Tor Anonymous Traffic Identification Based on Parallelizing Dilated Convolutional Network

The widespread use of the onion browser (Tor) has provided a breeding ground for the proliferation of cybercriminal activities and the Tor anonymous traffic identification method has been used to fingerprint anonymous web traffic and identify the websites visited by illegals. Despite the considerable progress in existing methods, problems still exist, such as high training resources required for the identification model, bias in fingerprint features due to the fast iteration of anonymous traffic and singularity in the definition of traffic direction features. On this basis, a Tor anonymous traffic identification model based on parallelizing dilated convolutions multi-feature analysis has been proposed in this paper in order to address these problems and perform better in website fingerprinting. A single-sample augmentation of the traffic data and a model combining multi-layer RBMs and parallelizing dilated convolutions are performed, and binary classification and multi-classification of websites are conducted for different scenarios. Our experiment shows that the proposed Tor anonymous traffic recognition method achieves 94.37% accuracy and gains a significant drop in training time in both closed-world and open-world scenarios. At the same time, the enhanced traffic data enhance the robustness and generalization of our model. With our techniques, our training efficiency has been improved and we are able to achieve the advantage of bi-directional deployability on the communication link.