RobTC: Robust Spatio-Temporal Trajectory Classification via Collaborative Learning in IoT

With the explosion of crowd mobility data generated by universal mobile devices equipped with spatial positioning modules, deep neural networks (DNNs) have been extensively applied to trajectory data mining and modeling. However, recent studies have shown that DNNs are vulnerable to certain adversarial examples, which are ingeniously crafted by introducing minute and imperceptible perturbations to original examples, but can fool classifiers with high confidence. To enhance the robustness of DNN-based trajectory classification, we propose a novel collaborative learning method for robust spatio-temporal trajectory classification, named RobTC, which consists of an autoencoder-based self-representation network (SRN) for robust latent feature learning and a gated recurrent unit (GRU)-based classification network by sharing parameters with the SRN to safeguard against various adversarial attacks. Furthermore, we introduce feature-level constraints between the original input and the corresponding adversarial examples instead of the point-level denoising strategies to effectively suppress the potential “error amplification effect.” Extensive experiments on the Geolife and Beijing taxi traces datasets demonstrate that our method yields significant improvements (white-box 15% and black-box 13%) over the state-of-the-art methods, suggesting that our proposed method can significantly enhance the model’s robustness against various adversarial attacks while preserving the model’s prediction accuracy on original examples.