Minimizing Communication Overhead via Wavelet-Based Model Compression in Federated Learning

Federated Learning (FL) is a decentralized approach for training machine learning models across distributed clients, while maintaining data privacy. Despite its advantages, it faces significant challenges, particularly due to the non-independent and identically distributed (non-IID) nature of the data across clients, which can impact the accuracy of the global model. Furthermore, the participation of numerous clients often results in substantial communication overhead and complex resource management. In this paper, we propose FedWave, a novel method designed to reduce communication overhead without significantly increasing computational costs. Our methodology leverages Harten’s multiresolution analysis to apply wavelet transforms to machine learning models, extracting key coefficients. By transmitting only the most significant coefficients, we achieve efficient model compression, reducing the amount of data communicated between clients and the central server. Unlike Singular Value Decomposition (SVD), commonly used for dimensionality reduction, our method offers computational simplicity and faster compression, making it highly suited for real-time, large-scale FL applications. The results demonstrate that our approach achieves comparable accuracy to SVD-based models while significantly decreasing both communication and computational costs, even in distributed, non-IID data environments.