BACA: Bandwidth and CPU-Aware Adaptive Federated Learning for Wireless Environments

Federated Learning (FL) offers a privacy-preserving approach to machine learning (ML) by training models on devices without sharing the underlying data. However, FL faces challenges in heterogeneous environments such as wireless environments where devices vary in data availability, processing power, and network bandwidth. This paper proposes the BACA algorithm, specifically designed for FL in such settings. BACA dynamically adjusts training parameters based on client characteristics and network conditions to optimize training efficiency. By continuously monitoring CPU utilization, memory usage, bandwidth usage and model performance, BACA ensures adaptive resource allocation, leading to a more efficient and scalable FL process.Experimental evaluations are conducted on a real physical IoT devices in a wireless network to demonstrate the efficacy of the proposed algorithm. Results show adaptive behavior in CPU and memory utilization, alongside improvements in model accuracy and convergence over training rounds. The algorithm effectively manages resource allocation, even in scenarios of constrained CPU limits, network bandwidth, or training time. These findings under-score the algorithm’s potential for practical deployment in real-world edge computing environments, addressing challenges posed by heterogeneous device capabilities and resource constraints.