Client Selection Strategy Using Predicted Client Computation Time for Federated Learning

Machine learning is now seamlessly integrated into daily life. High accuracy demands extensive training with large data. Federated learning emerged in 2016, decentralizing training. Clients execute training locally, and the server aggregates their models without direct access to private data. Federated learning selects a subset of clients for each iteration, each with varying training times. The time gap between the slowest and fastest clients is termed a "straggler". Our study proposes a novel client selection strategy, utilizing prediction time based on model computation complexity and client resources. This strategy aims to balance minimizing straggler problems with ensuring fair client selection, thereby enhancing overall efficiency. Our experimental results demonstrate that our proposed client selection strategy can improve the straggler problem by over 40% without significant accuracy loss. Our study emphasizes balancing straggler mitigation and fair client selection for enhanced federated learning outcomes.