On the Effectiveness of Traffic Analysis Against Tor Networks Using Flow Records [pdf]

https://mice.cs.columbia.edu/getTechreport.php?techreportID=1545&format=pdf

by uptown • 12 years ago

289 68 12 years ago

mice.cs.columbia.edu

Fogest2• 12 years ago

More and more just keeps coming out and it honestly seems like Tor is not a flawless method at all for anonymity and it is not something I'd trust relying on. I'd put more trust on using Tor in combination with a logless VPN.

AndrewKemendo• 12 years ago

There is no flawless method of anonymity. The best you can do is have plausible deniability.

Fogest2• 12 years ago

Or use multiple layers so at least you are less likely to be targeted by larger more automated operations, and only so much at risk to a targeted attacked.

lazaroclapp• 12 years ago

Well, you can construct the following protocol for fully untraceable communication between n nodes, numbered i \in 1...n: Divide time into intervals of size t, where t is larger than the time it takes to propagate a message of constant size s from any node to all of the remaining n-1 (e.g. by flooding). For interval j, node i = j % n always transmits a message m of size s to all other nodes with the following characteristic: m is either the output of a PRG or a message encrypted with the key for a host k to which i wishes to communicate a message, the choice of which is entirely up to i. Under this scheme - assuming previously set up authentication between every pair of nodes and an encryption scheme in which encrypted messages are indistinguishable from random data without the decryption key - any node can send a message to any other node in such a way that no one else inside or outside this network can know the contents of the message or even that the communication took place. For any node not receiving communication, the protocol would be indistinguishable of one in which all transmissions are random noise.

Of course, the issue is that latency in this scheme is O(n) and per-node bandwidth is O(1/n), with large constants. Also, it's a reasonable suspicion in practice that no one would set up this scheme and then actually have zero communication going on over it, so it still reveals that "at least one of the n nodes is talking to at least another of the n nodes".

_nullandnull_• 12 years ago

Agreed. Your comments reminded me of the points in "Evil Exit Nodes Unmasked me, and all I got was this lousy jail term" by the grugq. http://grugq.github.io/blog/2013/06/14/you-cant-get-there-fr...

tedks5• 12 years ago

This link is blogspam and should be redirected to the source: https://mice.cs.columbia.edu/getTechreport.php?techreportID=...

dchest• 12 years ago

No, it's an article about research paper. Please don't mislead moderators.

tedks5• 12 years ago

The article adds nothing on top of the paper besides a few remarks about other attacks.

HN should uphold a higher standard than that. Link to the source.

dang• 12 years ago

Hmm. This is a borderline case. The article mostly lifts from the paper, especially the diagrams. On the other hand, HN often prefers the best popular article on a story, with the specialized paper linked in the comments. The reason is that specialized papers are less accessible to a general-interest audience.

That said, papers in computing are an exception because the audience here is informed in that field. And the thread, at this point, provides a lot of context.

Given the above I think the paper is probably the better URL, and we've changed to it from http://thestack.com/chakravarty-tor-traffic-analysis-141114.

Marking this subthread as obviously off-topic.

Thaxll• 12 years ago

Who had the terrible idea of the top scrolling menu...

hawkice2• 12 years ago

So, I'm a motivated privacy-seeker with enough technical chops to configure any not-yet-for-the-masses technology out there. I would very much prefer for my ideal strategy not to be "poop my pants in terror" -- any guidance from the knowledgeable HN-osphere?

sbierwagen• 12 years ago

Freenet.

http://en.wikipedia.org/wiki/Freenet

Upsides: Much more secure than Tor, (since it doesn't try to be a low-latency mixing router) not funded by USG.

Downsides: Real slow. Real slow. SLOW. Think effective throughput in the tens of kilobytes/s, latency measured in minutes.

higherpurpose• 12 years ago

What about i2p? I hear it has quite a few bugs, but so did Tor in the beginning I'm sure. Maybe it just needs better funding. Isn't it a better design at least?

Phantom also seems to have great design, but it was discontinued quite early in its implementation. No idea if it's because it was unworkable or for other reasons. Maybe someone with experience in such networks can take a look at it:

https://code.google.com/p/phantom/

mrsteveman1• 12 years ago

Freenet is faster than that in aggregate, total node throughput can be in the 150-250KB/s range even if single requests are, as you say, potentially in the tens of KB/s. File transfers can be fast, static webpages can be loaded reasonably quickly if they are popular (in the sense of being frequently accessed by users).

Big security downside is that in the purely opennet configuration, your peer nodes are both untrusted strangers AND your own traffic is visible to them. There is NO "onion" effect going on, just reliance on the possibility that any traffic a peer node observes from your node is not necessarily YOUR traffic as all nodes are also routing requests for other nodes.

tedks5• 12 years ago

Upsides: Much more secure than Tor, (since it doesn't try to be a low-latency mixing router) not funded by USG.

Most academic computer science is funded by the American government. Does that mean that all academic computer science is backdoored by the military?

To put it more specifically, most compilers researchers I know have at some point been on a DARPA grant, because DARPA has money and academics want money. I'm sure plenty of LLVM contributors have been paid from DARPA grants. Is LLVM backdoored?

===========================

On an entirely different point, how is Freenet even comparable to Tor? They address totally different use cases (Tor is an anonymizing TCP overlay; Freenet is a distributed censorship-resistant store), and have very different threat models.

Further, it seems very unlikely that Tor (which is a piece of very well-maintained software with some of the foremost privacy researchers working on it) would have fewer bugs than Freenet, which is a sprawling Java program maintained by one man. Further, the security of Freenet in abstract hugely depends on having a functional small-world network, which requires Freenet to have been widely adopted to start with.

The comment parent's request was vague, but you simply cannot replace Tor with Freenet because they do totally different things. It is nonsensical to compare them because they address different threat models and accomplish different goals.

doublec• 12 years ago

I haven't found Freenet latency to be bad for most keys these days. For example, I mirror my blog in Freenet and it's on the order of seconds rather than minutes. Even an image gallery I tried has reasonable response. Uploading data on the other hand is very slow.

Here's my blog key if you want to try:

    USK@1ORdIvjL2H1bZblJcP8hu2LjjKtVB-rVzp8mLty~5N4,8hL85otZBbq0geDsSKkBK4sKESL2SrNVecFZz9NxGVQ,AQACAAE/bluishcoder/3/

And the image gallery:

    USK@2LK9z-pdZ9kWQfw~GfF-CXKC7yWQxeKvNf9kAXOumU4,1eA8o~L~-mIo9Hk7ZK9B53UKY5Vuki6p4I4lqMQPxyw,AQACAAE/pitcairnisland/0/

vitriolix• 12 years ago

"funded by USG"

When it's developed entirely in the open I'm not sure I see the concern here. If you have reason to believe they are somehow compromised or backdooring Tor please do share it, otherwise this is just FUD.

0xdeadbeefbabe• 12 years ago

Test the latest attacks including this one against yourself continuously. Are you going to do this tinfoil?

Seems like a lot of security products demand trust from their users (even gnunet). This implicit bossing is no less pretentious than a gas pump that says "Thankyou".

tedks5• 12 years ago

This is an easier version of a traffic analysis attack, an attack that Tor expressly does not attempt to provide a strong defense against.

It relies on a malicious server and entry node. The contribution of this paper is that if you have the malicious server and entry node, you can use a less expensive data source (Cisco NetFlow data) rather than raw packets to perform a correlation attack.

The correlation they achieve in a private Tor network is impressive; however, if you look at the graphs in the actual paper[0], you can see that the differences in correlations are actually quite small in the wild.

The title of this post and article is actually incorrect; the technique demonstrated has an 81.4% accuracy. This means that the base rate fallacy will make it nearly unusable in practice, and more so as the scale of Tor traffic grows. For more on the Base Rate Fallacy, see [1].

So in summary:

* This is an incremental improvement of an already existing and known attack pattern on low-latency anonymity systems

* The technique presented in this paper is only a threat if your threat model is an adversary that can control your entry guard and the server you are trying to communicate with, but does not have the budget for packet-level correlation attacks

* This technique does not achieve sufficiently high accuracy and sufficiently low false positives to reliably identify arbitrary Tor users, but might be more successful if used in combination with a prior hypothesis that, say, a specific NSA employee is communicating with GlobalLeaks.

[0] https://mice.cs.columbia.edu/getTechreport.php?techreportID=...

[1] http://archives.seul.org/or/dev/Sep-2008/msg00016.html

Tepix2• 12 years ago

Do they really need to control the entry guard? Or is it "good enough" for this attack if they merely have access to the netflow data?

tedks5• 12 years ago

They need to be able to access netflow data at the entry point.

To put it another way, let's say someone is posting bomb threats over Tor on Google Plus. Google cooperates with the FBI and deterministically perturbs all traffic going into the Tor network. What router does the FBI get netflow data from in order to find the bomber?

devindotcom2• 12 years ago

One more reason to have open-developed, ground-up-built endpoint hardware. I'm surprised there hasn't been a credible Kickstarter or something. People wouldn't mind slightly fewer features or lower performance due to reverse engineering an older design - if it meant real, publicly auditable security at that link in the chain. Still lots to do to improve security but that's one that's always seemed neglected to me.

SwellJoe• 12 years ago

While I agree that open hardware is a good idea, Cisco has no malicious intent with NetFlow. It is a legitimate tool for maintaining and improving network quality. This isn't an instance of a vendor providing a back door to the NSA (or whoever). It's an accident of the protocol that it is easier to derive this data from than raw packet sniffing.

devindotcom2• 12 years ago

No, certainly I don't think them malicious, I'm just surprised the option doesn't exist.

tedunangst• 12 years ago

Well, you could use an OpenBSD firewall. But then you'll soon discover that it ships with pflow, which is compatible with netflow.

Tepix2• 12 years ago

As the summary mentions: Tor is susceptible to this kind of traffic analysis because it was designed for low-latency.

You give up a certain amount of security, but gain a lot of comfort.

Perhaps a future version of tor can offer a "comfort level" setting that introduces a varying amount of delays, bogus traffic and other concealment methods (ideally at every hop).

amelius• 12 years ago

You give up a certain amount of security, but gain a lot of comfort.

He who sacrifices security for comfort deserves neither :)

randomjoe2• 12 years ago

Well if that's the case then that should apply to anyone on the internet. It'd be much more secure to just not go on the web, so anyone on the web is sacrificing a huge amount of security. These aren't black and white arguments.

comboy• 12 years ago

It's obviously hard to verify it, but I'd love to hear tor devs take on this.

I may be paranoid, but it seems to be that there is some FUD around tor lately. And if many people will stop using it, it will indeed become less secure.

killface2• 12 years ago

i think it's largely a misinformation campaign to keep people who aren't doing anything illegal from anonymizing their traffic. I like the idea of using Tor, but from what I can see, there's nothing of value for me on there. I don't need to buy drugs or get illegal porn, and I don't really care about anarchist-cookbook style sites. I'm glad it exists for journalists and whatnot, but I don't find it useful for an average person.

readmission2• 12 years ago

If you have the technical and OPSEC wherewithal, it appears that running you own "private" (PublishServerDescriptor 0) exit node has become an extremely attractive anonymity tool.

There are lots of downsides to this (epistemic attacks), but if your anon use case makes sense for such a setup, it is a valuable tool to have in the toolbox.

Torgo• 12 years ago

Can you explain this? bridge relays are setup using "PublishServerDescriptor 0" but exit nodes? How does that work?

readmission2• 12 years ago

https://lists.torproject.org/pipermail/tor-relays/2011-Augus...

If you just config it with PublishServerDescriptor as 0 and someone else knows the IP (middle relays) they will be able to use it.

It's essentially a function of not announcing the node to anyone.

Edit: And to be clear, priv exit nodes don't prevent the timing attack in the article.

paulhauggis• 12 years ago

I stopped trusting Tor a couple of years ago.

SoftwareMaven• 12 years ago

So encrypting content as it travels through Tor is generally considered best practice, but how does one get/manage/pay for a VPN such that the VPN itself doesn't lead directly back to you?

hawkice2• 12 years ago

So, to clarify something: you connect to the VPN, and the VPN connects to TOR. If you do it the other way around (as you point out) it's pretty worthless. If you connect to TOR from the VPN then the most de-anonymized you are going to get is "customer of this VPN provider".

As for paying for services such that "customer of this VPN provider" can't be linked to you, I'd look at places that specifically accept bitcoin (I'm not a fan in general, but in this case, it signals their willingness to avoid collecting normal billing data like name, address, cc#, etc.).

https://proxy.sh/

I have no affiliation but they popped up in the googlings. Seems to be simple enough to get their service without giving away anything personally identifiable.

EDIT: the reason I tease out those two parts is because I'm concerned people will think TOR can protect against entities that can do a global customer-of-VPN dereference. TOR doesn't protect against global super-adversaries, consult your local security practitioner before feeling safe, etc. etc.

onewaystreet• 12 years ago

VPN -> TOR is wrong. Think about what happens if the VPN gets owned or seized. The bottom line is if your need for anonymity isn't just fantasy (i.e. you are actually a target of law enforcement) then no layer of your protection should be an IP that is connected to your real identity or location.

chatmasta• 12 years ago

Anonymous VPN is pretty easy. Bitcoin, or if not, VCC is all you need.

mcintyre1994• 12 years ago

My immediate thought is cryotocurrency. You could probably lose direct links quite easily by using some sort of tumbler process - that should stop the trivial tracing back route. Bitcoin is not strictly designed to be anonymous but I do wonder whether something built on bitcoin or another crypto will achieve that in the near future.

elif• 12 years ago

There are dozens of vpn providers that are not based in the US, who accept bitcoin and don't even require an email address for an account. Many of them proudly do not record logs, let alone retain them.

TorrentFreak did a really good comparison of them recently: http://torrentfreak.com/which-vpn-services-take-your-anonymi...

e12e• 12 years ago

Bitcoin isn't anonymous (it's not a design goal, either). It might be pseudonymous, but that depends entirely of how you acquire bitcoins in the first place.

Second, "no logs"... well: https://www.bestvpn.com/blog/8383/earthvpn-user-arrested-cop...

Good luck with your "no logs" operation in any NATO member/friendly country. Not sure about Russia and its allies, but I expect you'd be hard pressed to achieve "no logs" in China as well. South-Africa and Israel is probably out too. Not sure what that leaves you.

newaccountfool2• 12 years ago

Use bitcoin, pay for the bitcoins with cash. Or buy a VPN and use a giftcard to pay for it. When connecting to the VPN, use a public WiFi or a WiFi without a password, use a long range antenna so your very far away.

exo762• 12 years ago

And as I understand this attack it is possible only if connection between target server and client is not protected by SSL.

colordrops• 12 years ago

Even in incognito mode, on my machine Chrome puts out enough unique info identify it across 4 million other users. You can test your browser here:

https://panopticlick.eff.org/

How well does Tor do with panopticlick?

Forbo• 12 years ago

Last I had checked, using the browser bundle in Tails with javascript disabled puts the fingerprint to about 1 in 22,000, but it's been a couple months since I last tested. Going to fire up my machine and see what it shows now, will report back with results shortly.

Update: Tor Browser in Tails 1.2 with javascript disabled returns a fingerprint that's shared by 1 in 2,615 with 11.35 bits of identifying information.

newaccountfool2• 12 years ago

Is that not rather narrow? How is it not say 1 in 100k?

kissickas• 12 years ago

How do I have more identifying bits of information, but it's shared by more browsers?

Within our dataset of several million visitors, only one in 4,955 browsers have the same fingerprint as yours.
Currently, we estimate that your browser has a fingerprint that conveys 12.27 bits of identifying information.

I figured it would be more unique due to my running Tor Browser on a Mac - but I don't see how the math works out. Unless it actually has a count of machines with information identical to mine?

mbrubeck2• 12 years ago

Try it and see! Tor Browser contains modifications specifically to make fingerprinting harder. Panopticlick gets 11.36 bits of information from Tor Browser on my computer, compared to 22.16 bits from unmodified Firefox on the same machine. With Tor, Panopticlick says "Within our dataset of several million visitors, only one in 2,620 browsers have the same fingerprint as yours."

That said, browser fingerprinting is very hard to fight, and it's a lot easier to come up with new fingerprinting techniques than it is to mitigate them. So I expect that determined fingerprinters have the edge in this arms race.

killface2• 12 years ago

Your browser fingerprint appears to be unique among the 4,697,679 tested so far.

Currently, we estimate that your browser has a fingerprint that conveys at least 22.16 bits of identifying information.

Well, shit. I'm a snowflake.

schoen• 12 years ago

The current Panopticlick measurements are not as numerically accurate as they were were it was first created. That's because the commonness of the things that it measures changes over time (for example, how common it is to have a particular browser version) and the Panopticlick site took a lot of measurements from the public when it was first released, but is taking comparatively few now. However, it's still calculating the statistics based on a database of everyone who has ever visited the site.

Also, Panopticlick hasn't been updated with some possible distinguishing measurements that have been identified since it was created. In that sense it was always meant as a lower bound on uniqueness, and should still be understood that way.

On the Effectiveness of Traffic Analysis Against Tor Networks Using Flow Records [pdf]

Related Stories