Stop, Don't Click Here Anymore: Boosting Website Fingerprinting By Considering Sets of Subpages

A type of traffic analysis, website fingerprinting (WFP), aims to reveal the website a user visits over an encrypted and anonymized connection by observing and analyzing data flow patterns. Its efficiency against anonymization networks such as Tor has been widely studied, resulting in methods that have steadily increased in both complexity and power. While modern WFP attacks have proven to be highly accurate in laboratory settings, their real-world feasibility is highly debated. These attacks also exclude valuable information by ignoring typical user browsing behavior: users often visit multiple pages of a single website sequentially, e.g., by following links. In this paper, we aim to provide a more realistic assessment of the degree to which Tor users are exposed to WFP. We propose both a novel WFP attack and efficient strategies for adapting existing methods to account for sequential visits of pages within a website. While existing WFP attacks fail to detect almost any website in real-world settings, our novel methods achieve F1-scores of 1.0 for more than half of the target websites. Our attacks remain robust against state-ofthe-art WFP defenses, achieving 2.5 to 5 times the accuracy of prior work, and in some cases even rendering the defenses useless. Our methods enable to estimate and to communicate to the user the risk of successive page visits within a website (even in the presence of noise pages) to stop before the WFP attack reaches a critical level of confidence.