JustMix JustMix[A]II

Can Websites Be Identified from Tor Traffic?

This article examines “Online Website Fingerprinting: Evaluating Website Fingerprinting Attacks on Tor in the Real World,” presented at USENIX Security in 2022. Giovanni Cherubin, Rob Jansen, and Carmela Troncoso investigated whether the website a user visits can be inferred from the external characteristics of encrypted Tor traffic—and how well earlier experimental results reflected real-world conditions.

The idea behind website fingerprinting is that encryption hides the contents of a transmission but does not necessarily make every transmission look the same. Different websites can produce different patterns in the volume, direction, and sequence of network exchanges. A model uses those characteristics to guess which website was visited. It does not read the page or identify the user: the task being studied is to classify the visited website from observable traffic.

Tor has exit relays, which carry traffic out of the network and onward to websites. The researchers used traffic collected at an exit relay to obtain real data with a known destination: the exit relay can determine which website a connection is going to. This provided labeled examples for training the model.

The authors then tested whether the website could be identified from traffic observed at a different point in the route: a guard relay, Tor’s entry point. Put simply, the research strategy was to first gather labeled examples—where the corresponding website was known—and then test recognition on a part of the Tor connection closer to the user. The researchers adapted Triplet Fingerprinting for an online setting, evaluating it under conditions intended to better approximate ongoing network traffic rather than relying only on previously prepared laboratory traces.

One important difference from many earlier tests was the use of an open-world setting. In a closed-world test, the model always chooses from a small, known list of websites: it is assumed that the user visited one of them. In reality, a user might visit any website. In the open world, the model must not only distinguish the target sites from one another; it must also avoid mistaking the diverse background traffic for traffic to a target site.

For a set of five popular websites, classification accuracy exceeded 95%. When the number of monitored websites increased to 25, the figure fell below 80%. The authors concluded that website fingerprinting may be feasible, but scaling it to monitor a large number of websites in the real world is likely to be difficult.

This does not mean that anyone can identify the website visited by any Tor user with 95% certainty. That figure applies to a particular model and experimental setup. The number of websites in the target set and the amount and variety of background traffic matter considerably. Overall accuracy also does not, by itself, show how often the model mistakenly labels an ordinary visit as a visit to a target site.

First, identifying websites from Tor traffic is not a purely theoretical threat: under certain conditions, a model can extract information from the external characteristics of network connections. Second, high accuracy on a small test set does not mean that the same performance will hold when monitoring many websites amid a large volume of background traffic. The scale of the task and the conditions of the evaluation matter.

The study does not show that Tor has been decrypted or that a user can automatically be identified by name. It examines a narrower risk: whether an observer can infer which website was visited, given a suitable vantage point and data to analyze. The most accurate summary is that fingerprinting is possible, but its practical effectiveness depends on the scenario—and results from a small test should not be generalized to the entire web.

↑ 2

Обсуждение

Комментарии

1

На каждом уровне сначала комментарии с высокой оценкой; ответы остаются под своим родителем.

Написать Join — чтобы участвовать в обсуждении.

Researchers really are impressive. They've already managed to do something like this... As technology advances further, and AI develops too, I can definitely imagine the accuracy of identifying which websites we're visiting getting even higher. It's scary to think that someday, even who accessed them might be revealed...