Hi all,

My colleague Ethan asked me the fun question which post-quantum signature schemes have the following security property, which he called jpeg resistance.

Attacker wins if for a (partially specified) signature and full message, they can find a completed signature and public key, such that the completed signature verifies under the public key.

A naive hash-based signature is not jpeg resistant. Schoolbook Winternitz one-time signatures, forest-of-trees few-time signatures, and Merkle trees all validate signatures (/authentication paths) by recomputing the public key (/Merkle tree root) from the signature and the message, and checking whether the recomputed public key matches the actual public key. That means we can pick anything for the signature, and just set the public key to the recomputed public key.

The situation is more subtle for actual standardized hash-based signatures. RFC 8391 XMSS doesn’t sign the message itself, but first hashes in (among others) the public key. Basically the best we can do for XMSS (except for setting the signature randomizer) is to guess the public key. Thus it’s pretty much jpeg resistant.

The situation is different again for RFC 8391 XMSSMT. XMSSMT is basically a certificate chain of XMSS signatures. An XMSSMT public key is an XMSS public key. An XMSSMT signature is a chain of XMSS signatures: the XMSSMT public key signs another XMSS public key; which signs another public XMSS public key; …; which signs the message. Again the top XMSSMT public key is hashed into the message signed, but that only binds the first XMSS signature. We can’t mess with the first signature, but the other signatures we can choose freely, as those roots are not bound. Thus XMSSMT with two subtrees is only half jpeg resistant and it gets worse with more subtrees.

Similarly SLH-DSA (FIPS 205, née SPHINCS+) is a certificate chain of (a variant of) XMSS signing another XMSS public key, which signs another XMSS public key, etc, which signs a FORS public key, which signs the final message. The SLH-DSA public key is the first XMSS public key. From the message and the public key it derives the FORS key pair (leaf) in the hyper tree to use to sign, and the message to actually sign. This means we can’t mess with the first XMSS keypair. Thus to attack SLH-DSA we honestly generate the first XMSS keypair. Then given a message, we just pick the signature arbitrarily for all but the first XMSS signature. We run the verification routine to recompute the root to sign by the first XMSS keypair. Then we sign it honestly. It depends a bit on the parameters, but basically we get to pick roughly ⅞ of the signature for free.

ML-DSA (FIPS 204, née Dilithium) is a Fiat–Shamir transform of a (module-)lattice identification scheme. In the identification scheme the prover picks a nonce y, and sends the commitment w1 = HighBits(A y) to the verifier, where A is a matrix that’s part of the public key and HighBits drops the lower bits (of the coefficients of the polynomials in the vector). The verifier responds with a challenge c, to which the prover returns the response z = y + c s1, where s1 is part of the private key. The verifier checks, among other things, whether HighBits(Az-ct) = w1, where t = As1+s2 is part of the public key. As usual with Fiat–Shamir, in ML-DSA the challenge c is the hash of the commitment, message, and public key. The scheme has commitment recovery, so the signature itself consists of the response z and the challenge c. (There is also a hint h, but that’s small and we can ignore it.) If we set s1 to zero, then z=y, which is free to choose. So we can freely choose z, which is by far the largest part of the signature. Such a public key t is easy to detect, as it has small coefficients. Instead we can set s1 to zero on only a few components. That allows us to choose z arbitrarily for those components, still breaking jpeg resistance, while being hard to detect. There could well be other approaches here.

Falcon. A Falcon private key are small polynomials f,g. Its public key is h = g f-1. With the private key, for any polynomial c, we can compute small s1 and s2 with s1 + s2h = c. A Falcon signature is a pair r, s2 where s1 = H(r, m) - s2 h is small. s2 is Guassian distributed, and is encoded using an Elias–Fano approach. It’s then padded to make signatures fixed-length. Clearly the randomizer r can be set arbitrarily, but it’s only 40 bytes. Putting arbitrary bytes in most of the encoding of s2 will likely yield a sufficiently small s2. Now, I thought about using this s2 as a new g and construct a signature that way by finding s’1 and s’2 with s’1 + s’2s1f-1 = H(r,m), but my brother suggested a simpler approach. s2 is likely invertible and we can set h = H(r, m)/s2. Both approaches would be thwarted by using H(H(h), r, m) instead of H(r, m). I do not know if there is still another attack.

Best,

 Bas