HiddenText: Cross-Trace Website Fingerprinting over Encrypted Traffic

Website fingerprinting can infer which website a user visits in Tor networks by eavesdropping and analyzing encrypted traffic patterns. Recent attacks built upon deep neural networks can achieve more than 98% accuracy. To mitigate the privacy leakage under website fingerprinting, effective defenses, such as Walkie-Talkie, have been proposed, where the attack accuracy can be mitigated to 50% at most. In this paper, we propose a cross-trace website fingerprinting, which leverages the semantic correlation of the content of webpages across traffic traces to improve attack accuracy when existing defenses are enabled. Our experimental results on real-world datasets demonstrate that our proposed cross-trace website fingerprinting can completely defeat Walkie-Talkie, in which an attacker can still achieve more than 70% accuracy over defended data.