CRAS-FL: Clustered resource-aware scheme for federated learning in vehicular networks

Abstract

As a promising distributed learning paradigm, Federated Learning (FL) is expected to meet the ever-increasing needs of Machine Learning (ML) based applications in Intelligent Transportation Systems (ITS). It is a powerful tool that processes the large amount of on-board data while preserving its privacy by locally learning the models. However, training and transmitting the model parameters in vehicular networks consume significant resources and time, which is not suitable for applications with strict real-time requirements. Moreover, the quality of the data, the mobility of the participating vehicles, as well as their heterogeneous capabilities, can impact the performance of FL process, bringing to the forefront the optimization of the data selection and the clients resources. In this paper, we propose CRAS-FL, a Clustered Resource-Aware Scheme for FL in Vehicular Networks. The proposed approach bypasses (1) communication bottlenecks by forming groups of vehicles, where the Cluster Head (CH) is responsible of handling the communication and (2) computation bottlenecks by introducing an offloading strategy, where the availability of the extra resources on some vehicles is leveraged. Particularly, CRAS-FL implements a CH election Algorithm, where the bandwidth, stability, computational resources, and vehicles topology are considered in order to ensure reliable communication and cluster stability. Moreover, the offloading strategy studies the quality of the models and the resources of the clients, and accordingly allows computational offloading among the group peers. The conducted experiments show how the proposed scheme outperforms the current approaches in the literature by (1) reducing the communication overhead, (2) targeting more training data, and (3) reducing the clusters response time.