A two-sided client-server matching mechanism for resilient Federated Learning

Federated Learning (FL) is a collaborative learning paradigm that enables multiple clients to train models collectively without sharing raw data, thereby addressing privacy concerns inherent in traditional centralized machine learning approaches. The process of selecting the clients that would participate in the model training is challenging due to their inherent heterogeneity, each having unique data characteristics, and computational resources. In FL, ensuring high-quality datasets is fundamental for achieving a better performance for the global model, as low-quality and noisy data can significantly degrade the global model accuracy. Several existing methods have been proposed for the client selection problem in FL, considering different client parameters, such as reputation and computational power. However, data-based selection that utilize data quality attributes to assess the suitability of the client’s data for FL tasks, was not tackled. This article proposes a novel two-sided client and server matching mechanism that considers data-, client-, and server-based parameters to match clients with servers in FL. Our approach includes the design of server and client preference functions that enable the assessment of servers and clients based on their respective parameters. Additionally, we propose a matching algorithm that incorporates the preferences of both in the selection. Experiments show that our proposed approach is resilient to variabilities in data, such as outliers and imbalance, consistently outperforming existing benchmarks. On the MNIST image-classification dataset, our proposed approach achieves 97.46% accuracy, outperforming DSCS (96.32%), FedMint (97.27%, ), and Vanilla FL (95.31%). For the Diabetes dataset, it attains 87.01%, exceeding DSCS, FedMint, Vanilla FL ((Formula presented) ). On Rain prediction, it reaches 84.69%, above DSCS, FedMint, Vanilla FL ((Formula presented) ).