Guard nodes (was: Re: [or-cvs] r13101)

[Paul Syverson]

On Fri, Jan 11, 2008 at 08:21:29AM -0500, Roger Dingledine wrote:
> 
> It's actually a bit more subtle than the above -- Tor clients do use
> alternate guard nodes if their preferred guards go down or become
> unguardworthy, and the set of guard-worthy nodes in the network does
> change over time. So you have to consider the timeframe of the unlimited
> number of requests. Perhaps a better way to say it is that the success
> of the attack during a given interval doesn't improve with the number
> of requests.
> 
> Though a correlation attack to identify guards and an active attack to
> knock them down would shorten the interval.
> 

Lasse and I talked about this in our Oakland 96 paper where we
introduced the attack and the guard node response. We had experiments
on the then Tor network rather than analytic results however. This was
the motivation for the layered guard node design we presented
(something I don't think is worth pursuing implementing without more
study of several aspects.)

> And even then it's not quite right, because I bet somebody could show,
> for a given statistical attack, that it gets more accurate results with
> two requests than one. So maybe I mean that it doesn't improve with
> large numbers of requests. :)
> 
> But yes, I still make the basic assumption that there exists a statistical
> attack that's good enough with just a single request -- even if we
> haven't discovered it quite yet.
> 

Sorry. Do you mean an attack that is statistical but where the
probability is practically speaking high enough for an effective
attack after a single request? If you mean an attack that
deterministically works with a single request, I can't understand how
the latter could be called "statistical".

aloha,
Paul