Understanding Feature Discovery in Website Fingerprinting Attacks
Published in IEEE Western New York Image and Signal Processing Workshop (WNYISPW), 2018
In this work, we investigate the use of convolutional neural networks (CNNs) in state-of-the-art website fingerprinting (WF) attacks. Our primary focus is on understanding how CNNs discover relevant features from packet traces to identify visited websites in the Tor network. Using neural network attribution techniques, we create visual explanations (heatmaps) that show which parts of the network traffic are most important to the model’s decision-making. Based on these insights, we propose a novel WF defense that selectively applies random padding to highly fingerprintable regions, reducing attack accuracy from 98% to 66% with a packet overhead of approximately 80%.
Recommended citation: Nate Mathews, Payap Sirinam, Matthew Wright. (2018). "Understanding Feature Discovery in Website Fingerprinting Attacks." IEEE Western New York Image and Signal Processing Workshop (WNYISPW), 2018. DOI: 10.1109/WNYIPW.2018.8576379.
Download Paper | Download Slides