A Review on Research Trends of Optimization for Client Selection in Federated Learning

Federated Learning (FL) presents a paradigm shift in data-driven model training, allowing for the collaborative learning of a shared model while keeping all the training data on the client side, hence avoiding central data accumulation. Within this framework, the selection of clients becomes a pivotal element impacting learning efficiency, model performance, and communication overhead. Engaging all clients simultaneously in the learning process can lead to inefficiencies, particularly when considering resource constraints and diverse data distributions. Consequently, there has been a burgeoning interest in research dedicated to developing optimization techniques tailored for strategic client selection to address these challenges. This paper delves into the various optimization strategies proposed in the realm of FL for user selection, scrutinizing their objectives, methodologies, and outcomes. [1] Our review indicates that these techniques not only contribute to alleviating bandwidth constraints and reducing computational loads but also enhance model performance by intelligently choosing clients that provide the most informative updates. Furthermore, we discuss the trade-offs involved in such optimization processes, like the balance between model accuracy and training time, and highlight potential paths for future research that may pave the way for more sophisticated and efficient federated systems.