[tor-dev] Defending against guard discovery attacks by pinning middle nodes

[George Kadianakis]

Roger Dingledine <arma at mit.edu> writes:

> On Sat, Sep 13, 2014 at 04:07:13PM +0300, George Kadianakis wrote:
>> So let's say that along with our guard, we also pick 6 second-tier
>> guards (middle nodes) that also get pinned for 2-3 months. This makes
>> us look like this:
>> 
>>             -> middle1
>>             -> middle2
>> HS -> guard -> middle3 -> <exit> -> RP
>>             -> middle4  
>>             -> middle5  
>>             -> middle6  
>> 
>>   (where the <exit> is chosen from the set of all relays (not static).)
>> 
>> Since guard picks are independent, the probability of one of your 6
>> middles being compromised is the same probability as the one from the
>> section above: of your main guard getting owned after 6 rotations.
>> This means that every 2-3 months, your guard has 1/4 chance of getting
>> discovered [2]. This means that after a few rotations your guard is
>> guaranteed to be leaked.
>
> Yeah, I think having so many middle-guards will undermine your goal of
> rarely picking a bad middle-guard.
>
> Incidentally, it looks like the math is about the same for the case
> where you pick only one middle-guard but you rotate it every 1/3 months?
>

Indeed.

I suggested multiple middle-guards so that the HS traffic gets load
balanced. However, I mainly mentioned it to demonstrate the flaws of
this scheme. I think the correct solution is having a single
middle-guard :)

>> Some thoughts:
>> 
>> a) To reduce the ownage probabilities we could pick a single middle
>>    node instead of 6. That will greatly improve guard discovery
>>    probabilities, and make us look like this:
>> 
>>       HS -> guard -> middle -> <exit> -> RP
>>       (where <exit> is chosen from the set of all relays)
>> 
>>    However, that will definitely degrade HS performance. I'm not sure
>>    if Tor relays are able to handle all that concentrated HS
>>    traffic. Specifically, the guards/middles that get picked by
>>    popular HSes will get flooded with traffic that is never accounted
>>    for in Tor's load balancing equations (since HS traffic is not
>>    counted) and they will get hammered both by HS traffic and regular
>>    Tor traffic.
>
> Seems like in this case you'd want to pick your middle hop the same way
> you picked your guard hop.
>
> If approximately no Tor users (x=0%) picked their paths this way, then
> the current load balancing should still work fine.
>
> If all Tor users (x=100%) picked their paths this way, we would basically
> push out all middle (non-guard non-exit) relays, and we should double
> up the load that our load balancing algorithms assign to guards.
>
> Were you thinking of doing this modification for just hidden service
> users? If so, and assuming we could measure (an approximation of) how
> many total Tor circuits, and how many total Tor bytes, are hidden service
> related, then it seems like we could weight the shift from middle to
> guard load according to this value x.
>

Yes, I was mainly considering this modification just for hidden
service traffic but it's still unclear to me whether this is a good
idea.

If _all_ Tor clients (both clients and HSes) were to use single
middle-guards, we would indeed have to do major modifications to the
network's load balancing.

OTOH, just changing the HS logic might be easier for the network even
though we would still need to perform load balancing changes similar
to the ones you suggested.

A more important question is what's the anonymity impact of changing
the HS circuit establishment logic. That will make circuits much
easier to distinguish as HS-related, by their guard node, middle-guard
or exit node. However, this is also currently possible if those actors
carefully inspect the rates of incoming/outgoing cells, etc.

More research is required...

>> b) To reduce the ownage probabilities we need to extend the guard
>>    lifetime period. If we switched guards every 9 months, instead of
>>    every 2 months, we would have much lower chance of getting
>>    compromised in the long run. Actually, the probabilities would even
>>    look encouraging (but still nowhere near cryptographic quality).
p>
> Don't get distracted by the fact that some other fields can get great
> security results. :)
>
>>    However, if guard discovery attacks are still possible, having
>>    guards last for 9 months gives the attacker a much larger window
>>    for attacks. Maybe we need to solve guard discovery attacks and
>>    increase the lifetime period at the same time?
>
> What does HS -> guard -> guard -> <middle> -> RP look like, where
> the first two hops are rotated every 9 months?
>

Let's consider HS -> guard -> guard -> <middle> -> RP. That's the
single middle-guard case:

In this case, the chance of getting your guard discovered is the same
as picking a compromised guard: which is optimal.

That is, against a 5% adversary an HS does a coin flip (with 1/20 fail
chance) every 9 months. An HS that hopes to be up for 5 years (that's
around 7 coin flips) has a 0.7 probability (that's (19/20)^7) of being
safe.

Of course, now we are entering the game where the adversary becomes
the <middle> node, and learns your second-tier guard node. and then
the attacker has to own the second-tier guard node, and then own the
guard node to win. This is definitely an improvement over the current
situation (one more box to own), but 9 months is a long time...

Maybe the optimal situation here is the static tunnel:

  HS -> guard -> guard -> guard -> RP

which increases the number of nodes such an attacker needs to
compromise to three. However, such a setup will increase the guard
load even more, and it might also allow other weird attacks (since the
last node is pinned now, RPs know better which HS they are serving).