Client-Aware Gradient Compression for Communication-Efficient Federated Learning on Edge Devices

Authors

  • Lipin Xiong Author
  • Asuguan Johnson Author
  • Xiping Liu Author

Abstract

Federated learning enables a population of edge devices to train a shared model without exposing their raw data, yet its practical deployment on wireless and mobile networks is constrained by the enormous communication volume of gradient exchanges. Most existing compression pipelines adopt a fixed sparsification or quantization budget for every client, which ignores the pronounced heterogeneity of device capability and data distribution across the federation. This paper proposes Sparse Client-aware Gradient Compression (SCGC), an adaptive pipeline that couples an error-feedback sparsifier with a client-aware budget controller and an 8-bit stochastic quantizer. The controller observes per-client bandwidth, straggling statistics and local gradient norms, and adjusts the transmitted fraction of each update so that heavily loaded or statistically distant clients are compressed more aggressively while accurate clients retain a finer resolution. A closed-form budget update rule is derived from a first-order analysis of the aggregation error, and the scheme is evaluated on CIFAR-10, FEMNIST and Speech Commands under a Dirichlet non-IID partition. Experimental results show that SCGC reaches the target accuracy of FedAvg with 7.2-11.5 times less uplink traffic, and outperforms fixed-ratio baselines by 3.1-6.4 accuracy points at an equal communication budget. The results indicate that client-aware budgets are a practical route to accurate federated training on bandwidth-limited edge networks.

Downloads

Published

2026-03-25

Issue

Section

Articles