Pere, Sadoune Ait Kaci Azzou

A model-agnostic post-processing framework enhances the quality and empirical privacy of synthetic tabular data, successfully addressing categorical mode collapse and proximity-based disclosure risks. This approach restored 100% of dropped categorical modes and maintained predictive utility within 1-1.5% of unfiltered baselines, demonstrating a more favorable fidelity-utility-privacy trade-off.

Key insights

  • Post-processing provides a flexible and complementary approach to enhance synthetic data quality and empirical privacy without requiring changes to the core generative model architecture.
  • Targeted fine-tuning of upper layers can effectively restore missing categorical modes while preserving the overall learned data distribution.
  • Distance-based filtering, when applied with moderate thresholds, can significantly improve empirical privacy and refine data distributions without substantially degrading predictive utility.
  • The framework can improve both univariate and multivariate fidelity, particularly for categorical structures, while preserving downstream utility.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.