[TLS] Re: New Version Notification for draft-sullivan-tls-xof-ciphers-00.txt

[Joan Daemen <[e-mail]>]

Dear Nick,

Thanks for your mail.

Op 24-07-2026 om 15:30 schreef Nick Sullivan:
> Joan and the Keccak team,
>
> Thank you for the review. Both of your suggestions were more useful 
> than a first read of my reply might suggest, and I want to be clear 
> that neither is closed.
>
> On overwrite-duplex for the derivation: I have not taken it into this 
> revision, but that is a statement about what I have been able to work 
> out so far, not a judgement on the idea. Two things stopped me. The 
> recursion in Lemma 2 appears to break the One-Shot Equivalence the 
> design relies on,

Indeed, Lemma 2 is there to show that you can reduce an attack on OD to 
an attack the underlying XOF ( (Turbo)SHAKE128/256) and therefore that 
the security of OD is covered by the cryptanalysis of the underlying 
XOF. For the (standard) duplex (as introduced in our paper Duplexing the 
sponge: single-pass authenticated encryption and other applications 
https://eprint.iacr.org/2011/499) this reduction is much simpler.

The main reason for using OD instead of standard duplex is that in 
between duplex calls OD has a much smaller state, e.g. for 
(Turbo)SHAKE128 the state of duplex is 200 bytes and that of OD is only 
40 bytes. Our modes on top of keyed OD, UpperDeck/Deck-BO and DWrap do 
quite some cloning and that is of course lighter with a small state. An 
additional benefit is that duplexing in OD where the adversary does not 
get the full output gives forward secrecy: to recover the previous 
state, the missing part of the output must be guessed.  But I guess 
forward secrecy is not essential during the key derivation phase. If it 
is, it can be realized by duplex too by feeding part of the output of a 
duplex call back to the next duplex call input, effective zeroizing part 
of the state.

> and the construction needs raw Keccak-p, which common libraries do not 
> expose,

Access to raw Keccak-p would certainly result in the most efficient 
code. But if a library exposes only the sponge interface the duplex may 
also be inefficient. Namely sponge first does absorb, then switches to 
the squeezing phase and does not allow returning to the absorbing phase. 
You could still do it but for every duplexing call you would have start 
a new sponge instance and feed it with all the duplexing call inputs up 
to that point (see Figure 3 in https://eprint.iacr.org/2011/499) If the 
library already exposes the duplex interface, then building OD on top of 
that could be done by forming the input to a duplexing call as the XOR 
of the output of the previous duplexing call plus the OD input. But this 
would just complicate things and then it would be better to go for 
duplex. But in the long run, if a useful/popular mechanism is defined 
for TLS, then the libraries can be expected to follow.

> plus a |K| <= rho bound on the input.
This limit comes from the fact that we see keyed OD as a keyed primitive 
that takes a key with high entropy per bit, while this is not 
necessarily the case for in key derivation. So what you would use is 
unkeyed OD or duplex and feed it with the secret that can be long and/or 
have low entropy per bit. As this may require multiple blocks, it is 
important to not give access to intermediate duplexing call outputs to 
the adversary as this may allow him to do a divide and conquer attack on 
the long secret. This can all be specified but in our paper we did not 
want to go into that. So you can absorb secrets of any length as long as 
you don't expose intermediate duplexing call outputs to the adversary.
> I would rather have your reading than my own here: if the equivalence 
> can be recovered, or if the bound is looser than I have assumed, I 
> would like to see it. On your ground, so I expect I am the one who is 
> wrong, and I may well be leaving performance on the table. What I 
> would most value is your view on whether a duplex derivation is 
> feasible here at all, and what it would look like.
>
> On the round count: the draft offers both 12-round TurboSHAKE256 and 
> 24-round SHAKE256 and deliberately leaves the choice open, with 
> neither named as the default. You called 12 rounds a comfortable 
> margin. Would you frame the margin any differently for this use: an 
> unkeyed sponge, injective framing, indifferentiable up to 2^256?

Indifferentiability is not a property of a XOF or a hash function, it is 
a property of a construction to build a XOF or a hash function, as it 
assumes an ideal underlying function. So differentiating the sponge 
construction with a random permutation from a random oracle has 
advantage at most N^2/2^{c+1} with N the number of calls to the random 
permutation by the adversary and c the capacity. As soon as you replace 
the random permutation by a concrete permutation you can no longer speak 
of indifferentiability.

This being said, also for this use we think 12 rounds offers a 
comfortable margin based on published third-party cryptanalysis. A lot 
of that is listed at https://keccak.team/third_party.html

>
> On the AEAD: I agree it is worth having, and it is a follow-on rather 
> than part of this document, which is about the key schedule. Perhaps 
> bringing the AEAD itself to CFRG is a route forward.

Indeed, that is separate thing.

Kind regards,

Joan and Gilles on behalf of the Keccak team

>
> Thanks for taking the time to review my rough -00 draft. Hopefully the 
> new draft is closer to a feasible secure design.
>
> Nick
>
> On Tue, Jul 14, 2026 at 1:09 PM Joan Daemen 
> <jda=[e-mail]> wrote:
>
>     Dear all,
>
>     We are enthusiastic about Nick Sullivan's announcement of his RFC
>     draft for a TLS 1.3 key schedule based on Keccak and happy with
>     the many reactions on the mailing list, so we thought it would be
>     good to give you our 2 cents.
>
>     # Including a Keccak-based AEAD option
>
>     In Table 1, the draft proposes AES-GCM and ChaCha20-Poly1305 as
>     AEAD schemes, but no Keccak-based scheme. As suggested by other
>     participants, it would be nice to also offer the option of a
>     Keccak-based AEAD scheme. This would allow one to potentially
>     reduce the code size (or area) and trust surface even further.
>
>     We did the exercise recently in our paper "Shaking up
>     authenticated encryption" presented at EuroS&P
>     (https://eprint.iacr.org/2024/1618) It defines two fully
>     committing AEAD schemes, both with security provably reducible to
>     (Turbo)SHAKE128/256.
>
>
>     # Instantiating the key derivation
>
>     The EuroS&P paper also defines a duplex object and a deck
>     function, both also reducing to the security of
>     (Turbo)SHAKE128/256. Thanks to this reduction, the former could be
>     used as primitives in the key derivation, solving much of the
>     domain separation. The use of "trailer" bytes that accumulate all
>     domain separation bits the final functions are very simple to
>     implement. Moreover, by overwriting input blocks (instead of
>     XORing them in), they have a nice property that each call to the
>     underlying permutation can be a ratchet: the only requirement is
>     that at least 128/256 bits of the output shall not be returned.
>
>
>     # Kravatte vs (Turbo)SHAKE
>
>     Kravatte is a very fast primitive that could be used for AEAD.
>     However, it needs a secret key upon initialization and is
>     therefore not suited for key derivation.
>
>     Kravatte is a deck function obtained by applying the Farfalle
>     construction with Keccak-p[6 rounds] and two rolling functions. It
>     is **not** built on top of Keccak and therefore it security cannot
>     be reduced to that of (Turbo)SHAKE. Actually, its security cannot
>     be reduced to a simpler primitive, so the security of Kravatte
>     must be established by the cryptanalysis of Kravatte itself.
>
>
>     # On the number of rounds
>
>     MarsupilamiFourteen (M14) was given as an example of a function
>     calling Keccak-p with a number of rounds that is not a multiple of 6.
>
>     M14 dates back from 2018 as a 256-bit version of (now called)
>     KT128. The reasoning for adding two rounds was to allow for extra
>     safety margin while giving more budget to the adversary. Since
>     then, we have seen the number of rounds that can be attacked under
>     cryptanalysis slow down, and now we think that 12 rounds provides
>     a comfortable safety margin for Keccak, even when targeting
>     256-bit security with a capacity of 512 bits. So, RFC 9861
>     proposes TurboSHAKE256 and KT256 on top of Keccak-p[12 rounds] and
>     not 14 rounds.
>
>     Note by Joan: I answered Markku indeed that I could not think of
>     any proposal where the round count would not be a multiple of 6,
>     thereby dismissing MarsipulamiFourteen but also our CAESAR AEAD
>     candidate Ketje that does single-round calls in the encryption
>     phase. My mindset was that both were proposals for which we think
>     we have more interesting alternatives.
>
>
>     Kind regards,
>
>     The Keccak team
>     Guido Bertoni, Joan Daemen, Seth Hoffert, Silvia Mella, Michaël
>     Peeters, Gilles Van Assche and Ronny Van Keer
>
>     Op 08-07-2026 om 18:01 schreef Nick Sullivan:
>>     Hi Hannes,
>>
>>     As Thom noted below in the chain, the motivation is to modernize
>>     the key schedule, which has two main advantages:
>>
>>     1. Efficiency gains: As the analysis on-list spells out, it’s a
>>     dramatic improvement to the number of hashes/permutations needed.
>>     But as you noted, is not the hot path at all.
>>     2. Removing a hard dependency on SHA-2 from future designs, as
>>     Thom noted. This gain isn’t immediate, but it clears the way for
>>     future configurations that don’t rely on SHA-2 for the
>>     CertificateVerify to drop SHA-2 completely from the code base.
>>
>>     Nick
>>
>>     On Wed, Jul 8, 2026 at 3:58 PM Thom Wiggers <[e-mail]>
>>     wrote:
>>
>>         Hi Hannes,
>>
>>         I don’t think runtime performance is an issue, but rather
>>         code size (or area), by getting rid of SHA2. (Of course, this
>>         is long into the future). The sponge-based constructions also
>>         have theoretical benefits.
>>
>>         Cheers,
>>
>>         Thom
>>
>>
>>>         Op 8 jul 2026, om 13:47 heeft Hannes Tschofenig
>>>         <hannes.tschofenig=[e-mail]> het volgende
>>>         geschreven:
>>>
>>>         Hi Markku, Hi Nick!
>>>
>>>         I will certainly look closer into the details but it appears
>>>         that you are optimizing TLS in the wrong place. The key
>>>         derivation is the least expensive part in TLS and spending
>>>         time optimizing it will bring little benefit. I am saying
>>>         this because I have for years been looking at optimizing
>>>         different parts of the TLS protocol with constrained IoT in
>>>         mind.
>>>
>>>         This brings me to the core question: What is the problem you
>>>         are trying to solve in the first place? I do not recall that
>>>         anyone has voiced performance problems with the key
>>>         derivation in TLS before this draft was published.
>>>
>>>         Ciao
>>>         Hannes
>>>
>>>
>>>         Am 08.07.2026 um 12:16 schrieb Markku-Juhani O. Saarinen:
>>>>         Hi,
>>>>
>>>>         Thanks for this. I quickly put together an implementation
>>>>         of draft-sullivan-tls-xof-ciphers-00.txt around Rustls to
>>>>         do some measurements:
>>>>
>>>>         https://github.com/mjosaarinen/altkdf-rs
>>>>
>>>>         ( Editorial comments in
>>>>         https://github.com/mjosaarinen/altkdf-rs/blob/main/FINDINGS.md
>>>>         )
>>>>
>>>>         The theoretical side of the design seems very defensible --
>>>>         clean proof target. In terms of concrete security, the
>>>>         Keccak variants have a much larger security margin than the
>>>>         SHA-2 family.
>>>>
>>>>         Given how much work we put into reducing the number of
>>>>         permutation calls with ML-KEM and Hybrid combiners --
>>>>         carefully debating and analyzing each permutation -- this
>>>>         one yields a staggering reduction, making the key schedule
>>>>         much faster (and the handshake probably too.)
>>>>
>>>>         For the representative full handshake: PSK + (EC)DHE +
>>>>         0-RTT leaves + NewSessionTicket + one KeyUpdate each
>>>>         direction + one exporter, the per-endpoint counts over
>>>>         24-round Keccak-f[1600] are:
>>>>
>>>>         41  * f1600: Deck implementation, measured stateful
>>>>         46  * f1600: Deck implementation, measured recompute
>>>>         52  * f1600: Section A.1 in draft-sullivan-tls-xof-ciphers-00
>>>>         156 * f1600: HKDF-SHA3-256 / RFC 8446 baseline
>>>>         117 * f1600: Appendix D "FIPS" KMAC256 schedule
>>>>
>>>>         So 41 vs 156 permutations by my count.
>>>>
>>>>         ( Note: The draft slightly overcounts permutations in its
>>>>         estimates. )
>>>>
>>>>         It's a quick prototype built with extensive AI assistance,
>>>>         but it includes basic correctness measures: primitive KATs
>>>>         (RFC 9861 TurboSHAKE256, FIPS 202 SHAKE256, SP 800-185
>>>>         KMAC256, including multi-block and long-output), 73
>>>>         self-generated Appendix C/D vectors, and byte-for-byte
>>>>         reproduction of all of them by an independent Python
>>>>         implementation written from the draft alone.
>>>>
>>>>         - Keccak-p[1600,nr] permutation and the
>>>>         rate-136/capacity-512 sponge
>>>>         - Five framed deck operations
>>>>         (Init/Absorb/Fork/Squeeze/Ratchet)
>>>>         - KMAC-layout MAC
>>>>         - Three-stage E/H/T schedule with its two ratchets
>>>>         - Section 5 derivations (record keys, Finished/PSK binders,
>>>>         exporters, resumption and key-update, and the §10
>>>>         external-PSK importer with ImportedIdentityV2).
>>>>         - All five cipher suites (0xFF01–0xFF05, both profiles,
>>>>         three AEADs)
>>>>
>>>>         Plus for comparisons:
>>>>
>>>>         - Appendix D FIPS-component schedule (RFC 8446 with KMAC256
>>>>         as the PRF)
>>>>         - a permutation-count benchmark reproducing §A.1,
>>>>         live-secret zeroization (§15.7.2.2)
>>>>
>>>>         Cheers,
>>>>         -markku
>>>>
>>>>         Dr. Markku-Juhani O. Saarinen <[e-mail]>
>>>>
>>>>
>>>>         On Tue, Jul 7, 2026 at 2:34 AM Nick Sullivan
>>>>         <[e-mail]> wrote:
>>>>
>>>>             Dear TLS,
>>>>
>>>>             I'm sharing a draft for the group's consideration.
>>>>             draft-sullivan-tls-xof-ciphers-00 runs the entire TLS
>>>>             1.3 key schedule
>>>>             on a single Keccak permutation, instead of HKDF built
>>>>             on HMAC built on
>>>>             the cipher suite's hash, which today is always SHA-2.
>>>>             This is newly
>>>>             practical because deployments using SHA-3, ML-KEM, or
>>>>             ML-DSA already
>>>>             carry a Keccak permutation, so the primitive is already
>>>>             in the stack.
>>>>
>>>>             Each derived value comes out in one pass, so a full
>>>>             handshake costs
>>>>             about a third of the permutation calls an HKDF schedule
>>>>             over the same
>>>>             permutation would spend.
>>>>
>>>>             A cipher suite names an AEAD plus a schedule profile,
>>>>             and nothing else
>>>>             changes. There is no new extension, and the state
>>>>             machine, record
>>>>             layer, and wire format are untouched. Two profiles are
>>>>             defined, one on
>>>>             the standard SHA-3 function and one on a faster
>>>>             reduced-round variant.
>>>>             Test vectors are pinned to cipher-suite values, so the
>>>>             final vectors
>>>>             will follow the code point assignment.
>>>>
>>>>             https://datatracker.ietf.org/doc/draft-sullivan-tls-xof-ciphers/
>>>>
>>>>             This is a big change to the key schedule, and the draft
>>>>             is very
>>>>             preliminary. Feedback on the approach, or interest in
>>>>             implementing it,
>>>>             would help a lot.
>>>>
>>>>             Best,
>>>>             Nick
>>>>
>>>>             On Mon, Jul 6, 2026 at 7:03 PM
>>>>             <[e-mail]> wrote:
>>>>             >
>>>>             > A new version of Internet-Draft
>>>>             draft-sullivan-tls-xof-ciphers-00.txt has been
>>>>             > successfully submitted by Nick Sullivan and posted to the
>>>>             > IETF repository.
>>>>             >
>>>>             > Name:  draft-sullivan-tls-xof-ciphers
>>>>             > Revision: 00
>>>>             > Title:    TLS 1.3 Cipher Suites with Alternative
>>>>             Key-Schedule Profiles
>>>>             > Date:     2026-07-06
>>>>             > Group:    Individual Submission
>>>>             > Pages:    46
>>>>             > URL:
>>>>             https://www.ietf.org/archive/id/draft-sullivan-tls-xof-ciphers-00.txt
>>>>             > Status:
>>>>             https://datatracker.ietf.org/doc/draft-sullivan-tls-xof-ciphers/
>>>>             > HTML:
>>>>             https://www.ietf.org/archive/id/draft-sullivan-tls-xof-ciphers-00.html
>>>>             > HTMLized:
>>>>             https://datatracker.ietf.org/doc/html/draft-sullivan-tls-xof-ciphers
>>>>             >
>>>>             >
>>>>             > Abstract:
>>>>             >
>>>>             >    TLS 1.3 builds its key schedule on HKDF over the
>>>>             cipher suite's hash.
>>>>             >    This document defines TLS 1.3 cipher suites that
>>>>             build it on a deck
>>>>             >    function over a single permutation instead, the
>>>>             one a deployment
>>>>             >    already carries when it uses SHA-3, ML-KEM, or
>>>>             ML-DSA. One
>>>>             >    permutation then runs the whole schedule, and a
>>>>             full handshake takes
>>>>             >    about a third of the permutation calls an HKDF
>>>>             schedule over that
>>>>             >    permutation would.  Such a cipher suite names an
>>>>             AEAD algorithm
>>>>             >    together with a schedule profile that defines
>>>>             every key-schedule
>>>>             >    function the connection uses.  The profile follows
>>>>             from the
>>>>             >    negotiated cipher suite alone, so no new extension
>>>>             is defined and the
>>>>             >    TLS 1.3 state machine and wire format are
>>>>             unchanged. Two profiles
>>>>             >    are defined, one on the standard SHA-3 function
>>>>             and one on a faster
>>>>             >    reduced-round variant of it.
>>>>             >
>>>>             >
>>>>             >
>>>>             > The IETF Secretariat
>>>>             >
>>>>             >
>>>>
>>>>             _______________________________________________
>>>>             TLS mailing list -- [e-mail]
>>>>             To unsubscribe send an email to [e-mail]
>>>>
>>>>
>>>>         _______________________________________________
>>>>         TLS mailing list [e-mail]
>>>>         To unsubscribe send an email [e-mail]
>>>         _______________________________________________
>>>         TLS mailing list -- [e-mail]
>>>         To unsubscribe send an email to [e-mail]
>>
>>
>>     _______________________________________________
>>     TLS mailing list [e-mail]
>>     To unsubscribe send an email [e-mail]
>     _______________________________________________
>     TLS mailing list -- [e-mail]
>     To unsubscribe send an email to [e-mail]
>