TableDiffusion: A Diffusion-Based Framework for Differentially Private Mixed-Type Tabular Data Synthesis
Year: 2025
Type: article
Abstract: With growing concerns about data privacy, differential privacy (DP) has become a central technique in privacy-preserving data sharing and machine learning. While many DP generative models have been developed for continuous data such as images, generating synthetic tabular data—especially with mixed data types—remains a challenging task. In this work, we present TableDiffusion, the first diffusion-based generative model specifically designed for differentially private tabular data synthesis. TableDiffusion introduces a hybrid data generation framework capable of handling both continuous and categorical features. It uses a cosine noise scheduler to control the diffusion process and adopts a composite loss function that combines mean squared error (MSE) and Kullback-Leibler (KL) divergence to better adapt to different data types. The model is implemented using the Opacus library to provide formal DP guarantees, and privacy accounting is performed through the Rényi Differential Privacy (RDP) mechanism. To assess the effectiveness of TableDiffusion, we implemented a set of benchmark DP generative models, including attention-based methods such as DPattentionVAE and DPattentionGAN, as well as classical approaches such as PATE-GAN and DP-WGAN. Experiments are conducted on a real-world tabular dataset in the cybersecurity domain, containing more than 50,000 samples and 45 features. We evaluated model performance using metrics such as marginal distribution similarity, PMSE ratio, and Alpha-Beta membership inference scores, focusing on both data utility and privacy preservation. The experimental results are currently being analyzed. Preliminary findings suggest that TableDiffusion offers promising performance in a range of privacy budgets and shows advantages in training stability and diversity over GAN-based methods. Our main contributions include: (1) proposing the first diffusion-based method for differentially private tabular data generation; (2) designing a unified architecture for mixed-type feature synthesis; and (3) providing a comprehensive implementation and evaluation framework to facilitate further research.
Author: Zichen Xu
Institution: University of Minnesota
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416125782