Distributed Synthetic Time-Series Data Generation With Local Differentially Private Federated Learning

Devices such as cell phones, vehicles, and smart home systems often generate time-series data that are crucial for developing effective AI applications. However, directly analyzing and using these local data could violate user privacy. Although synthetic data offer a promising alternative to real data, current methods for synthesizing time-series data either require servers to first collect large amounts of real user data or fail to offer adequate privacy protection for training generative models. In order to address the limitations in existing solutions, we introduce FedSTDG, an innovative framework for the distributed synthesis of time-series data under strict local differential privacy (LDP) guarantees. Our approach combines federated learning (FL) with a recurrent Wasserstein autoencoder (RWAE) to train generative models without the need to collect real local data directly. Additionally, we propose an enhanced LDP-FL mechanism that incorporates a private dimension selection algorithm and an adaptive learning rate adjustment algorithm. Our approach substantially reduces privacy risks during FL training and improves the model utility compared to existing LDP-FL methods. Extensive experiments on open-source datasets show that synthetic data generated by our framework can reduce downstream prediction errors by 20% to 85% compared to baseline models under equivalent privacy constraints. Furthermore, the framework offers stronger privacy protection and robustness compared to traditional nonprivate FL mechanisms.