Re: [saag] SSH & Ntruprime

["D. J. Bernstein" <[e-mail]>]

Paul Wouters writes:
> Should the IETF really recommend a dropped candidate at this stage?

Yes. IETF policy prefers algorithms with no known patent claims. BCP 79
does not authorize delegating IETF's patent-related decisions to NIST.

Furthermore, the notion that NIST is speaking for a unified community
is easy to disprove. For example,

   https://web.archive.org/web/20230401090854/https://secdev.ieee.org/wp-content/uploads/2022/10/LaMacchia-Keynote-IEEESecDev2022.pdf

revealed that ISO's crypto group agreed---in October 2022, months after
NIST announced its selection---to initiate a preliminary work item on a
very different list of algorithms. There are three algorithms on that
list; one matches a NIST selection (Kyber), one is under consideration
by NIST for possible future standardization (Classic McEliece), and one
was dropped by NIST (FrodoKEM).

> Patent claims are not the issue, as long as the conditions for using
> the patents are not encumbered.

As I wrote before: "there are other patents in the area, including at
least one patent holder publicly claiming Kyber coverage, with no public
response from NIST or from the Kyber team". I quoted and cited a message
that says "Kyber is covered by our patents"; I commented that the author
of that message "holds patents CN107566121 etc., filed before Kyber was
published".

Clearly this qualifies as a "known IPR claim" under BCP 79. I see no
evidence of an "offer of royalty-free licensing" under BCP 79.

> It seems that those will not be an issue as otherwise the NIST chosen
> algorithm would not be useful.

My message already cited examples of the patent minefield to some extent
delaying and to some extent deterring Kyber deployment. If "will" is
alluding to the activation of the patent licenses once NIST actually
issues a standard: sure, that deals with two patents (assuming NIST has
been correctly summarizing the license terms), but the minefield is
bigger than that, as illustrated by the further patent claim above.

> The Crypto Panel review also listed some technical points, which you
> seem to have left out in your latest email

No, I didn't leave them out. I explicitly focused on the Crypto Review
Panel comments regarding sntrup---because I was explicitly replying to
comments you made regarding sntrup. Here's your text (followed by many
more recent references to NIST's actions regarding sntrup):

   With this NTRUprime case, we have a less clear example. It’s not
   broken but the IETF Crypto Panel also said the cryptographic method
   used was somewhat dated and would no longer be recommended by the
   larger cryptographic community at this point.

Your SAAG presentation at IETF 119 claimed that the review had said "we
would have done it like this 15 years ago but these days we wouldn't do
it like this anymore so we shouldn't really like standardize that". 

Looking broadly at how the review as a whole is being used, I see four
basic issues:

   * The review and the followup action both failed to consider the
     patent situation. This is not in line with BCP 79.

   * The portion of the review regarding sntrup was completely
     non-technical, with no evident content beyond delegating IETF/IRTF
     cryptographic decisions to NIST. The review was not "critical,
     objective, timely and consistent review of cryptographic
     algorithms".

   * While I agree that the review did make technical comments regarding
     an issue beyond sntrup (the choice of combiner), those comments are
     not even marginally consistent with how combiners are being handled
     elsewhere in IETF and IRTF. (In case readers are interested in the
     details, see postscript below.)

   * The text of the review does not match what it has been portrayed in
     SAAG as saying.

As an example of the last issue: The SAAG portrayal is that the review
text expressed opposition to documentation and/or standardization of
what has been deployed in real-world SSH. The actual review text

   https://mailarchive.ietf.org/arch/msg/crypto-panel/kDiLLcVOhwoix5BUDdv4r91ZhfY/

sounds much less extreme, with mere "suggestions" to "describe much more
explicitly the combiner use", to add citations, and to "consider"
including Kyber.

As another example, I see nothing in the review text assigning a
positive/negative rating, so it's improper to attribute such a rating to
the review. This rating appears to be something that a particular AD
projected onto the review. The source should be properly labeled.

> The fact that the cryptographic research communities are focusing on
> NIST candidates does mean that those proposed algorithms will see a
> lot more scrutiny and research.

The hypothesis and conclusion of this circular argument are both easily
disproven by the available data. Skimming https://eprint.iacr.org/2024
from top down right now for the ten most recent post-quantum papers, I
find the following:

   https://eprint.iacr.org/2024/564 (attacking isogenies generally)
   https://eprint.iacr.org/2024/561 (an isogeny proposal)
   https://eprint.iacr.org/2024/555 (attacking lattices generally)
   https://eprint.iacr.org/2024/551 (Kyber and NewHope)
   https://eprint.iacr.org/2024/548 (NTRU)
   https://eprint.iacr.org/2024/530 (an NTRU variant)
   https://eprint.iacr.org/2024/523 (Kyber)
   https://eprint.iacr.org/2024/512 (Dilithium)
   https://eprint.iacr.org/2024/500 (SPHINCS+)
   https://eprint.iacr.org/2024/490 (new MPC-based signatures)

A solid half of these are on algorithms that have been either removed by
NIST or that are newer than anything submitted to NIST. Another two are
_overlapping_ NIST but also including other cryptosystems. Only three
fit within the alleged "focus".

> that is not a political argument

The text I quoted from the Crypto Review Panel regarding sntrup is
purely making claims about politics (again, dictionary definition:
"competition between competing interest groups or individuals for power
and leadership"). Making claims that _aren't_ in the text, and saying
that _those_ claims aren't political, doesn't contradict this.

More to the point, my description of the review had nothing whatsoever
to do with the identity of the reviewer, so it wasn't an ad-hominem
attack. Please withdraw your claim to the contrary.

> Some people prefer to not engage with you due to previous negative
> experiences with your method of discussion.

Now _that's_ an ad-hominem attack. Please (1) apologize and (2) keep
yourself under control in the future. Thanks in advance.

Getting back to sntrup: You've referred to secret "informal
conversations" as supposedly justifying opposition to sntrup. Let me
point out that this provides an easy explanation for the gaps between

   * the Crypto Review Panel text and
   * your description of that text.

Specifically, couldn't it be that what you're attributing to the Crypto
Review Panel is actually what's coming from those secret conversations,
and you simply lost track of the source?

Also, have you considered the possibility that the conclusions in those
conversations come from underlying errors that would be corrected if the
arguments were raised in public? Look at the above "scrutiny" claim:
it's the sort of error that can easily be repeated because it _sounds_
reasonable, but transparency allows the claim to be rapidly debunked.

> your statement that Roman promised publication
  [ etc. ]

I don't know what statements you're referring to here; certainly they're
not from me. If you're mixing up the NTRU Prime team, the OpenSSH team,
the author list for this I-D, etc., then please be more careful.

---D. J. Bernstein

P.S. In case readers are interested, here's the combiner issue.

One way to combine pre-quantum and post-quantum shared secrets into a
key for (e.g.) AES-256 or ChaCha20 is to hash the concatenation of the
secrets. This is typically just fine, the main risk being that

   * quantum computers break the pre-quantum system and
   * a bad choice of post-quantum system is also breakable (as in the
     CECPQ2b experiment, which used SIKE to encrypt real user data).

However, there are various papers pointing out contexts where stopping
attacks requires hashing more than the shared secrets. All security
recommendations in these papers are handled by a combiner that hashes
the shared secrets and the full transcript (pre-quantum and post-quantum
public keys and ciphertexts).

Someone reviewing a combiner with anything less than transcript hashing
has to look at the context and ask whether skimping on the hashing is
safe in that context. It's easier for the reviewer to skip this review
and just say "Why aren't you hashing more?". That's what happened in the
Crypto Review Panel review of the combiner in this SSH draft---it wasn't
reviewing whether this is safe in SSH; it was pointing out that this is
doing something that in _some_ contexts is unsafe.

Transcript hashing is cheap. I like making cryptographic choices that
save time for reviewers. So I'd like to see new proposals settling on
_one_ combiner that includes transcript hashing. (To be clear, I don't
see this as an argument against documenting something that has been
widely deployed for two years now.)

Meanwhile there are other people saying that transcript hashing costs
millions of dollars in aggregate and that any unnecessary hash should be
skipped---even if this means that reviewers have to look at different
combiners for different contexts, and check that each of the faster
combiners is safe in the contexts where the combiner is being used.

Here are three examples of combiners not using full transcript hashing:

   * A proposal called "X-Wing" uses an ad-hoc "QSF" combiner. This
     combiner is unsafe in some contexts.

   * draft-ietf-tls-hybrid-design uses a simple concatenation combiner.
     This combiner is unsafe in some contexts.

   * draft-josefsson-ntruprime-ssh uses a simple concatenation combiner,
     This combiner is unsafe in some contexts.

Now compare how this context switch is being handled:

   * X-Wing is currently under consideration by CFRG.

   * My understanding is that draft-ietf-tls-hybrid-design has reached
     consensus except for settling some code points.

   * Meanwhile an AD is opposing draft-josefsson-ntruprime-ssh, where,
     as far as I can tell, the only _technical_ complaint is that it's
     using a concatenation combiner.

This is not even marginally consistent. Sure, X-Wing is in CFRG and
hasn't reached consensus, but this procedural distinction doesn't work
for the TLS example, and it's also missing the point about the content.
The Crypto Review Panel charter asks for "consistent review"; given that
new proposals are being allowed in CFRG and TLS with combiners that can
be unsafe in other contexts, why is draft-josefsson-ntruprime-ssh being
selectively targeted with a complaint that its combiner can be unsafe in
other contexts?