Decentralized Federated Learning for Extended Sensing in 6G Connected and Automated Vehicles

Federated Learning (FL) techniques have been emerging in the last few years to provide enhanced learning functionalities and facilitate the decision-making process in connected automotive tasks. Yet, much of the research focuses on centralized FL archi- tectures, which have been shown to be limited by latency and scalability. Decentralized FL tools, on the other hand, are based on a distributed architecture: rather than relying on a central orchestrator, vehicles are able to autonomously share the parameters of the Machine Leaning (ML) model via Vehicle-to-Everything (V2X) connections. In this paper, we present an overview of FL potentials in 6G vehicular networks for automated driving and we propose a modular FL approach for road actor classication in a cooperative sensing use case. Lidar point clouds are used as input to a PointNet compliant architecture. At training time, a subset of the model parameters is mutually exchanged among interconnected vehicles, namely selected ML model layers, to optimize communication efficiency, convergence and accuracy. Real data extracted from a publicly available dataset are used to validate the proposed method. Data partitioning policies target practical scenarios with highly unbalanced local dataset across vehicles. Experimental results indicate the FL complies with the extended sensors use case for high SAE levels, and outperforms ego approaches with minimal bandwidth usage.