Re: [TLS] 0RTT?

[Watson Ladd <[e-mail]>]

On Sun, Aug 3, 2014 at 1:22 PM, Martin Thomson <[e-mail]> wrote:
> This reads a like a halfway between the proposed 0RTT handshake and
> resumption.  The advantage here I'm guessing being that this is
> resumption, which means that you bind to a specific previous session.
>
> However, I keep hearing that the main benefit of resumption is that it
> doesn't require a new DH operation.  And yes, this is still important
> even with the higher performance EC provides.

So I was going to make rekey optional. This wasn't very clear in my
original email.

>
> This relies on an old session, whereas the proposed 0RTT uses a DH
> share that is essentially public, so that it can be cached and shared
> (DNS, signaling).

That has costs with key rotation and administrative version. Maybe
those aren't so bad. But I would like to see resumption be 0RTT also:
right now it is 2RTT in TLS 1.2.

>
> As far as security goes, I don't see any major issues with this.  The
> initial round of application data is going to be as well protected as
> the previous session was.  As long as you aren't able to choose a
> different cipher suite to what was previously used, and I can't see
> why you would want to even contemplate allowing that.

I don't think you are qualified to say that, and the reasoning is
fallacious. You need to mutter something about PRF distinguishing and
collisions, and ideal functionality and simulation. (In particular if
I didn't mix in the old PMS, there would be an attack)

>
> Why do you need a new message?  Since you are using DH, you could just
> use the remainder of the 1RTT handshake structure with
> ServerKeyExchange.  You could, I suppose eliminate the authentication
> messages in order to rely on the previous session.
>
> If the fallback story when the ticket/nonce/share is bad is the
> existing regular handshake, then this is quite close to the proposed
> 0RTT mode.  With a few differences in terms of the first flight keys
> are derived and their lifecycle.

It probably is, but I think I missed the email where that was
discussed. The design space is constrained enough I think we are
converging on a very similar solution.

>
> On Aug 3, 2014 10:09 AM, "Watson Ladd" <[e-mail]> wrote:
>>
>> Dear all,
>>
>> I think I came up with a decent idea for handling 0RTT handshakes,
>> namely making resumption 0RTT, and doing a side rekeying.
>>
>> To see this work we assume that the usual resumption ticket gets set
>> in the handshake, and determines the PMS. A resumption looks as
>> follows:
>>
>> -The ticket
>> -An anti-replay nonce: BLAH bytes
>> -A refresh key: since we know what the server wants, we can provide a
>> group it supports
>>
>> The PMS that gets used is PMS_interim= HMAC(old_PMS, nonce || refresh
>> key): data encrypted as usual under the old ciphersuite follows. (Note
>> that we may have to include the entire Client Hello hashed: this is
>> probably easier, except that the data goes in the Client Hello when
>> using the current inclusion mechanism)
>>
>> On receiving this a server looks up the anti-replay nonce and checks
>> it is fresh. To make this easier we include UTC time since the epoch (
>> midnight of 1 January 1970) and mandate some degree of synchronization
>> and a window. To avoid tagging with clock drift we truncate some low
>> order bits after adding a random small offset.
>>
>> If the nonce is not previously seem the server can send Application
>> Data as normal.
>>
>> We now introduce a new handshake type: Rekey Finish, containing a new
>> ticket and server key. CCS follows afterwards. These are sent
>> encrypted.
>>
>> The new PMS will be HMAC(PMS_interim, ECDH(server, client keys)).
>>
>> The big limitation is ticket keys are going to need rotation. This
>> also doesn't address the desire to put extra data in DNS to give some
>> degree of forward secrecy, but I don't think you can change DNS that
>> quickly without some problems.
>>
>> Furthermore, if we want rekeys without renegotiation, we can reuse
>> Rekey Finish and add a Rekey Initiate handshake type that will send
>> the client key.
>>
>> Open questions: how secure is this? We certainly need to hash the
>> anti-replay nonce into the keys: is that all we need? (Not if we want
>> to avoid attackers manipulating which method we use) Does noisy
>> truncation work to prevent fingerprinting while reducing storage
>> requirements?
>>
>> Sincerely,
>> Watson Ladd
>>
>> _______________________________________________
>> TLS mailing list
>> [e-mail]
>> https://www.ietf.org/mailman/listinfo/tls



-- 
"Those who would give up Essential Liberty to purchase a little
Temporary Safety deserve neither  Liberty nor Safety."
-- Benjamin Franklin