Automating differentially private tabular data synthesis via Bayesian optimization

Generating high-fidelity synthetic tabular data under strict differential privacy (DP) constraints is a critical challenge. The practical deployment of DP-GANs is often derailed by extreme hyperparameter sensitivity, creating a “parameter lottery” where manual or stochastic tuning struggles to balance data utility with rigorous privacy guarantees. To overcome this barrier, we introduce BO-DPCTGAN, an automated framework that formulates hyperparameter tuning as a privacy-aware black-box optimization problem. Leveraging Bayesian optimization with a Gaussian process surrogate, it efficiently navigates the highly non-convex parameter space. Central to our approach is a four-dimensional evaluation system that quantifies statistical fidelity, structural integrity, machine learning utility, and privacy risk. These metrics are dynamically fused via a novel Harmonic mean objective, preventing the optimizer from falling into “noise traps” where high privacy scores merely mask mode collapse. Extensive experiments across diverse real-world datasets demonstrate that BO-DPCTGAN significantly outperforms traditional DP models and strong random search baselines. Exhibiting superior sample efficiency, it consistently identifies safe, high-utility configurations that resist attribute inference attacks, even under stringent privacy budgets (e.g., $$\epsilon = 1.0$$). Ultimately, this work advances privacy-preserving data synthesis from an ad-hoc art into a principled, controllable engineering workflow.