Tor is an imperfect privacy platform. Ars meets the researchers trying to replace it. Since Edward Snowden stepped into the limelight from a hotel room in Hong Kong three years ago, use of the Tor anonymity network has grown massively. Journalists and activists have embraced the anonymity the network provides as a way to evade the mass surveillance under which we all now live, while citizens in countries with restrictive Internet censorship, like Turkey or Saudi Arabia, have turned to Tor in order to circumvent national firewalls. Law enforcement has been less enthusiastic, worrying that online anonymity also enables criminal activity.
Tor’s growth in users has not gone unnoticed, and today the network first dubbed “The Onion Router” is under constant strain from those wishing to identify anonymous Web users. The NSA and GCHQ have been studying Tor for a decade, looking for ways to penetrate online anonymity, at least according to these Snowden docs. In 2014, the US government paid Carnegie Mellon University to run a series of poisoned Tor relays to de-anonymise Tor users. A 2015 research paper outlined an attack effective, under certain circumstances, at decloaking Tor hidden services (now rebranded as “onion services”). Most recently, 110 poisoned Tor hidden service directorieswere discovered probing .onion sites for vulnerabilities, most likely in an attempt to de-anonymise both the servers and their visitors.
Cracks are beginning to show; a 2013 analysis by researchers at the US Naval Research Laboratory (NRL), who helped develop Tor in the first place, concluded that “80 percent of all types of users may be de-anonymised by a relatively moderate Tor-relay adversary within six months.”
Despite this conclusion, the lead author of that research, Aaron Johnson of the NRL, tells Ars he would not describe Tor as broken—the issue is rather that it was never designed to be secure against the world’s most powerful adversaries in the first place.
“It may be that people’s threat models have changed, and it’s no longer appropriate for what they might have used it for years ago,” he explains. “Tor hasn’t changed, it’s the world that’s changed.”
Tor’s weakness to traffic analysis attacks is well-known. The original design documents highlight the system’s vulnerability to a “global passive adversary” that can see all the traffic both entering and leaving the Tor network. Such an adversary could correlate that traffic and de-anonymise every user.
But as the Tor project’s cofounder Nick Mathewson explains, the problem of “Tor-relay adversaries” running poisoned nodes means that a theoretical adversary of this kind is not the network’s greatest threat.
“No adversary is truly global, but no adversary needs to be truly global,” he says. “Eavesdropping on the entire Internet is a several-billion-dollar problem. Running a few computers to eavesdrop on a lot of traffic, a selective denial of service attack to drive traffic to your computers, that’s like a tens-of-thousands-of-dollars problem.”
At the most basic level, an attacker who runs two poisoned Tor nodes—one entry, one exit—is able to analyse traffic and thereby identify the tiny, unlucky percentage of users whose circuit happened to cross both of those nodes. At present the Tor network offers, out of a total ofaround 7,000 relays, around 2,000 guard (entry) nodes and around 1,000 exit nodes. So the odds of such an event happening are one in two million (1/2000 x 1/1000), give or take.
But, as Bryan Ford, professor at the Swiss Federal Institute of Technology in Lausanne (EPFL), who leads the Decentralised/Distributed Systems (DeDiS) Lab, explains: “If the attacker can add enough entry and exit relays to represent, say, 10 percent of Tor’s total entry-relay and exit-relay bandwidth respectively, then suddenly the attacker is able to de-anonymise about one percent of all Tor circuits via this kind of traffic analysis (10 percent x 10 percent).”
“Given that normal Web-browsing activity tends to open many Tor circuits concurrently (to different remote websites and HTTP servers) and over time (as you browse many different sites),” he adds, “this means that if you do any significant amount of Web browsing activity over Tor, and eventually open hundreds of different circuits over time, you can be virtually certain that such a poisoned-relay attacker will trivially be able to de-anonymise at least one of your Tor circuits.”
For a dissident or journalist worried about a visit from the secret police, de-anonymisation could mean arrest, torture, or death.
As a result, these known weaknesses have prompted academic research into how Tor could be strengthened or even replaced by some new anonymity system. The priority for most researchers has been to find better ways to prevent traffic analysis. While a new anonymity system might be equally vulnerable to adversaries running poisoned nodes, better defences against traffic analysis would make those compromised relays much less useful and significantly raise the cost of de-anonymising users.
What Tor does really well
Tor was designed as a general-purpose anonymity network optimised for low-latency, TCP-only traffic. Web browsing was, and remains, the most important use case, as evidenced by the popularity of the Tor Browser Bundle. This popularity has created a large anonymity set in which to hide—the more people who use Tor, the more difficult it is to passively identify any particular user.
But that design comes at a cost. Web browsing requires low enough latency to be usable. The longer it takes for a webpage to load, the fewer the users who will tolerate the delay. In order to ensure that Web browsing is fast enough, Tor sacrifices some anonymity for usability and to cover traffic. Better to offer strong anonymity that many people will use than perfect anonymity that’s too slow for most people’s purposes, Tor’s designers reasoned.
“There are plenty of places where if you’re willing to trade off for more anonymity with higher latency and bandwidth you’d wind up with different designs,” Mathewson says. “Something in that space is pretty promising. The biggest open question in that space is, ‘what is the sweet spot?’
“Is chat still acceptable when we get into 20 seconds of delay?” he asks. “Is e-mail acceptable with a five-minute delay? How many users are willing to use that kind of a system?”
Mathewson says he’s excited by some of the anonymity systems emerging today but cautions that they are all still at the academic research phase and not yet ready for end users to download and use.
Ford agrees: “The problem is taking the next big step beyond Tor. We’ve gotten to the point where we know significantly more secure is possible, but there’s still a lot of development work to make it really usable.”