Leveraging Transformer Architecture for Effective Trajectory-User Linking (TUL) Attack and Its Mitigation

Trajectories, a specific type of mobility data, can be used for many useful data mining tasks. However, these trajectories also raises important privacy concerns due to their strong inference potential. In this work, we propose TUL-STEO, a novel deep learning approach, based on a combination of pretraining and fine-tuning, for performing a Trajectory-User Linking (TUL) attack (also called user re-identification attack). More precisely, TUL-STEO can be used to identify individuals associated with given anonymized trajectories. Traditional methods to achieve TUL, as well as its mitigation, usually rely heavily on manual feature extraction, which is less adapted to complex, high resolution and large-scale trajectory datasets. In contrast, our attack is based on an end-to-end machine learning pipeline employing advanced neural architectures derived from transformer neural architectures. Furthermore, we propose Priv-STEO, an adversarial regularization approach to mitigate TUL. We demonstrate how these architectures can be used to effectively mitigate TUL, through an adversarial regularization approach, without a significant degradation of trajectory data utility.