A Privacy-Preserving Data Sharing Framework Based on Generative Adversarial Networks

Abstract

This paper proposes a privacy-preserving data sharing framework based on Generative Adversarial Networks (GANs), integrating a multi-discriminator mechanism, a dynamic differential privacy adjustment strategy, and a controllable generation module. The framework aims to balance data utility and privacy protection across high-risk domains. In medical data sharing (MIMIC-III) and cross-institutional financial analysis, experiments show that the proposed approach outperforms standard GANs, Differential Privacy Logistic Regression, and Federated Learning in generation quality, downstream task performance, and resistance to inference attacks. The multi-discriminator design constrains the generator from statistical, semantic, and temporal perspectives to mitigate mode collapse, while the dynamic privacy strategy adapts noise levels during training to optimize the privacy-utility trade-off. The controllable generation module enables tailored data distributions for specific business needs, improving minority-class performance. Although the framework introduces computational overhead, it offers a viable solution for secure, high-quality data sharing. Future work will focus on lightweight architectures, automated parameter tuning, and multimodal, cross-domain extensions to enhance adaptability and scalability.