Mitigating the Communication Straggler Effect in Federated Learning via Named Data Networking

Federated learning (FL) has emerged as a prominent solution that enables distributed training of machine learning (ML) models at multiple end-devices with their own data samples. The FL performance, and in particular, the training convergence speed, are limited by the device with the lowest computation and communication capabilities typically referred to as a straggler. In this work, we address the issues related to the presence of communication stragglers, that is, devices experiencing poor channel conditions. By leveraging named data networking (NDN) and customizing its forwarding fabric to improve the ML model delivery in the presence of lossy communications, our solution allows potential stragglers to beneficially participate in the FL application. Results show the advantages of the conceived solution in terms of reduced training time when compared to a conventional host-centric FL approach, and higher accuracy with reference to the case in which stragglers are not selected.