Performance Analysis of Tor Website Fingerprinting over Time using Tree Ensemble Models

Tor (The Onion Router) ensures network anonymity by encrypting contents through multiple relay nodes. Recent studies on website fingerprinting (WF) showed that websites can be identified with high accuracy by analyzing traffic data. However, websites are changing over time by updating contents, which can significantly reduce the accuracy of WF attacks. This study analyzes the performance over time by using ensemble models with excellent WF attack performance. The experiment are conducted in two cases with the initial model. The not updated analyzes the accuracy of models made from initial data over time, whereas the updated adds data that has changed over time to update the model to analyzes the accuracy. The average accuracy of the initial ensemble models is over 90.0% and the Rotation Forest algorithm shows high performance of 93.5%. Comparing the models trained after 30 days with the initial model, the classification performance dropped in both cases; the not updated dropped by more than 30.0% and the updated dropped by about 10.0%. The experimental results suggest that WF using machine learning may require model learning on a regular basis.