LCO-AGQ: A Lightweight Client-Oriented Adaptive Gradient Quantization Algorithm for Federated Learning

Quantization is one common approach to achieve communication-efficient federated learning (FL) via compressing the gradients uploaded by clients. However, most existing approaches use a uniform quantization level, neglecting factors such as data quality and resource heterogeneity of clients. While some studies applied customized quantization strategies, they always introduce significant computation cost and overlook the impact of training convergence. To address these issues, we propose a novel algorithm called LCO-AGQ (Lightweight Client-Oriented Adaptive Gradient Quantization), which enables each client to adaptively select its quantization level based on its data quality and communication capability, without substantially increasing computation cost. The core idea of LCO-AGQ is modeling the relationship between quantization levels and the impact degree of client quality, which remains consistent across clients and adjacent training rounds. This allows the impact of quantization levels on training convergence to be estimated in a lightweight manner, and such impact is also demonstrated in our theoretical analysis. Consequently, we adjust each client's quantization level to achieve the efficiency of the FL system. Compared to state-of-the-art methods, our approach achieves comparable accuracy and overall training time while only using 62.11 % of the communication cost and 64.52% of the computation cost.