GANs for Privacy-Aware Mobility Modeling

Human mobility modeling is crucial for many facets of our society, including disease transmission modeling and urban planning. The explosion of mobility data prompted the application of deep learning to human mobility. Along with the growth of research interest, there is also increasing privacy concern. This study first examines the cutting-edge approaches for trajectory generation, classification, and next-location prediction. Second, we propose a novel privacy-aware approach for predicting next-week trajectories. The approach is based on two modules, a Generative Adversarial Network used for generating synthetic trajectories and a deep learning model for user identification which safeguards privacy. These two modules are combined with a next-week trajectory predictor that uses privacy-aware synthetic data. The experiments on two real-life datasets show that the generator creates trajectories similar to the real ones yet different enough to safeguard privacy. The low user-recognition recognition accuracy of models trained on the generated data demonstrates privacy awareness. Statistical tests confirm no significant difference between the original and the generated trajectories. We further demonstrate the utility of the synthetic data by predicting week-ahead trajectories based on the synthetic trajectories. Our study shows how privacy and utility can be managed jointly using the proposed privacy-aware approach.

Automated summary

Human mobility modeling is essential for various societal aspects, such as disease transmission modeling and urban planning. The application of deep learning to human mobility has been facilitated by the availability of vast mobility data. This study explores cutting-edge methods for trajectory generation, classification, and next-location prediction, and proposes a privacy-aware approach for predicting next-week trajectories by combining a Generative Adversarial Network and a deep learning model for user identification. Experimental results show that the generated trajectories preserve privacy without significant deviation from the original ones.