Reproduce “Generating Private Synthetic Data with Genetic Algorithms” on OpenResearch

About this paper

Terrance Liu, Jingwu Tang, Giuseppe Vietri, Zhiwei Steven Wu

PRIVATE-GSD introduces a new approach for differentially private synthetic data generation using genetic algorithms, which enables direct optimization of both differentiable and non-differentiable statistical queries. The method outperforms existing techniques on non-differentiable queries and remains competitive on differentiable ones, while generating synthetic data useful for downstream machine learning applications.

Key insights

• Genetic algorithms, particularly with tailored selection, crossover, and mutation operations, can effectively generate differentially private synthetic data.
• Zeroth-order optimization methods, like those used in PRIVATE-GSD, can directly optimize a wider range of statistical queries, including non-differentiable ones, without needing surrogate functions or data discretization.
• The iterative nature of genetic algorithms, combined with adaptive query selection and private mechanisms, allows for robust privacy guarantees while maintaining data utility.

Summary by alphaXiv.