Federated Learning (FL) in scenarios involving vehicle-to-infrastructure (V2I) communications is a promising solution for training machine learning models collaboratively across vehicles while preserving data privacy and enhancing communication efficiency. A major issue of FL in V2I scenarios is the substantial communication resources consumed during the frequent exchange of model parameters between the vehicles and the central server responsible for model aggregation. As vehicles are connected through wireless channels with limited uplink bandwidth, the frequent transmission of large model updates can cause high communication load which translates into inefficient use of communication resources. In this work, we propose a novel scheme, called Layer-Split Distributed Learning (LSDL), that reduces the communication and computational load onboard the vehicles which are the resource-constrained components of V2I scenarios. LSDL divides the training of the neural network between the vehicles and the central server. Also, LSDL introduces a technique which selects the vehicles that transmit the results of their training to the central server with the aim of further reducing the usage of communication resources. Experimental results show that the proposed approach achieves comparable accuracy to traditional FL while offering significant reductions in communication costs and vehicle-side training loads.
LSDL: Layer-Split distributed learning in vehicle-to-infrastructure communication scenarios
Mertens, Joannes Sam;Galluccio, Laura;Morabito, Giacomo
2026-01-01
Abstract
Federated Learning (FL) in scenarios involving vehicle-to-infrastructure (V2I) communications is a promising solution for training machine learning models collaboratively across vehicles while preserving data privacy and enhancing communication efficiency. A major issue of FL in V2I scenarios is the substantial communication resources consumed during the frequent exchange of model parameters between the vehicles and the central server responsible for model aggregation. As vehicles are connected through wireless channels with limited uplink bandwidth, the frequent transmission of large model updates can cause high communication load which translates into inefficient use of communication resources. In this work, we propose a novel scheme, called Layer-Split Distributed Learning (LSDL), that reduces the communication and computational load onboard the vehicles which are the resource-constrained components of V2I scenarios. LSDL divides the training of the neural network between the vehicles and the central server. Also, LSDL introduces a technique which selects the vehicles that transmit the results of their training to the central server with the aim of further reducing the usage of communication resources. Experimental results show that the proposed approach achieves comparable accuracy to traditional FL while offering significant reductions in communication costs and vehicle-side training loads.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


