Partition-based differentially private synthetic data generation

Abstract

Private synthetic data sharing is beneficial as it better retains the distribution and nuances of the original data compared to summary statistics such as means and frequencies. Current state-of-the-art methods follow a select-measure-generate paradigm, but measuring large-domain marginals often leads to significant errors, and managing the privacy budget poses challenges. Our partition-based approach addresses these issues, effectively reducing errors and improving the quality of synthetic data, even with a limited privacy budget. Experimental results show that our method outperforms existing approaches, yielding synthetic data with enhanced quality and utility, making it a preferred option for private data sharing.