Federated averaging in noisy asynchronous settings

Abstract

This research report investigates the robustness of Federated Learning (FL) by understanding and examining the performance of Federated Averaging (FedAvg) and Federated Proximal (Fed- Prox), along with the study of fading and noise in wireless communication. The study begins with an introduction to wireless systems and the role played by noise and fading in them. Further, MATLAB simulations are performed to understand the effect of noise and fading on a signal by creating BER vs SNR plots for signals. Then, Federated Learning is introduced along with its basic functioning, and various FL algorithms such as FedAvg and FedProx are examined. Simulations are performed using different datasets and models, and key observations are made based on them. Finally, our problem statement based on convergence in FedAvg with a proximal constraint in a noisy asynchronous environment is introduced along with our research motivation to work in this domain.