Abstract: The rapid development of mobile Internet and LBS technology allows location service providers to easily collect an ocean of user location trajectory data. Recent studies have shown that deep learning methods can extract user privacy such as user identity from trajectory datasets. However, the existing work mainly focuses on the check-in trajectories collected by social networks, and the de- anonymization research of GPS trajectories is relatively lacking. Therefore, the research on the de- anonymization technology of GPS trajectory based on deep learning is carried out. Firstly, a pre-training method of GPS trajectory data is proposed. After sub-trajectory division, location conversion and location embedding, the spatial distance and context information of GPS coordinates in original trajectories are embedded into fixed-length vectors, so that the GPS trajectory data can be used as the input of neural network. Secondly, a GPS trajectory de-anonymization method based on deep neural network training is proposed. Based on the pre-trained vector sequences obtained in data pre-training, neural networks such as LSTM and GRU are used as encoders to train and fit user identification to achieve trajectory-user link from anonymous trajectories. Finally, the above methods are verified on Geolife trajectory dataset. In the experiment, the accuracy and Top5-accuracy of trajectory de-anonymization reach 56.73% and 7348%. The experimental results demonstrate that the GPS trajectory de-anonymization method based on deep learning can obtain more accurate user identification from the anonymous trajectory data.

Key words: deep learning, recurrent neural network, trajectory de-anonymization, GPS trajectory, data pre-training