MoFoC-VANET: MObility FOrecasting for Connected Vehicles in a Fully Decentralized and Communication-Efficient VANET Environment

Numerous studies aim to enhance the prediction accuracy of the autonomous vehicle behavior to promote safer and more efficient road traffic. The most promising approaches leverage deep learning techniques, with Federated Learning (FL) used to protect data privacy and improve communication efficiency between vehicles. FL enables collaborative trajectory prediction without sharing individual data. To address limitations of FL, such as single point of failure and communication bottlenecks, decentralized federated learning (DFL) is increasingly explored for mobility models. DFL eliminates the need for a central device by enabling direct vehicle communication. However, existing DFL-based trajectory prediction models often prioritize prediction accuracy or communication efficiency. To address these limitations, a new DFL-based trajectory prediction approach, termed MoFoC-VANET, is proposed in this study for a fully decentralized vehicular ad-hoc network (VANET) environment. This approach leverages a Transformer-based model to capture complex spatio-temporal dependencies and employs a gossip-based communication algorithm with weighted aggregation, which considers the quality of the models being merged. Evaluations on two open datasets show that our fully decentralized approach outperforms existing methods while offering improved resource usage efficiency over both traditional FL and DFL-based models and improve the privacy as it is performed in fully distributed way.