PerFL: A Personalized Privacy-Enhanced Federated Learning Approach for Vehicle Dynamics Prediction

Abstract

Deep learning has significantly improved vehicular dynamics prediction such as travel speed, acceleration, and heading. Centralized training, however, raises concerns about data privacy and communication cost. Federated Learning (FL) addresses these by transmitting model updates instead of raw data, but the updates themselves can still leak sensitive personal information. Local Differential Privacy (LDP) offers a complementary defense through calibrated noise injection, yet applying a uniform LDP budget across drivers ignores their heterogeneous privacy needs and degrades prediction accuracy unnecessarily. In this paper, we propose PerFL, a personalized privacy-preserving vehicle dynamics prediction method that introduces LDP into FL with a per-driver budget tailored to individual privacy sensitivity. Motivated by a questionnaire study revealing that drivers’ privacy preferences correlate with observable mobility attributes, PerFL derives each vehicle’s sensitivity from passively observable mobility statistics. The framework pairs a Gated Recurrent Unit (GRU) for trajectory prediction with a Double Deep Q-Network (DDQN) for per-round privacy-budget allocation, and aggregates the resulting perturbed updates via an SNR-aware FedAvg that down-weights noisier contributions. Comprehensive experiments on a real-world public-transit trajectory dataset from Helsinki show that PerFL Pareto-dominates uniform-DP and random-noise FL baselines, attaining lower per-tier prediction error while consuming under half of the cumulative privacy budget.