Measuring Privacy Risks and Tradeoffs in Financial Synthetic Data Generation

TL;DR

The work tackles the privacy-utility tradeoff in generating synthetic tabular financial data under differential privacy. It compares four representative generators—Gaussian Copula, TabDiff, CTGAN, and TVAE—with DP adaptations DP-CTGAN and DP-TVAE to assess data quality, downstream utility, and privacy across imbalanced datasets. Key findings show that non-private TabDiff and Gaussian Copula yield strong distribution-matching metrics, while DP noise can degrade downstream performance and yield inconsistent privacy signals, with MIAs offering limited insight into formal DP guarantees. The study emphasizes the need for domain-specific privacy auditing and synthetic data methods tailored to regulated financial data, guiding future development of privacy-preserving tools with tangible practical impact.

Abstract

We explore the privacy-utility tradeoff of synthetic data generation schemes on tabular financial datasets, a domain characterized by high regulatory risk and severe class imbalance. We consider representative tabular data generators, including autoencoders, generative adversarial networks, diffusion, and copula synthesizers. To address the challenges of the financial domain, we provide novel privacy-preserving implementations of GAN and autoencoder synthesizers. We evaluate whether and how well the generators simultaneously achieve data quality, downstream utility, and privacy, with comparison across balanced and imbalanced input datasets. Our results offer insight into the distinct challenges of generating synthetic data from datasets that exhibit severe class imbalance and mixed-type attributes.

Figures (2)

• Figure 1: A visual description of our DP-CTGAN framework, contrasted with the original CTGAN. The privacy leak sites of CTGAN are shown in red boxes, with the privacy-preserving modifications in DP-CTGAN shown in the corresponding green boxes below.
• Figure 2: Comparison of the shift (absolute difference) in minority class percentage of the synthetic dataset over original, and resulting balanced accuracy of a downstream model trained on the synthetic dataset. Each plot represents mean and standard deviation over all datasets. As $\epsilon$ increases (and privacy decreases), we see a clear increase in accuracy and decrease in the variance of minority class balance shift.

Theorems & Definitions (1)

• definition 1