[Paper Review] LSTM-TrajGAN: A Deep Learning Approach to Trajectory Privacy Protection

Jinmeng Rao, Song Gao+2 more|arXiv (Cornell University)|Jun 14, 2020

Human Mobility and Location-Based AnalysisSocial Sciences27 references36 citations

TL;DR

Introduces LSTM-TrajGAN, an end-to-end deep learning model that generates privacy-preserving synthetic trajectories to protect user privacy while preserving analytical utility.

ABSTRACT

The prevalence of location-based services contributes to the explosive growth of individual-level trajectory data and raises public concerns about privacy issues. In this research, we propose a novel LSTM-TrajGAN approach, which is an end-to-end deep learning model to generate privacy-preserving synthetic trajectory data for data sharing and publication. We design a loss metric function TrajLoss to measure the trajectory similarity losses for model training and optimization. The model is evaluated on the trajectory-user-linking task on a real-world semantic trajectory dataset. Compared with other common geomasking methods, our model can better prevent users from being re-identified, and it also preserves essential spatial, temporal, and thematic characteristics of the real trajectory data. The model better balances the effectiveness of trajectory privacy protection and the utility for spatial and temporal analyses, which offers new insights into the GeoAI-powered privacy protection.

Motivation & Objective

• Motivate trajectory privacy protection in location-based services and the need for preserving analytical utility.
• Propose an end-to-end deep learning framework to generate privacy-preserving synthetic trajectories.
• Develop a TrajLoss metric to jointly optimize privacy protection and trajectory utility.
• Evaluate privacy protection against trajectory-user linking and assess the utility of synthetic trajectories through spatial-temporal analyses.

Proposed method

• Encode trajectories with spatial deviations from a dataset centroid and one-hot temporal/category features.
• Use a three-component LSTM-TrajGAN: trajectory encoder/embedding, a generator that fuses embedded features with noise, and a discriminator to distinguish real vs synthetic trajectories.
• Train with TrajLoss, a composite loss combining BCE from the discriminator and spatial, temporal, and categorical similarity losses.
• Decode synthetic trajectories to latitude/longitude deviations and one-hot temporal/categorical attributes via dense layers.
• Compare with Random Perturbation and Gaussian geomasking on a trajectory-user linking task and analyze spatial-temporal utility.

Experimental results

Research questions

• RQ1 RQ1: How effective is LSTM-TrajGAN at protecting trajectory creators from re-identification in TUL tasks?
• RQ2 RQ2: Do synthetic trajectories preserve essential spatial, temporal, and thematic characteristics compared to real trajectories?

Key findings

• Synthetic trajectories substantially reduce TUL accuracy, lowering ACC@1 from 0.938 to 0.459 and ACC@5 from 0.976 to 0.722, indicating strong privacy protection.
• LSTM-TrajGAN outperforms baseline geomasking methods in balancing privacy and utility, while still reducing TUL metrics more than random perturbation and Gaussian methods.
• Spatial and temporal characteristics of synthetic trajectories are preserved reasonably well, with temporal/category distributions aligning closely to originals (Pearson correlations reported in analyses).
• Hausdorff and Jaccard-based spatial analyses show LSTM-TrajGAN achieves better spatial similarity than Gaussian geomasking and stronger privacy protection.
• Model sensitivity analyses reveal learning rate and embedding dimension significantly influence privacy-utility trade-offs.