Attackers controlling entry nodes on the Tor network could find the location of hidden services or unmask users visiting them.
Researchers have developed a new technique that could allow attackers to determine with a high degree of accuracy which Tor websites users are accessing and where those websites are hosted.
The new attack, which improves upon previous traffic fingerprinting techniques, was devised by researchers from the Massachusetts Institute of Technology (MIT) and the Qatar Computing Research Institute (QCRI), who found ways to differentiate between different types of connections in a user's encrypted Tor traffic.
The Tor anonymity network was built to hide from network snoopers which websites or other Internet resources that user is accessing. It does this by wrapping the user's requests in several layers of encryption and routing them through multiple computers that run the Tor software.
Each of those computers, known as nodes or relays, peel off one layer of encryption, before passing on the request to the next node. In this way the final node, called the exit relay, knows the request's destination, but not its original source, while the first node, known as the entry guard, knows the original source, but not the final destination.
It has long been known that if an attacker controls both the entry guard and the exit relay used for a Tor connection, or circuit, he could use traffic correlation techniques to deanonymize the user. However, that's hard to do, because Tor relays are chosen at random for every connection so an attacker would have to control a very large number of entry guards and exit relays to have a good chance of success.
In the past researchers also proposed another type of attack known as website fingerprinting that only requires controlling the entry guard. The premise is that attackers could build a list of websites they want to monitor, then set up a Tor client and access those websites through an entry guard they control in order to observe the differences in traffic patterns and use them to build so-called fingerprints.
Those fingerprints could later be used with some degree of success to tell if other users passing through the same entry guard are accessing one of the monitored websites.
This technique does have significant drawbacks. For example, websites have third-party ads and scripts that change frequently so the fingerprints quickly become unreliable. Also there is a lot of background noise in traffic originating from a Tor client and it's hard to isolate only the circuits that are interesting for analysis.
The new technique developed by the MIT and QCRI researchers solves the second problem, especially as it relates to hidden services -- websites that are only accessible inside the Tor network and not on the larger Internet.
Hidden services are popular with political activists who are under the threat of surveillance in certain countries and who want to operate untraceable online communities, but also with criminals who have used them to set up marketplaces for illegal goods or websites that host illegal pornographic content.
These websites are naturally a target for repressive governments or law enforcement agencies who have an interest in knowing who visits them.
With the researchers' new circuit fingerprinting technique, an attacker in control of an entry guard could determine with 99 percent accuracy if a Tor circuit passing through it is used to rendezvous with a hidden service or is used for general Internet browsing. This helps eliminate the background noise and focus on hidden service circuits only.
The researchers also argue that targeting hidden services with website fingerprinting techniques is easier than general Internet websites because their content doesn't change too often.
"In our attack, we show that in the realm of hidden services, we do not have those limitations that exist in the previous attacks," said Mashael AlSabah, an assistant professor of computer science at Qatar University and one of the research's authors, via email. "This makes the previous website fingerprinting attacks more serious in the particular case of hidden services."
The researchers gathered fingerprints for 50 hidden services and found that they could determine with 88 percent accuracy when a Tor client using their entry guard was visiting one of them. They also applied the same technique with a similar rate of success to de-anonymize hidden services when the computers that hosted them used their entry guard.
Hidden services run on computers that are Tor clients themselves so they need to connect to the network through entry guards. However, the entry guards for those computers should not be able to tell which hidden services run on them, because the whole point of hidden services is to hide the IP addresses of the computers hosting them.
Instead, Tor users connect to hidden services through nodes that acts as rendezvous points and are selected according to a special algorithm.
Attackers could increase their chances of success by creating multiple entry guards. A Tor client typically chooses three entry guards and uses them for a period of 45 days on average. Every time a new connection is established, one of the three entry guards is selected.
The more entry guards under their control, the more chances attackers would have of identifying users visiting fingerprinted websites or de-anonymizing particular hidden services.
In their paper, which will be presented at the 24th USENIX Security Symposium next month, the researchers also propose changes to the Tor network that in their opinion would make circuit fingerprinting much harder.
"It's a known issue that hidden service circuits are noticeable, but this attack is very difficult to execute," the Tor Project said in an emailed statement. "The countermeasures described in the paper are interesting since the authors claim that deploying some of them would neutralize their attack and better defend against hidden service circuit fingerprinting attacks in general. This has yet to be proven."
Several Tor developers and privacy researchers will be attending USENIX and are interested to see the research published. "We encourage peer-reviewed research into both attacks against and defenses of the Tor network."