Driving Towards Efficiency: Adaptive Resource-aware Clustered Federated Learning in Vehicular Networks

Guaranteeing precise perception for fully autonomous driving in diverse driving conditions requires continuous improvement and training. In vehicular networks, federated learning (FL) facilitates this by enabling model training without sharing raw sensory data. As an extension, clustered FL reduces communication overhead and aligns well with the dynamic nature of these networks. However, current literature on this topic does not consider critical dimensions of FL, including (1) the correlation between perception performance and the networking overhead, (2) the limited vehicle storage, (3) the need for training with freshly captured data, and (4) the impact of non-IID data and varying traffic densities. To fill these research gaps, we introduce AR-CFL, an Adaptive Resource-aware Clustered Federated Learning framework. AR-CFL utilizes clustered FL to collectively model the environment of connected vehicles, integrating models from all vehicles and ensuring universal accessibility to the refined model. AR-CFL dynamically enhances system efficiency by adaptively adjusting the number of clusters and specific in-cluster participant selection strategies. Using AR-CFL, we systematically study the scenario of online car detection model training on non-IID data across varied conditions. The evaluation results highlight the robust detection performance exhibited by the trained model employing the clustered FL approach, despite the constraints posed by limited vehicle storage capacity. Furthermore, our investigation unveils superior training performance with clustered FL in comparison to specific classical FL scenarios, increasing the training efficiency in terms of participating nodes by up to 25% and reducing cellular communication by 33%.