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

Federated Learning (FL) techniques have been emerging in the last few yearsto provide enhanced learning functionalities and facilitate the decision-making process inconnected 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 FLtools, on the other hand, are based on a distributed architecture: rather than relying on acentral orchestrator, vehicles are able to autonomously share the parameters of the Ma-chine 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 driv-ing and we propose a modular FL approach for road actor classification in a cooperativesensing use case. Lidar point clouds are used as input to a PointNet compliant architec-ture. At training time, a subset of the model parameters is mutually exchanged amonginterconnected vehicles, namely selected ML model layers, to optimize communication ef-ficiency, convergence and accuracy. Real data extracted from a publicly available datasetare used to validate the proposed method. Data partitioning policies target practical sce-narios with highly unbalanced local dataset across vehicles. Experimental results indicatethe FL complies with the extended sensors use case for high SAE levels, and outperformsego approaches with minimal bandwidth usage