Website Fingerprinting For TOR Network Through Neural Network

TOR is designed to be a low-latency system to support interactive such as instant messaging and web browsing.However,TOR has also been used to access criminal websites such as those dealing with illegal drugs.Right now,most of website fingerprintings rely on manually extracted features and finding distinctive features is essential for accurate recognition of websites.All of website fingerprinting have selected their features mostly based on expertise and their technical knowledge.This paper analyzes the advantages and disadvantages of the current website fingerprinting model and puts forward a new model based on deep learning that can avoid the manually extracted features.The paper first proposes a data representation on the TCP layer,which can better reflect the data transmission of TOR anonymous network.Then,based on the new data representation,a website fingerprinting classification model based on long-term and short-term memory neural network is proposed in order to fit the ordered features of the network packet sequence and the time characteristics.Finally,based on the experiment of long-term and short-term memory neural network model and the data restriction problem that neural network generally faces,a new website fingerprinting model based on convolutional neural network is constructed.The key idea is to use the dilation convolution to improve the perception of the neural network and improve the performance of the model.According to key ideas,this paper implements the prototype system and tests the model in closed_world test and open_world test.Results show that the models based on neural network gain the same performance as well traditional methods in closed_world test and 3%-4% higher in open_world test.That is to say,neural network's characteristics are more robust than the artificial features.In comparison with other neural network models,the proposed models are more excellent.Finally,we test the validity of dilated convolution and prove that the TOR series should be regarded as sequences.In summary,the use of neural networks to learn the characteristics of TOR network packet sequences is efficient and feasible.