Mashup of TweetNaCl with ed25519-dalek aimed towards embedded use cases on microcontrollers.

For more context, see also the salty book.

Originally, this library was a transliteration of the C implementation of Ed25519 signatures in TweetNaCl to Rust, "with helpful explanations".

Iterating over the not-very-nice API surface of NaCl, we ended up with a close relative of the "dalek" APIs, where things are modeled as, for instance, "compressed y-coordinate of an Edwards25519 curve point", instead of raw bytes.

One reason the current ed25519-dalek library in its current state is not ideal for microcontrollers is that it includes ~40kB of pre-computed data to speed things up. Moreover, its implementations are optimized for PC.


The main entry point of the API is either a keypair, or a public key.

For keypairs, an external trusted source of entropy is assumed, letting us deterministically construct a keypair as:

let seed: [u8; 32] = [42; 32]; // 32 actually entropic bytes
let keypair: salty::Keypair = salty::Keypair::from(&seed);

Any byte slice of data that fits in memory can then be signed deterministically via

let data: &[u8] = &[1, 2, 3]; // some data
let signature: salty::Signature = keypair.sign(data);

Thereafter, the signature can be checked:

let public_key: salty::PublicKey = keypair.public;
assert!(public_key.verify(data, &signature).is_ok());

For serialization purposes, the entropic seed is the private key (32 bytes). Both public keys and signatures have to_bytes() methods, returning 32 and 64 bytes, respectively.

let serialized_public_key: [u8; 32] = public_key.to_bytes();
let serialized_signature: [u8; 64] = signature.to_bytes();

Conversely, PublicKey implements TryFrom (verifying the alleged point actually lies on the curve), and Signature implements From.

use core::convert::TryInto;
let deserialized_public_key: salty::PublicKey = (&serialized_public_key).try_into().unwrap();
let deserialized_signature: salty::Signature = (&serialized_signature).into();
assert!(deserialized_public_key.verify(data, &deserialized_signature).is_ok());

Please note that Ed25519 signatures are not init-update-finalize signatures, since two passes over the data are made, sequentially (the output of the first pass is an input to the second pass). For cases where the data to be signed does not fit in memory, as explained in RFC 8032 an alternative algorithm Ed25519ph ("ph" for prehashed) is defined. This is not the same as applying Ed25519 signature to the SHA512 hash of the data; it is is exposed via Keypair::sign_prehashed and PublicKey::verify_prehashed. Additionally, there is the option of using "contexts" for both regular and prehashed signatures.


The bulk of time generating and verifying signatures is spent with field operations in the base field of the underlying elliptic curve. This library has two features tweetnacl and haase, which determine the implementation of these field operations. One of these must explicitly be selected. The tweetnacl implementation is portable but quite slow, whereas the haase implementation makes use of the UMAAL assembly instruction, which is only available on Cortex-M4 and Cortex-M33 microcontrollers.

This UMAAL operation is a mapping (a, b, c, d) ⟼ a*b + c + d, where the inputs are u32 and the output is a u64 (there is no overflow). In the future, we hope to offer a third implementation, which would do "schoolbook multiplication", but using this operation, e.g. as a compiler intrinsic. The idea is to have a similarly speedy implementation without the obscurity of the generated assembly code of the haase feature.

Current numbers on an NXP LPC55S69 running at 96Mhz, with "tweetnacl" feature:

  • signing prehashed message: 52,632,954 cycles
  • verifying said message: 100,102,158 cycles
  • code size for this: 19,724 bytes

Obviously, this needed to improve.

Current numbers on an NXP LPC55S69 running at 96Mhz, with "haase" feature:

  • signing prehashed message: 8,547,161 cycles
  • verifying said message: 16,046,465 cycles
  • code size: similar

In both cases, we suggest compiling with at least minimal optimization, to get rid of the zero-cost abstractions.


Future plans include:

  • rigorous correctness checks
  • rigorous checks against timing side-channels, using the DWT cycle count of ARM MCUs
  • ensure dropped secrets are zeroized
  • add the authenticated encryption part of NaCl
  • add X25519, i.e., Diffie-Hellman key agreement
  • speedy yet understandable field operations using UMAAL