Unveiling the Wireless Network Limitations in Federated Learning

With the advent of 5G and beyond (5GB) communications, decentralized Machine Learning models in various 5GB use cases have become critical. However, wireless network settings are often overlooked in federated learning although they are the crucial factors for edge devices to send their gradient updates. In this paper, the consequences of the background traffic on the performance of the federated learning are evaluated profoundly. Simulation results point out that the convergence is delayed by 60 communication rounds to reach 0.9 accuracy when the background network traffic is dense. Moreover, it is observed that the background traffic with packet interarrival time below 400 ms reduce the accuracy by half in the first 30 communication rounds.