Residence and Workplace Recovery: User Privacy Risk in Mobility Data

Mobility data has been collected through mobile devices and cellular networks used in academic research and commercial purposes for the last decade. Since releasing individual's mobility records or trajectories gives rise to privacy issues, datasets owners tend to only publish encrypted mobility data, which doesn't contains users' identification symbol like telephone number. However, we argue and prove that even publishing encrypted mobility data could lead to privacy problem, of which the critical problem is users' residence and workplace identification. We develop an attack system that is able to identify users' important locations by a semi-supervised learning model. In addition to traditional time features, our system takes the users' mobility and living patterns into consideration, which are important and affect each other. Our model demands for less ground truth labels and produces considerable improvement in learning accuracy. With large-scale factual mobile data and long-time tracking ground truth data captured from a big city, we reveal that our attack system is able to identify users' residence and workplace with accuracy about 98%, which indicates severe privacy leakage in such datasets. And we provide advice for this kind of privacy-preserving problem.