How to make data differentially private: a practical guide to generating differentially private synthetic data

The paper provides a practical guide to generating differentially private synthetic data, covering the current state of techniques for images, tables, text, and decentralized settings. Differentially private synthetic data aims to preserve the overall trends of the source data while providing strong privacy guarantees for the individuals who contributed the data. The article reviews the relevant techniques, the types of privacy protection these techniques provide, and current practices for each modality, and lists the components required for a generation system, including sensitive data handling and preparation, usage tracking, and empirical privacy testing.