DP-LTGAN: Differentially private trajectory publishing via Locally-aware Transformer-based GAN

Abstract

Trajectory data has a wide range of applications in various domains but also raises serious privacy concerns. To address these concerns, the integration of deep learning with differential privacy for trajectory publication has gained widespread attention. However, existing solutions are mostly based on temporal neural networks and the Generative Adversarial Networks (GAN) framework, which intrinsically faces the ”forgetting” problem, leading to their failure to capture and simulate the multi-time-scale behavioral patterns of traffic participants, thereby reducing the utility of published trajectories. Moreover, the computational cost and the noise introduced by the widely used differentially private gradient perturbation method are proportional to the model size, which compromises model quality. To address these problems, we propose a Differentially Private Trajectory Publishing method via Locally-aware Transformer-based GAN (DP-LTGAN), achieving high-utility trajectory publishing while providing differential privacy protection. Specifically, our method features a Locally-aware Transformer, whose attention mechanism is refined by incorporating local state encoding and a multi-scale temporal encoding mechanism. This enhancement significantly improves the modeling of both long- and short-term trajectory patterns. Furthermore, a differentially private gradient perturbation method named Common Term Perturbation (CTP) has been developed, which effectively reduces the amount of noise and the computational cost by utilizing a designed local noise addition pattern and an adaptive noise addition mechanism. Extensive experiments on several real trajectory datasets show that our method enhances the utility and efficiency of synthetic trajectories by 57.7% and 46.88%, respectively, significantly outperforming current state-of-the-art approaches.