FEDBAT++: COMMUNICATION-EFFICIENT FEDERATED LEARNING VIA IMPROVED BINARIZATION-AWARE TRAINING

Abstract

Federated learning enables collaborative model training across distributed clients while preserving data privacy, but its practical deployment is often hindered by high communication overhead. Recent approaches such as FedBAT employ binarization-aware training to reduce local computational costs; however, the communication pipeline still relies on full-precision parameter transmission, limiting overall efficiency. In this paper, we present FedBAT++, an improved communication-efficient federated learning framework that extends binarization-aware training to the communication process. Through detailed communication cost analysis, we show that the baseline FedBAT transmits approximately 2.99 GB over 100 training rounds using full-precision parameters, despite maintaining binarized representations during local optimization. FedBAT++ addresses this inefficiency by enabling the transmission of binarized updates with learned scaling factors, significantly reducing uplink communication while preserving convergence stability and model accuracy. Extensive experiments under both IID and non-IID data distributions demonstrate that FedBAT++ achieves substantial communication savings without compromising performance. These results highlight the effectiveness of integrating binarization-aware training with communication compression, making federated learning more practical for bandwidth-constrained and resource-limited environments.