Trajectory Pattern Mining Based On Deep Learning And Its Security Confrontation Research

With the continuous development of Internet of Things technology,most of the existing transportation has GPS positioning equipment.Through GPS positioning equipment,the cloud can obtain travel trajectory data of moving objects.A large amount of trajectory data provides an important analysis basis for traffic planning and has potential commercial value.How to fully mine the mode information of spatio-temporal trajectory data and apply it to applications such as traffic planning and route recommendation is an urgent problem to be solved.To fully mine the pattern information of the trajectory data,this paper carries out the mode classification task of the trajectory data and the anomaly detection task of the trajectory mode.Among them,the mode classification task of trajectory data is dedicated to extracting time and space information of trajectory data.The anomaly detection task of trajectory mode is used to mine the difference information between trajectory modes.To ensure the safety of deep learning in the field of trajectory mode classification,this paper conducts adversarial sample research on spatio-temporal trajectory data.The main research content of this paper is as follows.Based on the trajectory mode classification task,this paper proposes a trajectory mode classification model based on deep feature fusion technology.The model extracts shallow features from different granularities of trajectory data;based on the above shallow features,different convolutional networks are constructed for deep feature extraction;trajectory mode classification is performed by multi-headed attention mechanism,deep feature fusion layer and fully connected layer.The experimental results show that the deep feature fusion model has high accuracy,and the local features of the trajectory data and the trajectory embedding vector have more trajectory data information.In order to mine the difference information between trajectory modes,this paper proposes an unsupervised anomaly detection model based on LSTM autoencoder.The model uses the LSTM network to capture the timing information of the trajectory data.The experimental results show that the anomaly detection model can fully learn the data information of normal trajectory mode and detect other types of trajectory modes.Based on the security of deep learning in the field of spatio-temporal trajectory data,this paper proposes a research on adversarial examples in the field of trajectory mode classification.The research designed a convolutional autoencoder to convert the trajectory mode data into image data format,and the image format data is used to train the deep trajectory mode classification model.Based on the trained deep model and trajectory mode data,the FGSM(Fast Gradient Sign Method)algorithm is used to generate the adversarial samples.The experimental results show that adversarial examples exist widely in the field of trajectory mode classification.To investigate whether the adversarial samples can be generated directly based on the trajectory data,based on the original trajectory data and the shallow convolutional neural network,this paper conducts white-box attack and black-box attack experiments,and the results show that the adversarial samples can be successfully generated in the original trajectory data.To improve the defense ability of deep trajectory mode classification,adversarial training is performed to improve the robustness of the deep model.