Privacy-Preserving Vehicle Trajectory Prediction for Smart Cities Using Federated Learning
Keywords:
Federated Learning (FL), Vehicle Trajectory Prediction, LSTM-Transformer Hybrid Model , Differential Privacy (DP) , Secure Aggregation, Real-Time InferenceAbstract
The swift expansion of smart vehicles and urban traffic sensors has become real-time vehicle trajectory prediction an essential element of intelligent transportation systems (ITS). Conventional centralized learning methodologies necessitate extensive aggregation of sensitive mobility data, prompting apprehensions regarding privacy, data breaches, and adherence to regulations. This research presents FedTrack, a federated learning system that facilitates the collaborative training of a hybrid LSTM–Transformer trajectory prediction model among cars and roadside units (RSUs) without the need to upload raw GPS or sensor data. FedTrack incorporates differential privacy, secure aggregation, and gradient compression to guarantee privacy-preserving, fast, and scalable training. Assessments on benchmark datasets (GeoLife, NGSIM, TDrive) indicate a 25–30% enhancement in prediction accuracy relative to centralized baselines, attaining an RMSE of less than 15 meters for a 60-second horizon and an inference latency under 200ms on edge devices. The framework accommodates diverse clients, non-IID data, and optional blockchain-based immutable logging, rendering it appropriate for practical implementation in smart cities.
