> One other point here, is that we should try at least one interop experiment 
> with the PQC algorithm, whether it's composite mldsa44-ed25519-sha512 or 
> mldsa44-sha512.   This choice should be discussed at the interim Wednesday.  
> -Wei


Some implementation experience that may help before Wednesday's interim:

I have implemented mldsa44-ed25519-sha512 in mail-auth, the Rust library behind 
Stalwart's DKIM2 support, as part of updating it to 
draft-ietf-dkim-dkim2-spec-07. It is available for interop testing in the 
mail-auth playground at https://mail-auth.stalw.art, where you can generate 
keys, sign and verify messages with rsa-sha256, ed25519-sha256 and 
mldsa44-ed25519-sha512. I am happy to share keys and test messages with anyone 
who wants to try it against their own implementation.

On the "implementing your own crypto" concern: I did not have to implement any 
primitive myself. ML-DSA-44 and Ed25519 come from AWS-LC (and, in a second 
backend, from the RustCrypto ml-dsa and ed25519-dalek crates). The composite 
part from draft-ietf-lamps-pq-composite-sigs is only the construction of M' (a 
fixed prefix, the label, the context length and context, and the SHA-512 
pre-hash) plus splitting and concatenating keys and signatures. Both backends 
pass the test vectors in Appendix E of the LAMPS draft. The one library feature 
it needs is pure ML-DSA with a caller-supplied context string, since the LAMPS 
label is used as the ML-DSA context. AWS-LC supports that in its C API, 
although the aws-lc-rs wrapper does not expose it yet, so I call AWS-LC 
directly for that step. OpenSSL has the same building blocks: it has no 
composite algorithm as such, but OpenSSL 3.6 signs and verifies pure ML-DSA-44 
with a caller-supplied context string (for example "openssl pkeyutl -sign 
-rawin -p
 keyopt context-string:COMPSIG-MLDSA44-Ed25519-SHA512"), so the composite can 
be built on it in the same way. Performance is not a concern either: on my 
machine a composite signature takes about 46 microseconds to sign and 30 
microseconds to verify with AWS-LC, against about 18 microseconds for an 
Ed25519 verification.

While implementing it I found some points the current texts leave open. I think 
they need answers whichever PQ algorithm the group picks, because most of them 
apply to a plain mldsa44-sha512 as well.

1. What is signed. Section 3.4 of -07 says the algorithm "computes a hash over 
all the Message-Instance and DKIM2-Signature fields as described in Section 
12.8 using SHA-512 [FIPS-180-4] as the hash-alg. It signs the hash with both 
the "post-quantum" MLDSA-44 algorithm [FIPS-204] and the PureEdDSA variant 
Ed25519 (see [RFC8032])." That mirrors Section 3.3 and does not cite LAMPS; 
only dns-01 says the algorithm "is specified in" the LAMPS draft, and lists it 
as an informative reference. I can see three readings that do not interoperate:

   a. The LAMPS composite, where its message M is the Section 12.8 signing 
input, and the SHA-512 in Section 3.4 is the LAMPS pre-hash computed outside 
the signer (LAMPS Section 10.5, External Pre-hashing). The input is hashed once.

   b. The LAMPS composite, where M is the SHA-512 digest of the Section 12.8 
input, by analogy with Section 3.3. M' then contains SHA-512(SHA-512(input)).

   c. No LAMPS combiner: plain ML-DSA-44 and Ed25519 each sign the digest and 
the two values are concatenated.

   I implemented (a), which matches Richard's pointer to the LAMPS draft and 
avoids hashing twice. It is also the "pre-hashed input model" from Wei's threat 
model statement, which LAMPS explicitly allows. For a plain mldsa44-sha512 the 
equivalent question is whether pure ML-DSA signs the SHA-512 digest (as Ed25519 
does in Section 3.3) or whether HashML-DSA is meant. Either way, a sentence 
stating exactly what bytes are passed to the signature primitive would remove 
the ambiguity.

2. Context string. Neither draft specifies an application context, so for the 
composite the LAMPS default (empty) applies, and that is what I use. Plain 
ML-DSA also takes a context string. Should DKIM2 define one (for example 
"DKIM2") for domain separation, or is empty deliberate?

3. Key encoding. I publish the raw composite public key in p=, that is the 
1312-byte ML-DSA-44 public key followed by the 32-byte Ed25519 public key, not 
a SubjectPublicKeyInfo structure, which is how I read dns-01 together with 
LAMPS Sections 4.1 and 5. Given the RSA format errata in dns-01 Appendix A, it 
would help to say explicitly whether SPKI is allowed for the new key types.

If the interim decides to go with plain ML-DSA-44 instead, adding it on my side 
is a small change, and I will add it to the playground as well so there is 
something to interop against.


Best,
Mauro

_______________________________________________
Ietf-dkim mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to