Source-linked AI summary
LSTM-TrajGAN: A Deep Learning Approach to Trajectory Privacy Protection
Jinmeng Rao, Song Gao, Yuhao Kang, Qunying Huang
TL;DR
Individual-level trajectory data support valuable analyses but remain vulnerable to re-identification because their spatial, temporal, and thematic characteristics can reveal users. The paper proposes LSTM-TrajGAN, an end-to-end LSTM-GAN model trained with TrajLoss to generate synthetic trajectories, and evaluates it on trajectory-user linking. The model suppresses re-identification performance relative to common geomasking methods while retaining some trajectory characteristics, though it requires greater computation and broader evaluation remains needed.
Problem
Existing trajectory-protection methods struggle to control privacy–utility trade-offs, account for temporal and thematic semantics, and avoid procedure disclosure that may enable reverse engineering.
Method
LSTM-TrajGAN combines LSTM and GAN structures to generate privacy-preserving synthetic trajectories, using TrajLoss to measure trajectory-similarity losses during training.
Results
The model generates spatial-temporal synthetic trajectories that prevent user re-identification to a certain degree while preserving some spatial, temporal, and thematic characteristics of original trajectories.
Takeaways & Limitations
Synthetic trajectories provide an alternative for trajectory data publication with a supported balance between privacy protection and spatial, temporal, and thematic utility.
Takeaways & Limitations
The approach requires substantially greater computational effort and pre-deployment training, focuses on trajectory-user linking and selected spatial-temporal analyses, and currently generates trajectories matching the originals’ length at city scale.
Abstract
from arXiv · showhide
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.
1 Introduction
Trajectory data enable mobility and human-environment research but create serious re-identification and geoprivacy risks. The paper proposes LSTM-TrajGAN to generate privacy-preserving synthetic trajectories while evaluating privacy protection, semantic utility, and their trade-off.
- Motivation: Location-based services generate individual-level trajectories that support mobility, disaster-response, and public-health research while raising geoprivacy and broader ethical, legal, and policy concerns.Trajectory privacy concerns preventing disclosure of trajectory identity and sensitive locations.
- Motivation: Removing identifiers does not eliminate privacy threats because spatial, temporal, and thematic characteristics can act as quasi-identifiers for linking trajectories to their creators.Aggregation can also provide little anonymity while reducing spatial resolution and analytical effectiveness.
- Prior approaches: Existing privacy-protection approaches include differential-privacy-related grouping, mixing, cloaking, and generalization, alongside geomasking that perturbs spatial locations.These methods seek to hide or modify original locations while retaining spatial patterns.
- Limitations of prior work: Prior methods make the privacy–utility trade-off difficult to control, rarely model temporal and thematic semantics, and may expose procedures that enable reverse engineering.The paper identifies black-box machine learning models as a possible way to address reliance on manually designed procedures.
- Proposed approach: LSTM-TrajGAN combines LSTM and GAN structures to generate synthetic trajectories for data sharing and publication.The paper also introduces semantic trajectory encoding, TrajLoss, and joint evaluation of privacy protection and utility using real-world location-based-service data.
2 Method
LSTM-TrajGAN combines semantic trajectory encoding, an LSTM-based generator and discriminator, and TrajLoss to generate synthetic trajectories that retain spatial, temporal, and categorical information. The workflow uses encoded trajectories and noise to produce and evaluate synthetic data against real trajectories.
- Trajectory Encoding: The framework encodes location, temporal, categorical, and other trajectory attributes into numerical representations for model training.Location is represented by latitude and longitude deviations from the dataset centroid, while temporal and categorical attributes use one-hot vectors.
- Trajectory Encoding: Variable-length trajectories are zero-padded to the longest trajectory, then masked during training and inference so padding does not affect results.Padding enables equal-sized batches and faster training while masked points do not influence weight updates or derived outputs.
- LSTM-TrajGAN Model: The discriminator distinguishes real from synthetic trajectories, while the generator learns to produce realistic synthetic data from original trajectories and random noise.The discriminator receives trajectory data and evaluates whether samples come from the training set or the generator.
- LSTM-TrajGAN Model: The generator embeds spatial, temporal, and categorical features, fuses them with random noise, models sequences using a many-to-many LSTM, and decodes synthetic trajectory points.Spatial, temporal, and categorical attributes are embedded separately before fusion into 100-dimensional features for sequence modeling.
- TrajLoss: TrajLoss measures spatial, temporal, and categorical similarity between real and synthetic trajectories to train the generator alongside binary cross-entropy.Its component weights can be assigned differently for different scenarios, allowing the loss to reflect application-specific priorities.
3 Experiments
The experiments evaluate LSTM-TrajGAN on trajectory privacy protection and utility using a Foursquare NYC weekly trajectory dataset. The model suppresses trajectory-user-linking performance while preserving spatial, temporal, and categorical characteristics better than key geomasking baselines.
- Experimental setup: The evaluation uses 193 users, 3,079 trajectories, and 66,962 trajectory points from a Foursquare NYC weekly dataset, split two-thirds for training and one-third for testing.The model is trained for 2,000 epochs, and synthetic trajectories are generated from test trajectories and random noise.
- Privacy protection: LSTM-TrajGAN suppresses TUL ACC@1, Macro-P, Macro-R, and Macro-F from over 0.900 to around 0.400, while Top-5 Accuracy decreases from over 0.976 to 0.722.Higher TUL accuracy indicates weaker privacy protection; Gaussian Geomasking performs better than Random Perturbation but remains less protective than the proposed model.
- Spatial utility: The spatial evaluation compares original and synthetic trajectories using Hausdorff Distance and Jaccard Index, including activity-space overlap based on convex hulls.These metrics assess spatial dissimilarity and similarity between corresponding trajectories.
- Spatial utility: Random Perturbation achieves the smallest average Hausdorff Distance (0.004) and largest average Jaccard Index (0.763), but sacrifices location privacy.LSTM-TrajGAN performs better than Gaussian Geomasking on these spatial metrics while also suppressing TUL metrics, indicating a stronger privacy–similarity balance.
- Temporal utility: LSTM-TrajGAN’s hourly POI-category distributions share substantial commonality with the original data, whereas both geomasking baselines generally lack temporal similarity and clear 24-hour patterns.Categories C and E approach zero visit probability because they rarely occur in the training data.
- Temporal utility: The model fits overall temporal visit frequencies with Pearson Coefficient 0.761 and categorical visit frequencies with 0.889, exceeding Random Perturbation and Gaussian Geomasking temporally.The reported results support preservation of temporal and categorical characteristics.
4 Discussion
The discussion examines how training choices, embedding design, and model optimization affect privacy–utility outcomes. It also identifies application-specific trade-offs and limitations of LSTM-TrajGAN.
- Training and Optimization Settings: Higher learning rates can accelerate convergence, while random-noise choices have small effects on TUL metrics.The baseline uses learning rate = 0.001, spatial dimension = 64, and TrajLoss; a learning rate of 0.002 converges faster.
- Training and Optimization Settings: Removing Spatial Loss or Temporal Loss sharply reduces metric scores, whereas removing Categorical Loss has limited impact.Removing the complete TrajLoss produces the lowest TUL metric scores, indicating that spatial and temporal characteristics are central to trajectory preservation.
- Spatial Embedding: 128-dimensional spatial embeddings improve TUL metric scores but trade location accuracy against computational effort.Higher-dimensional vectors can encode more information, but physical-device constraints limit the practical choice of embedding size.
- Privacy–Utility Trade-off: Privacy-protection methods must be evaluated against application-specific data types and analytical requirements rather than treated as universally suitable.Tasks may require POI-based or road-network-based data, and some analyses require precise trajectory-point locations.
- Privacy–Utility Trade-off: During training, increasing TUL Top-5 Accuracy coincides with decreasing Average Hausdorff Distance, exposing a privacy–utility trade-off.LSTM-TrajGAN can monitor this relationship to help identify balanced parameter settings.
- Limitations: The approach requires more computation and predeployment training, focuses on TUL and selected spatial-temporal analyses, produces fixed-length trajectories, and currently targets city-scale data.The deviation-based location representation may not suit global-scale trajectories, and broader evaluations remain future work.
5 Conclusion and Future Work
The paper proposes LSTM-TrajGAN to generate privacy-preserving synthetic trajectories for publication and evaluates its ability to reduce re-identification while retaining trajectory characteristics. Future work extends the loss function, scale, trajectory lengths, attack evaluations, and application coverage.
- Conclusion: LSTM-TrajGAN combines an LSTM recurrent neural network with a GAN to generate privacy-preserving synthetic trajectories for trajectory data publication.The model uses adversarial training and is trained on a Foursquare NYC weekly trajectory dataset.
- Conclusion: The model prevents trajectory creators from being re-identified to a certain degree while preserving spatial, temporal, and thematic characteristics.The evaluation addresses trajectory-user linking and reports potential support for further spatial or temporal analyses.
- Future Work: Future work will improve TrajLoss, extend the framework to global-scale datasets, generate variable-length trajectories, investigate privacy attacks and defenses, and evaluate additional analysis tasks.These directions broaden the model's scale, flexibility, security evaluation, and application coverage.