Research on vehicle trajectory privacy protection mechanism based on DP-transformer model

Abstract

Under the 6G integrated communication, sensing and computing framework, intelligent Internet of Vehicles (IoV) generates massive volumes of ultra-long sequential trajectory data persistently. The majority of prevailing trajectory privacy preserving algorithms are constructed upon recurrent neural networks, which bear inherent drawbacks including insufficient modeling capacity for long-range spatio-temporal dependencies, disassociated pipelines of temporal feature learning and privacy perturbation, and drastic utility degradation of trajectory records introduced by globally invariant noise magnitudes. To address the aforementioned limitations, this paper proposes an integrated trajectory privacy protection algorithm termed DP-Transformer. A multi-head self-attention autoencoder is deployed to reconstruct global spatio-temporal trajectory representations, a post-processing differential privacy architecture is elaborately devised, and a spatio-temporal feature-driven dynamic privacy budget allocation strategy is formulated to realize adaptive Laplace noise injection, yielding trajectory samples complying with rigorous differential privacy constraints. Comparative experiments are conducted on the public T-Drive taxi trajectory dataset. Experimental results verify that the overall performance of DP-Transformer remarkably outperforms a suite of baseline models, BiLSTM and DP-BiLSTM. Equipped with provable mathematical privacy guarantees, the proposed approach can faithfully retain trajectory spatial distribution, temporal continuity and traffic query precision, and thus delivers high-performance privacy-preserving technical underpinnings for secure IoV trajectory publication, urban road network analytics and vehicle-mounted location-aware service scenarios.