Hi Tor-Dev,
I'm curious what the timing is of Tor's opening of preemptive circuits. Specifically, consider the following attack:
- A new stream is assigned to a clean circuit.
- Because of the above, that clean circuit is now a dirty circuit.
- Because of the above, the number of clean circuits is now decreased
by 1. 4. Because of the above, the number of clean circuits is now lower than the number that Tor wants to have open. 5. Because of the above, Tor opens a new preemptive circuit. 6. An attacker who can observe the circuit in (1) and the circuit in (5) can deduce by temporal proximity that those 2 circuits belong to the same client.
This attack seemed obvious enough to me that I assumed that Tor must have some kind of countermeasure to it, e.g. random delays in opening preemptive circuits. However, the tor-path specification doesn't mention any such countermeasure, and based on a brief search through the Tor source code, all I can find is that Tor opens preemptive circuits using a function that always gets called once per second (with no mention of any delay beyond that one-second interval, random or otherwise).
So, does Tor make any effort to mitigate the above attack? If so, what mitigations are present, and where would I find them (in both the spec and the source code)? If not, is there any documented reason (e.g. "the attack is too hard to pull off" or "we want to mitigate it but haven't gotten to it yet") for the lack of mitigation?
Cheers,
Hi Jeremy,
When I read your checklist from 1 to 6 I remembered that there was a research made on this [1] (I think you are talking about the same thing, except not mentioning where your "attacker" is positioned). If a counter measure existed it would have been documented in the Tor spec for tor-path of course, so I guess that part is correct.
There is an obvious straight forward solution to fix it [2], except AFAIK nobody had time to work on this yet.
I guess this is because this threat is not very scary, it is nice to fix it of course, but correlating anonymous circuits to the same anonymous user is much less scary than: - guard discovery attack; - guard partitioning attacks / path-bias attacks; - routers netflow recording of traffic patterns; - v3 onion services;
There has been a lot of work into these directions.
[1]: https://lists.torproject.org/pipermail/tor-dev/2014-September/007517.html
[2]: https://lists.torproject.org/pipermail/tor-dev/2014-September/007518.html
If this thread model is interesting to you or your project(s), you can take Paul's ideas from [2] and write a patch. It is also going to need a proposal before it will be merged into Tor but at least there will be some action ;)