Towards Fair and Private Synthetic Health Data: A Hybrid Generative Framework under Zero-Concentrated Differential Privacy
Year: 2026
Type: conference-paper
Abstract: Machine learning for clinical prediction and decision support is constrained by limited, biased, and privacy-sensitive patients' data. While generative models offer a promising alternative, most fail to jointly ensure high utility, fairness, and formal privacy guarantees. This paper presents SF-GAN, a hybrid generative framework that integrates a Conditional Variational Autoencoder (CVAE) with a Wasserstein GAN with Gradient Penalty (WGAN-GP) and an adversarial fairness mechanism. The CVAE captures complex feature dependencies in mixed-type clinical data through tailored reconstruction objectives and class-conditional generation, enabling improved handling of class imbalance. To ensure privacy, zero-Concentrated Differential Privacy (zCDP) is incorporated during CVAE training via gradient clipping and calibrated noise injection, providing rigorous privacy guarantees while preserving data fidelity. Building on these private representations, the WGAN-GP stage enhances sample realism and diversity through stable adversarial training. A dedicated fairness critic is introduced to predict sensitive attributes from generated samples, while the generator is trained to minimize this predictability, thereby enforcing statistical independence and reducing demographic bias. Extensive experiments on five datasets demonstrate that SF-GAN consistently outperforms baselines in predictive utility, fairness, and substantially reduces privacy risks. Ablation studies confirm the complementary contributions of each component, and privacy-utility analyses show robustness across varying privacy budgets.
Source: HAL
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214141325