#pqc #post-quantum #cryptography #decoding

classic-mceliece-rust

Pure rust implementation of the PQC scheme Classic McEliece

5 stable releases

Uses new Rust 2021

2.0.1 Sep 8, 2022
2.0.0 Sep 5, 2022
1.1.0 Sep 5, 2022
1.0.1 Apr 12, 2022
1.0.0 Apr 1, 2022

#41 in Cryptography

Download history 2/week @ 2022-06-15 1/week @ 2022-06-22 1/week @ 2022-06-29 3/week @ 2022-07-06 4/week @ 2022-07-13 3/week @ 2022-07-20 9/week @ 2022-07-27 23/week @ 2022-08-03 1/week @ 2022-08-17 2/week @ 2022-08-24 40/week @ 2022-08-31 483/week @ 2022-09-07 405/week @ 2022-09-14 284/week @ 2022-09-21 221/week @ 2022-09-28

1,424 downloads per month

MIT license

10MB
5K SLoC

classic-mceliece-rust

This is a pure-rust safe-rust implementation of the Classic McEliece post-quantum scheme.

  • Classic McEliece is a code-based key encapsulation mechanism (KEM)
  • The implementation is based on the Classic McEliece reference implementation of NIST round 3
  • The implementation does not utilize any concurrency techniques (SIMD/threading/…, except maybe auto-vectorization on your CPU)
  • It depends on sha3 as SHA-3 implementation and aes as AES block cipher (used as RNG) implementation
  • It passes the 100 testcases of the C reference implementation
  • It implements all 10 variants of the Classic McEliece KEM
  • The implementation takes between 100 milliseconds (mceliece348864) and 500 milliseconds (mceliece8192128f) to run on a modern computer
  • The implementation is constant-time on software instruction level
  • The random number generator is based on AES256 in counter mode
  • First described in 1978, the cryptographic scheme has a rich history in security analysis. Its large public key size, however, often limits adoption.

The 10 variants have the following designated identifiers:

  • mceliece348864
  • mceliece348864f
  • mceliece460896
  • mceliece460896f
  • mceliece6688128
  • mceliece6688128f
  • mceliece6960119
  • mceliece6960119f
  • mceliece8192128
  • mceliece8192128f

Who should use it?

Anyone, how wants to use Classic McEliece to negotiate a key between two parties.

How does one use it storing keys on the heap (default feature alloc)?

Add this to your Cargo.toml:

[dependencies]
classic-mceliece-rust = "2.0"

To use a specific Classic McEliece variant, you need to import it with the corresponding feature flag:

[dependencies]
classic-mceliece-rust = { version = "2.0", features = ["mceliece6960119"] }

Assuming this dependency, the simplest and most ergonomic way of using the library is with heap allocated keys and the *_boxed KEM step functions. First, we import them:

use classic_mceliece_rust::{keypair_boxed, encapsulate_boxed, decapsulate_boxed};

Followingly, we run the KEM and provide generated keys accordingly. Here, we consider an example where we run it in a separate thread (be aware that the example also depends on the rand crate):

fn run_kem() {
  let mut rng = rand::thread_rng();

  // Alice computes the keypair
  let (public_key, secret_key) = keypair_boxed(&mut rng);

  // Send `secret_key` over to Bob.
  // Bob computes the shared secret and a ciphertext
  let (ciphertext, shared_secret_bob) = encapsulate_boxed(&public_key, &mut rng);

  // Send `ciphertext` back to Alice.
  // Alice decapsulates the ciphertext...
  let shared_secret_alice = decapsulate_boxed(&ciphertext, &secret_key);

  // ... and ends up with the same key material as Bob.
  assert_eq!(shared_secret_bob.as_array(), shared_secret_alice.as_array());
}

fn main() {
  std::thread::Builder::new()
    // This library needs quite a lot of stack space to work
    .stack_size(2 * 1024 * 1024)
    .spawn(run_kem)
    .unwrap()
    .join()
    .unwrap();
}

Pay attention that public keys in Classic McEliece are huge (between 255 KB and 1.3 MB). As a result, running the algorithm requires a lot of memory. You need to consider where you store it. In case of this API, the advantages are …

  • you don't need to handle the memory manually
  • on Windows, the call to keypair uses more stack than is available by default. Such stack size limitations can be avoided with the heap-allocation API (see Windows remark below).

How does one use it storing keys on the stack (disabled feature alloc)?

The other option is that you exclude the heap-allocation API and use the provided stack-allocation API. Its advantages are:

  • stack allocation also works in a no_std environment.
  • on some microcontroller platforms, a heap is not available.
  • stack [de]allocation in general is faster than heap [de]allocation

Thus, in this section we want to show you how to use this API without the heap. For this, you need to disable feature alloc which is enabled per default (this line retains default feature zeroize but removes default feature alloc):

classic-mceliece-rust = { version = "2.0", default-features = false, features = ["zeroize"] }

How does one use the API then (be aware that the example also depends on the rand crate)?

use classic_mceliece_rust::{keypair, encapsulate, decapsulate};
use classic_mceliece_rust::{CRYPTO_BYTES, CRYPTO_PUBLICKEYBYTES, CRYPTO_SECRETKEYBYTES};

fn main() {
  let mut rng = rand::thread_rng();

  // Please mind that `public_key_buf` is very large.
  let mut public_key_buf = [0u8; CRYPTO_PUBLICKEYBYTES];
  let mut secret_key_buf = [0u8; CRYPTO_SECRETKEYBYTES];
  let (public_key, secret_key) = keypair(&mut public_key_buf, &mut secret_key_buf, &mut rng);

  let mut shared_secret_bob_buf = [0u8; CRYPTO_BYTES];
  let (ciphertext, shared_secret_bob) = encapsulate(&public_key, &mut shared_secret_bob_buf, &mut rng);

  let mut shared_secret_alice_buf = [0u8; CRYPTO_BYTES];
  let shared_secret_alice = decapsulate(&ciphertext, &secret_key, &mut shared_secret_alice_buf);

  assert_eq!(shared_secret_bob.as_array(), shared_secret_alice.as_array());
}

Here, you can see how the keys are allocated explicitly.

A remark on Windows

If you want your program to be portable with stack allocation and not unexpectedly crash, you should probably run the entire key exchange in a dedicated thread with a large enough stack size. This code snippet shows the idea:

std::thread::Builder::new()
    .stack_size(4 * 1024 * 1024)
    .spawn(|| {/* Run the KEM here */})
    .unwrap();

Feature kem: RustCrypto APIs

If the kem feature is enabled, key encapsulation and decapsulation can also be done via the standard traits in the kem crate.

Feature zeroize: Clear out secrets from memory

If the zeroize feature is enabled (it is by default), all key types that contain anything secret implements Zeroize and ZeroizeOnDrop. This makes them clear their memory when they go out of scope, and lowers the risk of secret key material leaking in one way or another.

Please mind that this of course makes any buffers you pass into the library useless for reading out the key from. Instead of trying to fetch the key material from the buffers you pass in, get it from the as_array method.

#[cfg(not(windows))] {
    use classic_mceliece_rust::keypair;
    use classic_mceliece_rust::{CRYPTO_PUBLICKEYBYTES, CRYPTO_SECRETKEYBYTES};

    let mut rng = rand::thread_rng();

    let mut pk_buf = [0u8; CRYPTO_PUBLICKEYBYTES];
    // Initialize to non-zero to show that it has been set to zero by the drop later
    let mut sk_buf = [255u8; CRYPTO_SECRETKEYBYTES];

    // This is the WRONG way of accessing your keys. The buffer will
    // be cleared once the `PrivateKey` returned from `keypair` goes out of scope.
    // You should not rely on that array for anything except providing a temporary storage
    // location to this library.
    #[cfg(feature = "zeroize")]
    {
        let (_, secret_key) = keypair(&mut pk_buf, &mut sk_buf, &mut rng);
        drop(secret_key);
        // Ouch! The array only has zeroes now.
        assert_eq!(sk_buf, [0; CRYPTO_SECRETKEYBYTES]);
    }

    // Correct way of getting the secret key bytes if you do need them. However,
    // if you want the secrets to stay secret, you should try to not read them out of their
    // storage at all
    {
        let (_, secret_key) = keypair(&mut pk_buf, &mut sk_buf, &mut rng);
        assert_ne!(secret_key.as_array(), &[0; CRYPTO_SECRETKEYBYTES]);
    }
}

How does one run it?

This library comes with two examples:

$ cargo run --example basic

The output annotates messages with Alice/Bob to illustrate which data is processed by which party. The katkem example implements the classic request/response file structure which is part of the NIST PQC framework.

$ cargo run --example katkem PQCkemKAT_935.req PQCkemKAT_935.rsp
$ cargo run --example katkem PQCkemKAT_935.rsp

The different variants can be enabled through feature flags:

$ cargo run --example katkem --features mceliece6960119 -- PQCkemKAT_1450.req PQCkemKAT_1450.rsp

mceliece348864 is the default variant. You cannot enable two variants simultaneously.

How fast is it?

All data uses clock cycles as unit (the smaller the better). The rust implementation yielded the following runtime results:

complete KEMkeypairencdec
mceliece348864460,062,191439,682,143222,42442,046,357
mceliece348864f244,943,900203,564,820215,97141,648,773
mceliece4608961,326,425,7841,434,864,061487,522111,547,716
mceliece460896f789,636,856652,117,200553,301106,521,703
mceliece66881283,188,205,2662,596,052,574785,763202,774,928
mceliece6688128f1,236,809,0201,059,087,715826,899203,907,226
mceliece69601192,639,852,5732,532,146,1263,864,285203,959,009
mceliece6960119f1,165,079,187965,134,5463,416,795197,089,546
mceliece81921283,129,183,2622,754,933,130965,822247,083,745
mceliece8192128f1,342,438,4511,150,297,5951,068,317242,545,160

The C reference implementation yielded the following runtime results:

complete KEMkeypairencdec
mceliece348864434,103,000437,187,000187,55773,801,300
mceliece348864f252,423,000180,235,000189,52273,668,000
mceliece460896760,993,000894,497,000298,041154,507,000
mceliece460896f606,225,00044,906,000297,743154,013,000
mceliece66881281,568,900,0001,780,660,000425,50429,575,000
mceliece6688128f109,471,000760,298,000414,358298,173,000
mceliece69601193,405,730,0001,694,410,000840,598287,154,000
mceliece6960119f1,311,130,000942,987,000984,660303,543,000
mceliece81921281,635,550,000760,619,000428,112361,999,000
mceliece8192128f1,772,530,0001,222,720,000534,503392,729,000

The tests were done on a Lenovo Thinkpad x260 (Intel Core i5-6200U CPU @ 2.30GHz). In the case of rust, criterion 0.3.5 has been used as given in benches/ and in case of C, Google's benchmark with PFM support and disabled CPU frequency scaling. You can run the benchmark suite yourself with the bench subcommand and optionally some variant feature flag:

$ cargo bench --features mceliece348864

Is it correct?

Yes, besides passing unittests (derived from the C implementation), the generated KAT KEM test files have equivalent MD5 hashes. Namely …

variantexpected MD5 hash
mceliece348864d2def196fde89e938d3d45b2c6f806aa
mceliece348864f84b5357d8dd656bed9297e28beb15057
mceliece4608968aac2122916b901172e49e009efeede6
mceliece460896fd84d3b179e303b9f3fc32ccb6befb886
mceliece6688128b86987d56c45da2e326556864e66bda7
mceliece6688128fae1e42cac2a885a87a2c241e05391481
mceliece69601199d9b3c9e8d7595503248131c584394be
mceliece6960119fc79b1bd28fd307f8d157bd566374bfb3
mceliece8192128b233e2585359a1133a1135c66fa48282
mceliece8192128fd21bcb80dde24826e2c14254da917df3

Where is the source code?

On github.

What is the content's license?

MIT License

Changelog

  • 2022-09-08 version 2.0.1: fix README documentation
  • 2022-09-06 version 2.0.0: refined API with heap-allocated keys and RustCrypto integration
  • 2022-09-06 version 1.1.0: add CI, clippy, infallible SHAKE impl, forbid unsafe code
  • 2022-04-12 version 1.0.1: fix C&P mistakes in documentation
  • 2022-04-01 version 1.0.0: public release (no April fools though)

Where can I ask you to fix a bug?

On github.

Dependencies

~1.5MB
~14K SLoC