Research On Website Fingerprinting Attack Technology Of Tor Based On Deep Learning

In this era shrouded in digitization and informatization,personal privacy information is being stolen on a large scale.More and more Internet users choose Tor anonymous communication system to protect their network privacy.However,there are many criminals who abuse the Tor anonymous communication system to carry out various cybercrime activities,which have had extremely serious negative impacts on society.Website fingerprinting attack enables local listeners to identify the websites that users visit through anonymous communication links,which is an effective means to enhance network supervision.The deep learning method can automatically extract features during the training process,which improves the performance of machine learning algorithms that are difficult to improve due to the need to manually extract features.It has been widely used in the field of website fingerprinting attacks.This paper conducts research on how to improve the performance and efficiency of the deep learning-based website fingerprinting attack method and how to enhance the breakthrough ability of website fingerprinting attacks on existing defense technologies.The main research work is as follows:(1)Aiming at the poor performance of the website fingerprinting attack model based on deep learning when there is only a small amount of training data,a website fingerprinting attack model DSCWF based on deep separable residual network is proposed.Using depthwise separable convolution operation to replace convolution operation and introducing residual module can extract spatial and channel features in traffic data more effectively than convolutional neural network,and parameter amount and calculation amount are also less than convolution neural network.At the same time,dilated convolution is introduced to increase the model’s perception of long-distance dependencies in traffic sequences.The experimental results confirm that the accuracy of the model exceeds that of the comparison model when there is only a small amount of training data.At the same time,the convergence speed of the model is faster during training,which improves the training efficiency,and can effectively alleviate the problem of concept drift and enhance the model.robustness.(2)Aiming at the problem of reduced attack success rate after deploying website fingerprint defense technology in websites,a Spatial-Temporal Feature Fusion Model with Squeeze-and-Excitation Attention Mechanism is proposed.The model uses the deep CNN module to extract the spatial features in the input sequence,and uses the temporal convolutional network specially designed to solve the modeling problem of time series data to extract the temporal features in the input sequence,and fuses the two features.Improve the cross-modal performance of the model.In addition,the SE attention mechanism is introduced into the model to make the model pay more attention to its most important channel features and further improve the attack success rate of the model.The experimental results show that when the STFF-SE model is used to attack on the WTF-PAD defense dataset,an attack accuracy rate of 95.4% is achieved,showing good attack performance.