Efficient Model Propagation for Peer-to-Peer Federated Learning using Minimum Spanning Tree and Gossip Networks

Recently, Peer-to-Peer (P2P) Federated Learning (FL) algorithms have caught the attention of researchers due to their ability to extend classical FL through complete decentralization - thereby unlocking potentially new applications as well as addressing various constraints arising from individual nodes in a network. In this correspondence, we study the problem of efficient knowledge propagation in a distributed network. While this can be solved by a simplistic model of All-Pairs Shortest Path traversal, it does not leverage the available parallelism opportunities. On the other hand, there is a rich literature pertaining to rumour spreading, which, however, does not always account for connectivity, communication and node capacity constraints prevalent in P2P networks. Therefore, we present parallel Minimum Spanning Tree formulations, integrating gossip-based local updates and show that this algorithm can be directly applied to efficient P2P training with guaranteed bounds. Such a dynamic and communication-efficient approach is particularly suitable for highly dynamic and resource-constrained environments such as vehicular ad hoc networks (VANETs), industrial IoT (IIoT), and autonomous systems. Our studies on practical networks show a 30.7% reduction in communication costs across various graph sizes compared to pure gossip strategies using FL.