Federated Learning with Erroneous Communication Links

TL;DR

This paper analyzes federated learning over unreliable communication links modeled as packet erasure channels, demonstrating convergence properties and comparing update strategies.

Contribution

It introduces a convergence analysis for federated learning with communication errors and compares update strategies under packet erasure conditions.

Findings

Convergence to the same global model as error-free FL.

Using only fresh updates can lead to faster convergence when data is uniformly distributed.

Theoretical validation supported by simulation results.

Abstract

In this paper, we consider the federated learning (FL) problem in the presence of communication errors. We model the link between the devices and the central node (CN) by a packet erasure channel, where the local parameters from devices are either erased or received correctly by CN with probability $\epsilon$ and $1-\epsilon$, respectively. We proved that the FL algorithm in the presence of communication errors, where the CN uses the past local update if the fresh one is not received from a device, converges to the same global parameter as that the FL algorithm converges to without any communication error. We provide several simulation results to validate our theoretical analysis. We also show that when the dataset is uniformly distributed among devices, the FL algorithm that only uses fresh updates and discards missing updates might converge faster than the FL algorithm that uses past local updates.