Latency-rigid OTA federated learning

Abstract

This research report investigates the robustness of Federated Learning (FL) by understanding, examining, and comparing the performance of Federated Averaging (FedAvg) and our proposed approach, i.e., Latency Rigid Federated Learning, in the presence of noise in wireless commu nication. The study begins with an introduction to wireless systems and the role played by fading and noise in them. Federated Learning is introduced along with the steps involved in it and a brief summary of different FL algorithms that we have studied. We then introduce our system model, explaining how we derived our update equation based on regular FedAvg and the proofs for the corresponding lemmas that we had to incorporate to get the desired results. We then propose our new update equation based on the required unbiasedness conditions. We then propose our algorithm named Latency Rigid FL, mentioning the various steps involved in both the client and server side. We provide the results of the simulations that we performed based on our new update rule and algorithm by making the necessary changes in the code to compare and contrast the accuracy and time associated with regular FedAvg and our proposed approach.