Do you think it's worth doing this?

I think authentication of the sender at the HPKE level shouldn't be necessary because the inner activity is self-authenticating (includes integrity proof).

Think 5.1.3 (AuthEncap) is better since it still uses an ephemeral key, but does an extra DH to authenticate the sender. 5.1.1 (Encap) is simpler, but unless the aad contains at least the sender's identity, there is nothing preventing someone from extracting a ciphertext, and re-signing it (forging the origin).

Having a trivial amount of AAD is computationally cheaper than doing the extra DH, but the extra DH would be sub 100 usec, so this is "pick something to taste".

ps: Spent the day chasing LLVM bugs, errors in this would be all mine.

Ah , mea culpa, you're right. 5.1.1 avoids the nonce reuse issue and provides imperfect forward secrecy. I've been fighting u-boot on "new" exotic target the last day so my thinking isn't as sharp as it should be.

Replies should go through the same process (avoids sender having to store the shared secret, less state is more better in this case).

I've read a bit more about DHKEM and now I am even more confused. Isn't encap_key an ephemeral key? The library I am using describes it as such:

https://docs.rs/hpke/latest/hpke/trait.Kem.html#tymethod.encap_with_rng

Derives a shared secret and an ephemeral pubkey that the owner of the reciepint’s pubkey can use to derive the same shared secret.

This "ephemeral pubkey" is what I labeled as encap_key. It is later included in fep0806:cipherData

-----

I think it's okay to have no forward secrecy, the goal of this FEP is to describe a simple encryption mechanism that anyone can implement. I'll just warn readers about its weakness.

If the signature in proof covers fep0806:cipherData, you can do ephemeral_key_pub | ciphertext. and use Encryption to a Public Key.

Because in "Encryption to a Public Key" mode, sender's key is not used.

If I understand the idea correctly, the ephemeral key would need to be added to fep0806:cipherData. That is, instead of encap_key | ciphertext it will be ephemeral_key_pub | encap_key | ciphertext.

No, the recipient would also generates an ephemeral key, and uses that when replying.

So it's always a one shot ephemeral (sender)/static (recipient) HPKE exchange, with sender authenticity guaranteed by the signature. Could also use the shared secret from the original post but force a different IV, but "replies are just the sender and receiver being flipped, code path is the same" is easier.

having the sender generate an ephemeral keypair (include the public key envelope)

So the recipient will need to encrypt to this ephemeral key when replying?

Of an envelope. If actor being changed mid-transit isn't a problem, then omit it from the aad. Including the from/to as part of the AAD binds the ciphertext to the origin/recipient (the HPKE ciphertext does not carry any of this information).

It is worth noting that, as the FEP is specified now, using the sender and recipient's keypairs to do HPKE will result in the same shared secret being generated for every message. This is catastrophic to security as the nonce for the symmetric portion of the HPKE encryption is deterministic (RFC 9180 5.1).

This can be avoided by either: having the sender generate an ephemeral keypair (include the public key envelope), or using a unique HPKE info per envelope. The former is better as it gives imperfect forward secrecy (sender identity key compromise does not reveal past plaintexts), the latter is cheaper to do.

>If the proof covers all of that then it's not strictly needed, but it's cheap paranoia.

Do you mean id/actor/to of a plaintext activity, or of an envelope (EncryptedActivity)?

My understanding is that the properties of an envelope have no effect on security. It can even have a different actor.

I added an example of a plaintext activity: https://codeberg.org/silverpill/feps/src/branch/main/0806/fep-0806.md#plaintext-activity

Also added HPKE mode to the parameter list - mode_base.

The primitive choices are sensible.

The "aad" is sent in the clear (or just public as it is now) so you could do something like "fep-0806-aad" | ':' | <id> | ':' | <actor> |':' | <to>. If the proof covers all of that then it's not strictly needed, but it's cheap paranoia.

I was going to say add version info to the labels, but fep0806:cipherAlgorithm covers that.

https://codeberg.org/silverpill/feps/src/branch/main/0806/fep-0806.md#algorithms

KEM: DHKEM(X25519, HKDF-SHA256)
AEAD: ChaCha20Poly1305
KDF: HKDF-SHA256
"info" string: fep-0806-info
"aad" string: fep-0806-aad

Does it look right?

KEM/AEAD/KDF choices are not my preferences, I just copied them from another HPKE+ed25519 implementation.

WhatsApp.

It's biggest claim to fame is non-group messaging (WireGuard), and the fact that the designer based it on the Signal wire protocol design (same person). The latter research that adds concrete security proofs is an added bonus, as well as decent library availability.

But, there's only so many ways to do NIKE (Non-Interactive Key Exchange), and the IETF/IRTF has a preference for reusing their own standards.

TLDR: Noise over HPKE is a personal preference, I think it's harder to blow your foot off with Noise and it's easier to implement (having done multiple Noise impls, and dealt with HPKE more than I should), but honestly either is a reasonable choice, and if you want something more standardized, HPKE is reasonable.

As far as I can tell, HPKE is used in MLS

HPKE has the benefit of being standardized, and imperfect forward secrecy is the best you can do with 0-RTT (which probably fits the fediverse model better).

Just be very careful about the "use correctly" and it's "fine". I'm partial to Noise (https://noiseprotocol.org/) for these sort of use cases because it has less foot+guns and there's more flexibility, but I wouldn't freak out or scream if I saw HPKE in use.

As a side note, it is indeed safe to convert ed25519 keys to X25519 keys assuming a common sense step is taken when using the latter (https://eprint.iacr.org/2021/509).

HPKE seems to address the issue you pointed out back then: https://www.rfc-editor.org/rfc/rfc9180.html#name-forward-secrecy.

At least if used properly. The library I am using hard-codes "info" and "aad" strings: https://github.com/rustonbsd/ed25519-dalek-hpke/blob/27b010a9ae25ba4ce051510eadfd103a179dd715/src/lib.rs#L98-L104

Anyway, here's an updated FEP: https://codeberg.org/silverpill/feps/src/branch/main/0806/fep-0806.md

@raucao @lain @sozialwelten @koteisaev

Sorry for the delayed response, but yes.

The naive approach would be:

- Convert both party's Ed25519 keys to X25519 keys
- ECDH(Sender static, Recipient static) -> k
- HKDF(k) -> Shared secret

The problem with this approach is that, if either side's private key gets compromised, all historical traffic is trivially decryptable. (Static-Static ECDH)

Perfect forward secrecy (compromise of either key does not give attacker the ability to read past messages) requires at least one extra round trip, which is not great in this setting.

The tweak I suggested is:

- Convert both party's Ed25519 keys to X25519 keys
- Sender generates a ephemeral (throw away) X25519 key
- ECDH(Sender static, Recipient static) -> k_1
- ECDH(Sender ephemeral, Recipient static) -> k_2
- HKDF(k_1 | k_2) -> Shared secret
- Sender discards the ephemeral key

With this improvement, sender long term key compromise does not allow decryption of past traffic, because the ephemeral keypair is long gone into the void (one-way forward secrecy, Static-Static + Ephemeral-Static ECDH).

Note: While the theoretical safety of reusing a Ed25519 key for also doing X25519 was an open question, there are proofs that it is safe.

From what I learned so far, the core principle is pretty much the same in all modern encryption schemes, they differ in how shared symmetric key is generated. Is that correct?

As a note, FEP-0806 is overly simplistic in that it has no forward secrecy.

As an easy improvement, if the sender also generates a per-message ephemeral X25519 keypair, you can do static-static + static-ephemeral KEX and get imperfect forward secrecy (no additional round trips required).