This is a very simple implementation of an active timing attack on Tor. Please note that
the Tor developers are aware of issues like this –
https://blog.torproject.org/blog/one-cell-enough states:

> The Tor design doesn't try to protect against an attacker who can see
> or measure both traffic going into the Tor network and also traffic
> coming out of the Tor network.
> [...]
> The way we generally explain it is that Tor tries to protect against
> traffic analysis, where an attacker tries to learn whom to investigate,
> but Tor can't protect against traffic confirmation (also known as
> end-to-end correlation), where an attacker tries to confirm a
> hypothesis by monitoring the right locations in the network and then
> doing the math.

That page also links to this really scary paper:
<http://petworkshop.org/2007/papers/PET2007_preproc_Sampled_traffic.pdf>

So, this is a known problem, but I wanted to see how easy it really is to do this,
and I wanted to try it myself, so I built a PoC.

The requirements are:
 - The user points his browser to an attacker's webserver and stays on that server
   long enough (a bit over 4 minutes in my implementation)
 - An attacker controls the webserver or the exit node (or something between them)
   (in my implementation, he controls the webserver)
 - An attacker can measure the internet traffic of all possible users
 - The attacking machines have their time synced over NTP or so

It is NOT required, however, that the webserver is run by the same attacker who also
runs the passive traffic analysis near the users – they can be two distinct attackers
who decide to collaborate after-the-fact. The webserver owner only needs to save the
64-bit ID he generated, the traffic analysis attacker needs to save one bit every four
seconds for every connection.

Also, it is NOT required that the victim's browser supports JavaScript or so. curl would
already be sufficient.

In my implementation, the attacking server can encode 64 bits into a pattern
of data bursts – simplified, a zero becomes "first data, then nothing" and a one
becomes "first nothing, then data". The server then sends those data bursts back to
the client.
The attacker measures the traffic of all possible users and decodes every TCP data
stream back into bits using the data burst encoding. Then, he scans all the decoded
data for the bits he sent to the user to find out which user connected to his server
through Tor.
This is really just a simple one-day-project implementation without any sophisticated
stuff that would probably improve the accuracy and speed a lot.



===== PoC LINK AND USAGE =====
My proof-of-concept code is at <http://git.thejh.net/?p=detour.git;a=tree>.
It needs libpcap and works on Linux. It probably won't work on Windows.

Compile with "./compile.sh".
On the server, run "./pulser". This will open an HTTP server on port 4422.
On the monitoring device (just run it on your computer if you just want to
test it for yourself), run "./pulserecord eth0" as root (replace eth0 with the
right interface if it's wrong).
Generate a new bit pattern with "./random_bits" (or anything else).
In the victim browser that uses Tor, navigate to "http://<your server>/<random bits>".
After the page has loaded in the victim browser, run
"./pulsehunter <random bits> | sort -n | tail"
(with the same bits you used in the browser) on the monitoring device. The number on
the left side is how many bits matched, the TCP connection data (IPs and ports) is on
the right side.



===== TEST RESULTS =====
So, with my PoC ready, I tried it on myself in three different scenarios:

 - no other traffic over Tor
 - reading news articles over Tor
 - watching Youtube videos over Tor (with switching to another video after the first
   one had buffered completely)

Here are the results.


--- DOING NOTHING ---
$ ./random_bits 
1010101010110011101101110010110010011101000000000100101000100101
$ sudo rm out/*
$ sudo ./pulserecord eth0
Device: eth0
^C$ ./pulsehunter 1010101010110011101101110010110010011101000000000100101000100101 | sort -n | tail
[...]
64      ***.***.***.***:59254 -> ***.***.***.***:443
64      ***.***.***.***:443 -> ***.***.***.***:59254

=> All 64 bits correct in both directions – that's a really solid hit.



--- READING NEWS ARTICLES ---
$ ./random_bits 
1100001100001111110101110000101000111110001111010111110011011111
$ sudo rm out/*
$ sudo ./pulserecord eth0
Device: eth0
^C$ ./pulsehunter 1100001100001111110101110000101000111110001111010111110011011111 | sort -n | tail
[...]
60      ***.***.***.***:443 -> ***.***.***.***:59254
61      ***.***.***.***:59254 -> ***.***.***.***:443

60 bits for incoming Tor, 61 bits for outgoing Tor. How solid is that hit? Let's ask Wolfram Alpha:

http://www.wolframalpha.com/input/?i=p%3D0.5+n%3D64&a=*MC.p%3D0!.5+n%3D64-_*Formula.dflt-&a=FSelect_**BinomialProbabilities-.dflt-&f3=60&f=BinomialProbabilities.x_60&a=*FVarOpt.1-_***BinomialProbabilities.x--.***BinomialProbabilities.l-.*BinomialProbabilities.r---.*--

probability for 61 or more bits to appear at random: 2.331*10^-15
Wolfram Alpha says there are roughly 1 200 000 000 personal computers on earth (as of 2011)
1200000000*100*2.331*10^-15 = 0.00028

=> Even if you're monitoring the internet use of every personal PC on earth and every one of those PCs
   is communicating over 100 connections at the time of interest, you're still usually not going to get
   any false positives. This is a solid hit.



--- YOUTUBE ---
I watched youtube and switched to a new video when the first one finished buffering.

$ ./random_bits 
0101100100001100011011001001101101110000011110000011100011110000
$ sudo rm out/*
$ sudo ./pulserecord eth0
Device: eth0
^C$ ./pulsehunter 0101100100001100011011001001101101110000011110000011100011110000 | sort -n | tail
[...]
41      ***.***.***.***:40307 -> ***.***.***.***:443
43      ***.***.***.***:443 -> ***.***.***.***:40307

43 is incoming Tor, 41 is outgoing Tor

http://www.wolframalpha.com/input/?i=p%3D0.5+n%3D64&a=*MC.p%3D0!.5+n%3D64-_*Formula.dflt-&a=FSelect_**BinomialProbabilities-.dflt-&f3=42&f=BinomialProbabilities.x_42&a=*FVarOpt.1-_***BinomialProbabilities.x--.***BinomialProbabilities.l-.*BinomialProbabilities.r---.*--

0.4073% probability for 43 or more bits to appear at random

=> Significant, but if you look at >100 connections, you might well get some false positives.