Adaptive Control of Client Contribution and Batch Size for Efficient Federated Learning
Jinhao Ouyang, Yuan Liu · 2024
Federated learning is a promising distributed learning paradigm for protecting data privacy by delegating learning tasks to local clients and aggregating local models, instead of raw data, to a server. However, heterogeneous data and clients slow down learning performance and cause significant communication overheads, which hinder the application of federated learning to wireless networks. To address this issue, in this paper, we develop a novel federated learning framework with contribution-aware client selection and batch size selection to maximize learning efficiency in each round. Firstly, we analyze the impact of the client contribution-aware selection on the convergence rate. Then a learning efficiency maximization problem is formulated by jointly optimizing the contribution threshold and the data batch size in each round. We propose a two-layer iterative algorithm to solve the non-convex problem optimally. Experimental results demonstrate that the proposed scheme can effectively mitigate the influence of data and clients heterogeneity for learning efficiency maximization compared to conventional benchmark schemes.