Route selection

Roger Dingledine arma at mit.edu
Fri Dec 5 05:51:36 UTC 2003


On Sun, Nov 30, 2003 at 09:39:59AM -0500, Paul Syverson wrote:
> So I woke up some time ago thinking that we were wrong to bother with
> voting on network state in the directory servers, that they should
> just be voting on membership and maybe exit policy. Working on it for
> a while I now think we probably still want a voted network state at
> least c. every hour or so, since `simpler' ideas now seem even more
> complicated, but I think I uncovered another issue.

The reason we need to tell clients which nodes are up right now is so
they can choose working nodes for their circuits. Getting a consensus
prevents a bad directory server from partitioning clients. I think we
want a quick turnaround. Probably the right way to implement it is to
have a vote opportunity every n minutes (5 seems doable), and the vote
happens only if one of the dirservers has requested that it happen.

I think the issue of membership can be treated independently from the
issue of liveness. That is, we could have less frequent votes about what
descriptors to use for each router. Again, these votes need only happen
when a dirserv requests it, e.g. because a descriptor has changed or a
router has been added/removed from the approved list.

None of this has been specified in detail or implemented.

> I was thinking about connection breaks.  By breaking a circuit, under
> our current design and policy, information does not propagate back to
> earlier nodes that the break has happened. But, solving one attack
> creates another, and in this case I think much worse. What this does
> allow is a node breaking circuits until we choose a next hop they like,
> ideally just exiting from that last node before the break.  If
> deterministically iterated, this guarantees a path compromise any time
> the initiator is already identified with a circuit, but even if it is
> only done statistically, the advantage is there. Any bad node in a
> route who knows the initiator can get the responder with uncomfortably
> high probability. (We just moved from c^2/n^2 much closer to c/n).  We

Agreed, this is a problem. There's a broader problem though, which
is that many of our ideas about how to rebuild circuits from partway,
have streams leave circuits from intermediate nodes, etc seem to assume
that circuits are many hops long, whereas for usability they should be
as short as possible. The current planned compromise is around 3 hops,
plus one if the first OR is revealing, e.g. you, and plus one if the
last OR is revealing, e.g. it's your destination or it has a very limited
exit policy.

> should look at strategies such as pulling back at least to the
> penultimate to a break node before extending to reduce this threat.
> This brings it back from needing for this attack just first node and
> any other bad node in the path to needing the first node and two
> successive other bad nodes in the path. We need to establish a default
> circuit break policy cognizant of this attack.
> 
> Here's a stab. We should have a route selection strategy wherein you
> pick the far node (random wrt the exit policy or other constraints you
> have).  Then you get to that node however you do, but you don't let
> the network complete a route to an unknown point for you. If you are

What do you mean 'complete a route to an unknown point for you'?

> going enclave to enclave, then this far node is actually the
> penultimate node (the ultimate being the far-end enclave node).

In this case we could pick the 'far node' without attention to its exit
policy, because we're not going to be exiting from it.

> Whether it is better to rebuild a whole route from scratch or not
> will require much more thought. I suspect that for Onion Routing
> this is a bigger problem than the much ballyhooed predecessor attacks,
> but that's just a gut. Obviously there's a tension between how we
> deal with each of these. Various things we might try, longer routes
> with predesignated break points, etc.

I think longer routes are rarely an acceptable option, because of the
drop in usability they imply.

Below are some scribbles from a conversation I had with Nick about
path selection.

Why variable path lengths?
  - To keep adversary from knowing where on the path he is.
  - If he does know, he can guess how much work it takes to rove to
    an endpoint. Also, he can guess about clients' behavior, e.g.  from
    the exit policy of the last node if he thinks he's the penultimate.
  - Roving is not feasible for short-lived circuits, but may be for ssh.
  - Remember that randomized length just changes it from a certain attack
    to a probabilistic attack.

The big question: When you need to handle a new stream that the current
exit doesn't handle, do you exit from an intermediate node, truncate
and extend, just extend, or make a new circuit? What about when it's
time to rotate a circuit?

If the convention in response to a new unhandled stream is to extend one
hop, then the last node knows that you're about to open a stream which
is accepted by the new node's exit policy but not accepted by his. If
you choose to open a stream from an intermediate node, that node knows
there's a reason you didn't use the later node.

As Paul points out above, if when you're building/extending a circuit the
chosen next node is down -- or more subtly, if it goes down later in the
circuit's lifetime -- and you just blindly choose another node, you're
opening yourself up for the adversary to control the rest of your path.

The current Tor implementation doesn't do any fancy stuff with
circuits. You try to have a few clean (not yet used) circuits around;
periodically you check to see if you don't have enough clean circuits,
and make new ones if needed. Once a circuit that's been dirty for a while
is no longer in use, you close it. New streams prefer to be attached to
recently-dirtied circuits, but if no clean or not-dirty-for-long circuit
has an acceptable exit policy, you make a new circuit right then and
there. Newly built circuits pick an exit node that will handle as many
pending streams as possible; it's possible that several new circuits
will be on their way in parallel if no one exit node can handle them all.

So we don't exit from the middle, we don't reextend, etc. On the other
hand, I believe that currently if a circuit breaks, streams will exit
from the new last node in that circuit, but the circuit's dirtiness
is unaffected.

And since we let multiple streams exit from the same node, that node
can link them together. This new confirmation-style attack may help an
adversary to link two previously unsuspected actions.

More generally, our adversary could be aiming to do all sorts of
things. He might ask "Who is Alice talking to?" Does he need a complete
list, or is a partial list ok? Alternately, "Who talks to Bob?" "Who
else do people-talking-to-Bob talk to?" Adversaries may be positioned
to answer some of these questions better than others.

A fixed first hop means the predecessor attack points to that hop. As
long as the identity of that hop does not help the adversary learn your
identity, you force the adversary to rove to a new part of the system to
learn about you. (Unless he's already there, in which case you've lost.)

A fixed entry hop seems safer than a fixed exit hop -- the end websites
could be bad, so they can link your action now with a previous action
even if you thought they were separated. If we have a fixed entry hop,
also having a fixed exit hop does not help much (contrary to what wright03
says), again because the end websites could be bad.

Other questions revolve around Preferred Entry Nodes and Preferred Exit
Nodes: if the destination website is on the same IP as an OR (think of
this like the enclave model), should we require that a new circuit be
made that ends at that node? Or extend to it from a current circuit? Or
should the Preferred Exit Nodes option just be a recommendation, and if
you happen to be making a new circuit there use it, else don't worry
about it? How do we specify a service that lives at a certain enclave
OR -- there's only one IP space to share, after all. Indeed, since the
client doesn't know what www.freedom.net resolves to yet anyway, how
can he know that 18.244.0.188 is an enclave OR serving that site? Does
the directory have a list of hosts that are served from that OR? What
about ORs that lie to attract and monitor traffic for certain sites? And
if the preferred nodes are down, do you use a different one, or should
the stream fail? How can you notify the user in this situation that the
circuit is less secure than it could be?

Certainly there's a lot to be investigated here.

--Roger



More information about the tor-dev mailing list