A Practical Congestion Attack on Tor Using Long Paths

In 2005, Murdoch and Danezis demonstrated the first practical congestion attack against a deployed anonymity network. They could identify which relays were on a target Tor user’s path by building paths one at a time through every Tor relay and introducing congestion. However, the original attack was performed on only 13 Tor relays on the nascent and lightly loaded Tor network. We show that the attack from their paper is no longer practical on today’s 1500-relay heavily loaded Tor network. The attack doesn’t scale because a) the attacker needs a tremendous amount of bandwidth to measure enough relays during the attack window, and b) there are too many false positives now that many other users are adding congestion at the same time as the attacks. We then strengthen the original congestion attack by combining it with a novel bandwidth amplification attack based on a flaw in the Tor design that lets us build long circuits that loop back on themselves. We show that this new combination attack is practical and effective by demonstrating a working attack on today’s deployed Tor network. By coming up with a model to better understand Tor’s routing behavior under congestion, we further provide a statistical analysis characterizing how effective our attack is in each case.