[tor-dev] Notes from the prop224 proposal reading group

[George Kadianakis]

George Kadianakis <desnacked at riseup.net> writes:

> [ text/plain ]
> Hello,
>
> so we had a meeting about the future of "Next Generation Hidden Services" aka prop224.
> It was a good meeting.
>
> We spent most of the time discussing the topics brought up here:
>   https://lists.torproject.org/pipermail/tor-dev/2016-March/010534.html
> Please read the above mail to get up to speed with the topics of discussion.
>
> <snip>
>
> b) In prop224, why do intro points need to know the "intro point encryption key"
>    and also what's the point of UPDATE-KEYS-SUBCMD?
>
>    Nick told us that the main point of UPDATE-KEYS-SUBCMD is that so hidden
>    services can rotate their intro point encryption key periodically, so that
>    they can reset their replay caches. 
>
>    That's a fair point. The big question here is, is this worth the complexity
>    that MAINT_INTRO and UPDATE-KEYS-SUBCMD add to the protocol logic?
>
>    Unfortunately we don't know the answer to this yet, because we actually
>    don't have stats on how our replay caches are doing currently. Here are some
>    things we need to look into:
>
>    i) Speak with facebook or some other big onion hoster, and try to get some
>       information about their replay cache. Is it ready to bust? How much space
>       is it occupying?
>
>    ii) The above will help us gauge replay caches in terms of normal activity,
>        but we also need to think of how replay caches should act in cases of a DoS.
>
>    iii) Also think about INTRODUCE1 replay attacks in general, and understand
>         how dangerous they can be, so that we know how robust our replay caches
>         should be. Is it _that_ dangerous if an attacker manages to sneak in a
>         replay attack every once in a while? 
>
>         For example, imagine an attacker who fills up our replay cache on
>         purpose, so that she can sneak in one replay attack when we reset
>         it. If an attacker can send legit INTRODUCE2 cells to fill up our
>         replay cache, what does she need the replays for? What other types of
>         attacker are there?
>
>    After we learn more about the above, we will understand *how much we care*
>    about onion services being able to rotate intro pint encryption keys on the
>    fly using UPDATE-KEYS-SUBCMD. Here are some possible scenarios here:
>
>    - We end up deciding that INTRODUCE1 replay attacks are *dangerous* and we
>      need to seriously defend ourselves against them. In this case, having some
>      sort of UPDATE-KEYS-SUBCMD mechanism sort of makes sense, since onion
>      services should be able to rotate keys at will and let their clients know,
>      without them having to fetch a new HS descriptor.
>
>    - We end up deciding that INTRODUCE1 are a concern but not something we
>      should complicate our design considerably for. In this case, we might want
>      to consider improving the quality of our replay caches, by using more
>      compact designs (= more space for entries), or by using bloom filters or a
>      similar tech. A proposal will need to be made if so. Yawning posted the
>      following relevant paper: https://home.kookmin.ac.kr/~mkyoon/publications/2010TKDE.pdf
>
>    - We end up deciding it's no big deal if the attacker sneaks in a replay
>      attack every once in a while. In this case, we can just go with the most
>      minimal approach and ditch the UPDATE-KEYS-SUBCMD mechanism for now, and
>      have intro points be oblivious about encryption keys (which further
>      simplifies our cell formats, etc.). If we take this approach, and in the
>      future future we change our minds about the importance of replay attacks,
>      we can still add some sort of UPDATE-KEYS-SUBCMD mechanism using the
>      extensions mechanism of ESTABLISH_INTRO cells.
>    

Hm. Unfortunately I didn't receive any feedback on this matter, and we need to
move forward here, so I thought about this some more on my own. I'm still not
sure what's the right thing to do here, but here are my current thoughts:

First of all, the UPDATE-KEYS-SUBCMD mechanism does not actually prevent replay
attacks at all. It's a performance feature. It's there so that if the HS
rotates its replay cache and encryption key, the intro point can inform clients
about the new encryption key. The win here is performance, since that clients
don't need to fetch a new descriptor to learn the new encryption key.

Howevr, the performance gain from UPDATE-KEYS-SUBCMD does not seem substantial
since it only occurs in the case where a client has already fetched an HS
descriptor, and then the HS rotates its keys, and then the client wants to
revisit that HS.

If UPDATE-KEYS-SUBCMD is implemented, then the intro point will pass the new
keys to the client, and the client will be able to connect to the HS by issuing
a new INTRODUCE1 cell.

On the other hand, if we don't implement the UPDATE-KEYS-SUBCMD mechanism, then
the client will need to send an INTRODUCE1 cell to another intro point. This is
not terrible performance-wise. However, in the worst case, where the HS has
rotated the encryption keys of *all* its intro points, then the client will
need to try all of them, fail all of them, and fetch a new descriptor. This
worst case scenario should never really happen, except if the HS is under
INTRODUCE1 DoS and rotates its keys more frequently than usual. But why would
an attacker do that?  To delay the clients?

So basically, UPDATE-KEYS-SUBCMD does not offer much, except in the case of a
targetted DoS. This DoS attack does not seem very plausible to me, since its
only purpose is to delay clients. I don't see how it can completely block
access to an HS, except if the attacker is able to *instantly* exhaust the
replay cache of an HS which should not be the case.

Hence, my current intuition is to kill the UPDATE-KEYS-SUBCMD feature for the
sake of simplicity. It seems like a dirty and not trivial to implement feature
that so far only helps in the case of a hypothetical DoS attack which is not
very profitable (fwiw it's also possible with the current hidden services
subsystem)

Now, if I'm wrong and this hypothetical DoS attack ever becomes a real problem,
we can then implement a feature like UPDATE-KEYS-SUBCMD by using the extension
fields of ESTABLISH_INTRO and INTRODUCE_ACK. I feel that using extension fields
is cleaner than the way UPDATE-KEYS-SUBCMD is currently specified (as a weird
subtype of ESTABLISH_INTRO).

What do you people think? Do you feel this suggestion is too sloppy? I'm really
not sure myself.

PS: FWIW, if we are actually serious about INTRODUCE1 replay attacks and we
    want to block them no matter what, we need to either change our detection
    method from replay caches to something else, or at least make replay caches
    more compact so that they fit more elements.