In recent times, the application of Federated Learning (FL) solutions in connected and autonomous vehicles has gained significant attention due to its privacy-preserving nature. However, deploying FL solutions in Vehicle-to-Infrastructure (V2I) scenarios has criticalities, particularly due to bandwidth limitations and intermittent connectivity. The uplink, in which all vehicles must transmit model updates to the roadside unit (RSU) which operates as the aggregator, is especially resource-intensive. In contrast, the downlink is comparatively efficient, as the global model can be broadcast once to all vehicles within the RSU's communication range. To address these challenges, we propose SLiM-V2I, a communication-efficient FL framework designed to reduce uplink transmission overhead in V2I settings. In SLiM-V2I, each vehicle node compares its locally trained model with the most recent global model it received and selectively transmits only those layers that exhibit significant changes, rather than sending the entire model to the RSU for aggregation. At the RSU, these partial updates are merged to reconstruct the global model. We further investigate performance-aware participation of the vehicle nodes and its impact on both data transmission efficiency and overall model performance. We evaluate SLiM-V2I using a publicly available driving dataset and a convolutional autoencoder architecture for unsupervised learning of driving behavior, with potential applications in anomaly detection. Simulation results demonstrate that SLiM-V2I obtains substantial communication savings while maintaining competitive accuracy and convergence when compared to standard application of FL. SLiM-V2I offers a scalable and resource-efficient solution for FL in intelligent transportation systems, particularly in dense vehicular environments where uplink communication costs are high.
SLiM-V2I: Selective Layer Merging for Communication-Efficient FL in V2I Networks
Mertens, Joannes Sam;Galluccio, Laura;Morabito, Giacomo
2025-01-01
Abstract
In recent times, the application of Federated Learning (FL) solutions in connected and autonomous vehicles has gained significant attention due to its privacy-preserving nature. However, deploying FL solutions in Vehicle-to-Infrastructure (V2I) scenarios has criticalities, particularly due to bandwidth limitations and intermittent connectivity. The uplink, in which all vehicles must transmit model updates to the roadside unit (RSU) which operates as the aggregator, is especially resource-intensive. In contrast, the downlink is comparatively efficient, as the global model can be broadcast once to all vehicles within the RSU's communication range. To address these challenges, we propose SLiM-V2I, a communication-efficient FL framework designed to reduce uplink transmission overhead in V2I settings. In SLiM-V2I, each vehicle node compares its locally trained model with the most recent global model it received and selectively transmits only those layers that exhibit significant changes, rather than sending the entire model to the RSU for aggregation. At the RSU, these partial updates are merged to reconstruct the global model. We further investigate performance-aware participation of the vehicle nodes and its impact on both data transmission efficiency and overall model performance. We evaluate SLiM-V2I using a publicly available driving dataset and a convolutional autoencoder architecture for unsupervised learning of driving behavior, with potential applications in anomaly detection. Simulation results demonstrate that SLiM-V2I obtains substantial communication savings while maintaining competitive accuracy and convergence when compared to standard application of FL. SLiM-V2I offers a scalable and resource-efficient solution for FL in intelligent transportation systems, particularly in dense vehicular environments where uplink communication costs are high.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


