Enhancing Federated Learning Efficiency with Generative Model-Based Data Augmentation for Non-IID Data

With the wide application of machine learning, data privacy has been widely concerned. As a distributed machine learning method, federated learning has the characteristic of protecting user data privacy. However, the data between clients is often non-independent and identically distributed (non-iid), which leads to the efficiency of federated learning is greatly reduced. In this paper, we propose a data argumentation method using generative models to solve the problem of data non-iid problem by synthesizing high-quality data at the client. Specifically, each client pre-trains generative models locally according to local data, and then the client synthesizes the missing data according to the distribution of the local data, that is, performs local data argumentation. We take the original non-iid distribution and the normal data argumentation methods in machine learning as two control groups. Multiple experiments show that the data argumentation method using generative model improves the performance of federated learning to some extent.