FedVANET-TP: Federated Trajectory Prediction Model for VANETs

In recent years, deep learning techniques have been employed within Intelligent Transportation Systems (ITS) to outperform classical trajectory prediction models, aiming for greater precision and efficiency. Federated learning (FL) has attracted substantial attention for its ability to safeguard sensitive vehicle data while reducing communication overhead. However, FL-based trajectory prediction models face a prioritization dilemma. Some prioritize security, leading to longer computation times and higher resource consumption, while others trade off data privacy for increased accuracy, incurring additional computational costs. In this study, we introduce an approach named FedVANET-TP for predicting trajectories in Vehicular Ad hoc Networks (VANETs). Our method ensures a balanced utilization of computational resources, attains high accuracy and offers optimal privacy. The model is constructed within a FL framework and utilizes an encoder-decoder architecture. This architecture incorporates Convolutional Neural Network (CNN) layers to capture crucial spatial features and Long Short-Term Memory (LSTM) layers for capturing temporal dependencies from historical trajectories. FedVANET-TP has been trained and validated using the NGSIM US-101 dataset. The simulation results show that the model’s average root mean squared error (RMSE) and average overall accuracy are superior to those of the comparison models. This model offers higher accuracy than centralized benchmarks and lower resource consumption than federated-based benchmarks, all while ensuring optimal data confidentiality.