Re: Packet number encryption

["Brian Trammell (IETF)" <[e-mail]>]

> On 30 Jan 2018, at 20:31, Eric Rescorla <[e-mail]> wrote:
> 
> 
> 
> On Tue, Jan 30, 2018 at 6:53 AM, Brian Trammell <[e-mail]> wrote:
> hi Martin,
> 
> Thanks for the PR.
> 
> > On 30 Jan 2018, at 01:59, Martin Thomson <[e-mail]> wrote:
> >
> > Here's a PR:
> >
> > https://github.com/quicwg/base-drafts/pull/1079
> >
> > In short, this builds on Kazuho's excellent suggestion to use ECB-mode
> > to protect the packet number.  It does this only slightly differently,
> > using suggestions from ekr and others to allow for cryptographic
> > agility and use of primitives that match the negotiated AEAD.
> >
> > In short, it takes some of the ciphertext of the packet payload, and
> > uses that to encipher the packet number.  My intuition is that this is
> > pretty solid, but I want to check if the working group is happy that
> > this is generally OK before seeking cryptographic review.
> >
> > For instance, the requirement for side-channel resistance is already
> > controversial: https://github.com/quicwg/base-drafts/pull/1079/files#r164588705
> >
> > For the record, I doubt that this can be proven to be secure without
> > both breaking new ground and adding some constraints to the AEADs that
> > we use.  And that's before we get into actual breaks on the technique
> > that I'm too thick-headed to think of.  In any case, we're relying on
> > this to protect against linkability, so I'd like to ensure that we're
> > not completely crazy in doing so.
> 
> I'm not sure this isn't completely crazy.
> 
> Our model for linkage attacks seems a little too fuzzy to me to be useful for evaluating whether this proposal is an improvement on the status quo. Of course, there's an inherent tension among the identifiability of flows by 5-tuple, the explicit identifiability of flows by CID (which we'll work to remedy at the same time as we stick more bits in the CID, apparently), and the desire for two independent 5-tuple flows to remain unlinkable. So this is inherently hard, because we don't have internally consistent set of requirements for linkability support and avoidance.
> 
> Beyond this, though, we need to have a coherent idea of what we're trying to prevent. From discussions to date, I think these are the two cases we care about:
> 
> (1) An unprivileged on-path device that sees a packet from a flow that is migrated on purpose by an endpoint (i.e., due to connection migration or multipath) should not be able to associate that packet to the prior flow.
> 
> (2) An unprivileged on-path device that sees a packet from a flow that is unintentionally migrated (i.e., one of its globally routable addresses changes without the associated endpoint's knowledge) should not be able to associate that packet to the prior flow.
> 
> In the presence of any connection ID that has enough entropy to be used for load balancing, linkability avoidance in case 2 is impossible. Indeed, the CID exists in part to *allow* linkability in case 2. Any packet number encryption approach is necessarily useless for linkability avoidance in this case, unless the CID is also protected with a secret known only to the endpoints and the load-balancer. We'll see where the CID DT lands here.
> 
> When the connection ID is absent or protected, and the packet number is encrypted, case 2 is still problematic. In an unintentional migration, only two values in the five-tuple change, and it's impossible to make the first packet sent from the migrated endpoint to the non-migrated endpoint look like a handshake packet. So to find an unintentional migration, an observer only needs to look for all flows toward a 3-tuple without being preceded by a valid handshake that started within 30 seconds of another one ending on the same 3-tuple. Even for very large load balancers, this is not a giant anonymity set.
> 
> If you believe that greasing functions over packet numbers is necessary -- I'm not sure I do, but that's a different, future message -- then this proposal might be useful for that, but if linkability prevention is the goal, I'm not convinced it's fit for purpose.
> 
> This proposal also opens a trivial class of wasted-work attacks against QUIC servers by malicious clients, which can now craft packets which will require the server to partially decrypt them before realizing they're not useful. Obviously bad packet numbers can be avoided with simple heuristics. Less-obvious attacks might be crafted to incur significant nonproductive cost on the server side. This seems to be an unavoidable property of any approach to packet number encryption that would actually work, though, so that's not an observation about this specific proposal. I also haven't done the analysis to see if this is all that bad.
> 
> Why isn't this already the case? I just take a valid packet, change a bit in the MAC, and resend it a bunch of times.

It seems to me that the borked MAC replay is easy to distinguish and dynamically filter at the receiver: if you see this, someone is screwing with you, and you can safely blackhole them. On the other hand, a stream of packets that look valid to the decryption stage but that will be rejected as irrelevant but not insane seems harder to reject and blackhole in the general case

Cheers,

Brian


> -Ekr
> 
> 
> Note as well that this proposal comes with a manageability cost. On-path measurement of upstream loss and reordering would become impossible, and the fidelity of the spin bit under loss and reordering would be reduced.
> 
> > In case people were wondering, and because it already came up: I
> > intend to separately propose a change to the short header packet that
> > would remove the type field and move the packet number length under
> > encryption.  See my response to Ian's question on the PR for an
> > explanation of how this will likely work.
> 
> Yep, this makes sense. We'll need to consider these changes closely in concert with any change to the CID. As above, avoiding linkability with PN encryption while aiding it with a CID is hard, and a varlen CID plus a varint encrypted PN will make parsing fiddlier than it already is.
> 
> As Christian noted, using the short header type field to provide a sort of subsequence number or other packet-number independent explicit signal would mitigate the damage packet number encryption would to do loss and reordering observability, but not completely address them.
> 
> Cheers,
> 
> Brian
> 
>