A Review of Client Selection Methods in Federated Learning

• Published in: Archives of computational methods in engineering 2024-03, Vol.31 (2), p.1129-1152
• Main Authors: Mayhoub, Samara, M. Shami, Tareq
• Format: Article
• Language: English
• Subjects: Clients; Communication; Engineering; Federated learning; Heterogeneity; Machine learning; Mathematical and Computational Engineering; New technology; Privacy; Review Article; Selection
• ISSN: 1134-3060; 1886-1784
• DOI: 10.1007/s11831-023-10011-4

Federated learning (FL) is a promising new technology that allows machine learning (ML) models to be trained locally on edge devices while preserving the privacy of the devices’ data. FL, as an emerging technology, still suffers from a bunch of challenges, including the heterogeneity of its participating devices (clients) in the learning process. The heterogeneity of communication and computational resources or the data distributions of the devices could cause poor performance of FL. In addition, allowing all devices to participate in the FL process is not always possible since some clients may have poor wireless links or their battery energy is low. Therefore, selection of the best clients is a crucial step in FL. Client selection (CS) methods come to view as an essential stage of the FL process to deal with the heterogeneous devices and to avoid selecting clients that do not meet certain requirements. The goal of CS is to select the proper set of participating clients in a way that mitigates the negative effects of that heterogeneity in order to increase the performance of FL. This paper critically reviews recent CS methods for FL. The two main categories discussed in this paper are based on the used metrics and the used mechanisms in the CS methods. The paper also analyses the CS methods, their functionality, and their limitations. Moreover, it provides a comparison of the used approaches in terms of how they are evaluated. Several potential directions are identified that can further enhance FL performance.