Federated Learning with Erroneous Communication Links
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 $ε$ and $1-ε$, 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.