A Systematic Review of Communication-Efficient Federated Learning Through Lossy and Lossless Compression

Abstract

The shift toward decentralized machine learning has positioned Federated learning (FL) as an appealing solution for privacy-preserving collaborative model training, but its practical implementation is still limited by a fundamental bottleneck: the communication overhead. This expense is more than just a hassle in settings with constrained bandwidth and erratic network conditions; it is a constraint that determines whether FL can work at all. Reducing this cost while maintaining model performance has become a key research challenge in the FL community. This paper reviews compression techniques proposed to address communication overhead in FL, covering 25 studies organized under two compression strategies: lossy compression, encompassing quantization, pruning, sparsification, and knowledge distillation, and lossless compression. Each strategy is analyzed in terms of how it reduces the amount of data transmitted between clients and the server, and what impact it has on model accuracy. The review shows that each technique gives a different perspective of the problem and achieves different levels of communication reduction depending on the model, dataset, and system constraints. This reflects the larger truth that there is no single best solution and the best solution is dependent on the specific requirements of a particular deployment environment. This review aims to provide researchers with a better understanding of the available options, and helps guide more informed decisions in the construction of communication-efficient FL systems.