On the Orchestration of Federated Learning through Vehicular Knowledge Networking
Duncan Deveaux, Takamasa Higuchi, Seyhan Uçar, Chang‐Heng Wang, Jérôme Härri, Onur Altintas
Abstract
Federated Learning (FL) is a recent distributed technique to extract knowledge, i.e. an abstract understanding obtained from a set of information through experience and analysis. Vehicular networks are highly mobile networks in which a large spectrum of data types is distributed. So far, no mechanisms have been defined that distribute FL model updates in vehicular networks based on which nodes are likely to hold the right data for training, and when. In turn, this potentially limits FL model training speed and accuracy. In this paper, we describe protocols to exchange model-based training requirements based on the Vehicular Knowledge Networking framework. Based on this understanding, we define vehicular mobility and data distribution-aware FL orchestration mechanisms. An evaluation of the approach using a federated variant of the MNIST dataset shows training speed and model accuracy improvements compared to traditional FL training approaches.