Geolocated Data Generation and Protection Using Generative Adversarial Networks

Data mining techniques allow us to discover patterns in large datasets. Nonetheless, data may contain sensitive information. This is especially true when data is georeferenced. Thus, an adversary could learn about individual whereabouts, points of interest, political affiliation, and even sexual habits. At the same time, human mobility is a rich source of information to analyze traffic jams, health care accessibility, food desserts, and even pandemics dynamics. Therefore, to enhance privacy, we study the use of Deep Learning techniques such as Generative Adversarial Network (GAN) and GAN with Differential Privacy (DP-GAN) to generate synthetic data with formal privacy guarantees. Our experiments demonstrate that we can generate synthetic data to maintain individuals' privacy and data quality depending on privacy parameters. Accordingly, based on the privacy settings, we generated data differing a few meters and a few kilometers from the original trajectories. After generating fine-grain mobility trajectories at the GPS level through an adversarial neural networks approach and using GAN to sanitize the original trajectories together with differential privacy, we analyze the privacy provided from the perspective of anonymization literature. We show that such epsilon-differentially private data may still have a risk of re-identification.