Re: [lamps] Which PQC KEMs can be used for composite encryption?

["Markku-Juhani O. Saarinen" <[e-mail]>]

From: Mike Ounsworth <[e-mail]>

> I think (but could be wrong) that CCA security is orthogonal to whether
> the KEM output (ie shared secret) can be used directly as a symmetric key,
> or if it needs to be run through a KDF first. For example, I could imagine
> a KEM that outputs an IND-CCA shared secret prefixed with a fixed version
> byte; you can't use that directly as a one-time-pad key because of the
> fixed version byte.
>

Hi Mike and Others,

Yes .. fortunately contemporary (PQC) KEMs do not output such version
bytes; all output bits are significant "full entropy". This is due to the
Fujisaki-Okamoto transform generally used for CCA security and the related
QROM proofs. With "RSA KEMs" things were perhaps a bit different.

If you look at the implementations, the shared secret output is usually
directly from SHA3/SHAKE via the FO transform select logic. This mechanism
is very carefully designed by cryptographers. And as Uri noted, it is a
"part of the algorithm" and KAT test vectors, etc.

On the implementation front, much attention has been paid specifically to
shared secret output due to "decryption failure oracles". Implementations
of a couple of NIST PQC round 3 finalists were broken this way. It is okay
for encapsulation to fail in case of an invalid public key, but
decapsulation "never fails" explicitly, even in case of obviously invalid
ciphertext; a synthetic ("implicit failure") shared secret is generated
instead.

So decapsulation actually produces a "uniform" output even in the case of
nonsense ciphertext. If the shared secret is used with something like AEAD
(for purposes of public-key message encryption/decryption), the decryption
failure occurs only during MAC verification. Having an additional hash or
KDF there is possible, but it needs to be constant time.

Hardware implementations use the SHAKE/SHA3 output to their advantage, as a
masked Keccak module can output the shared secret in masked shares, which
is precisely what one needs for side-channel secure on-chip management of
SSPs.

In software, there's a generic attack on decryption failure oracles (in
shared secret generation).
For timing oracles: https://eprint.iacr.org/2020/743
The algorithm design teams made a bad fix that timing attack.. that allowed
key recovery even without the timing oracle on Round 3 FrodoKem and SIKE
reference implementations.
https://groups.google.com/a/list.nist.gov/g/pqc-forum/c/kSUKzDNc5ME/m/EMFYz9RNCAAJ
https://groups.google.com/a/list.nist.gov/g/pqc-forum/c/QvhRo7T2OL8/m/4gSe5coAAwAJ
Note that both implementations still passed "normal KATs" that do not have
decryption failures.

So I urge people to avoid messing with the internals of KEMs such as shared
secret generation, ciphertext serialization, or key serialization. I can
almost promise you that an innocent-looking functional "optimization" will
lead to security problems unless you're familiar with the security proofs..

(.. Mike's Kyber and SIKE examples..)

>
> So we have two examples of compatible KEMs, and one that isn't directly
> but could be made compatible. But we're not sufficiently expert in KEMs to
> know if this applies only to some PQC KEMs, to all PQC KEMs, or all KEMs
> present and future. Or, to your point, whether this is actually how KEM
> outputs are intended to be used, or if you need to hash them with protocol
> context values first. We believed, from looking at the Kyber and SIKE
> construction that an extra KDF step (and parameter) was unnecessary, but
> we're happy to add it if it improves security or makes this mode more
> generally applicable to more KEM primitives.
>

"Full entropy output" is an assumption that we're also making in the
hardware interface designs, and it seems to apply to all relevant NIST
finalist PQC KEM algorithms.

Cheers,
- Markku

Dr. Markku-Juhani O. Saarinen <[e-mail]> PQShield, Oxford UK.