OFPP-GAN: One-Shot Federated Personalized Protection–Generative Adversarial Network

Differential privacy techniques have shown excellent performance in protecting sensitive information during GAN model training. However, with the increasing attention to data privacy issues, ensuring high-quality output of generative models and the efficiency of federated learning while protecting privacy has become a pressing challenge. To address these issues, this paper proposes a One-shot Federated Personalized Protection–Generative Adversarial Network (OFPP-GAN). Firstly, this scheme employs dual personalized differential privacy to achieve privacy protection. It adjusts the noise scale and clipping threshold based on the gradient changes during model training in a personalized manner, thereby enhancing the performance of the generative model while protecting privacy. Additionally, the scheme adopts the one-shot federated learning paradigm, where each client uploads their local model containing private information only once throughout the training process. This approach not only reduces the risk of privacy leakage but also decreases the communication overhead of the entire system. Finally, we validate the effectiveness of the proposed method through theoretical analysis and experiments. Compared with existing methods, the generative model trained with OFPP-GAN demonstrates superior security, efficiency, and robustness.