DP-Select: Improving Utility and Privacy in Tabular Data Synthesis with Differentially Private Generative Adversarial Networks and Differentially Private Selection

Abstract

This thesis proposes DP-Select, a novel approach to tabular data synthesis that combines DP-GAN and differentially private selection. We develop a mutual information-based selection method that is flexible and scalable for high-dimensional data and large numbers of marginals while being differentially private. We evaluate DP-Select on various datasets and demonstrate its effectiveness and utility compared to existing DP-GAN methods. Our results indicate that DP-Select significantly enhances the utility of synthesized data while maintaining strong privacy guarantees, making it a promising extension of DP-GANs for privacy-preserving data synthesis in terms of differential privacy. We also show that DP-Select performs better for smaller privacy budgets, making it an attractive option for scenarios with limited privacy budgets.