Map-conditioned Generative Adversarial Networks for Automotive Trajectory Privacy Protection

Privacy protection methods based on pseudo-trajectory generation face vanishing gradients in large-scale sequence modeling and trajectory distortion due to insufficient modeling of geographic semantics. To address these problems, this paper proposes a map-conditioned trajectory generation framework (MCTG) based on generative adversarial networks. SN-GAN is employed to learn the global spatial distribution of trajectories. Constraints on the spectral norm of the discriminator stabilize training over long sequences and suppress mode collapse. A conditional GAN incorporating geographic semantics is further constructed, with an encoder-decoder architecture for extracting features of road networks. A symmetric bidirectional long short-term memory (LSTM) combined with a teacher-forcing mechanism jointly models temporal dependencies and boundary constraints of trajectory sequences. Experiments on the geolife and portugal datasets show that the cosine similarity between synthetic and real trajectories reaches up to 0.98, remaining stable within 0.90~0.95 under complex multi-modal conditions. The Jensen-Shannon divergence decreases from 0.22 to 0.105, indicating good distributional consistency. The Hausdorff distance of the synthetic trajectories is consistently lower than those of k-anonymity and differential privacy, reflecting superior preservation of spatial morphology. In simulated privacy attacks, the recognition rate for synthetic trajectories falls below 0.01 and the discriminator accuracy approaches 0.519, indicating that real and synthetic trajectories are nearly indistinguishable. MCTG enables effective privacy protection for vehicle trajectory data by generating synthetic trajectories with high geographic plausibility while preserving statistical distribution consistency and road-network semantic constraints.