A decision framework for privacy-preserving synthetic data generation

Abstract

Access to realistic data is essential for various purposes, including training machine learning models, conducting simulations, and supporting data-driven decision making across diverse domains. However, the use of real data often raises significant privacy concerns, as it may contain sensitive or personal information. Generative models have emerged as a promising solution to this problem by generating synthetic datasets that closely resemble real data. Nevertheless, these models are typically trained on original datasets, which carries the risk of leaking sensitive information. To mitigate this issue, privacy-preserving generative models have been developed to balance data utility and privacy guarantees. This paper examines existing generative models for synthetic tabular data generation, proposing a taxonomy of solutions based on the privacy guarantees they provide. Additionally, we present a decision framework to aid in selecting the most suitable privacy-preserving generative model for specific scenarios, using privacy and utility metrics as key selection criteria.