Comparative Analysis of Network Coding Algorithms in Centralized Federated Learning Over Unreliable Networks
Jungmin Kwon, Hyunggon Park · 2025
In wireless federated learning (FL) systems, the use of the User Datagram Protocol (UDP) has been explored to reduce communication overhead by avoiding retransmissions. However, UDP-based transmission is inherently unreliable and can lead to packet loss, causing parts of the model parameters to be lost and thereby degrading the overall training performance. To address this issue, network coding (NC) techniques have been proposed as a complementary solution that linearly combines packets to improve both reliability and efficiency. In this paper, we incorporate several NC algorithms into a centralized FL system and experimentally evaluate how they affect model accuracy and communication efficiency under unreliable communication conditions.