ML-KEM / CRYSTALS-Kyber Python Implementation

Caution

⚠️ Under no circumstances should this be used for cryptographic applications. ⚠️

This is an educational resource and has not been designed to be secure against any form of side-channel attack. The intended use of this project is for learning and experimenting with ML-KEM and Kyber

This repository contains a pure python implementation of both:

1. ML-KEM: The NIST Module-Lattice-Based Key-Encapsulation Mechanism Standard following FIPS 203 from the NIST post-quantum cryptography project.
2. CRYSTALS-Kyber: following (at the time of writing) the most recent specification (v3.02)

Note: This project accompanies dilithium-py which is a pure-python implementation of ML-DSA and CRYSTALS-Dilithium and shares a lot of the lower-level code of this implementation.

Disclaimer

kyber-py has been written as an educational tool. The goal of this project was to learn about how Kyber works, and to try and create a clean, well commented implementation which people can learn from.

This code is not constant time, or written to be performant. Rather, it was written so that the python code closely follows the Kyber specification specification and FIPS 203. No cryptographic guarantees are made of this work.

History of this Repository

This work started by simply implementing Kyber for fun, however after NIST picked Kyber to standardise as ML-KEM, the repository grew and now includes both implementations of Kyber and ML-KEM. I assume as this repository ages, the Kyber implementation will get less useful and the ML-KEM one will be the focus, but for historical reasons we will include both. If only so that people can study the differences which NIST introduced during the standardisation of the protocol.

KATs

This implementation currently passes all KAT tests for kyber and ml_kem For more information, see the unit tests in test_kyber.py and test_ml_kem.py.

The KAT files were either downloaded or generated:

1. For ML-KEM, the KAT files were download from the GitHub repository usnistgov/ACVP-Server/ release 1.1.0.35, and are included in assets/ML-KEM-* directories.
2. For Kyber, the KAT files were generated from the projects GitHub repository and are included in assets/PQCLkemKAT_*.rsp

Note: for Kyber v3.02, there is a discrepancy between the specification and reference implementation. To ensure all KATs pass, one has to generate the public key before the random bytes $z = \mathcal{B}^{32}$ in algorithm 7 of the specification (v3.02).

Dependencies

Originally this project was planned to have zero dependencies, however to make this work pass the KATs, we needed a deterministic CSRNG. The reference implementation uses AES256 CTR DRBG. I have implemented this in aes256_ctr_drbg.py. However, I have not implemented AES itself, instead I import this from pycryptodome. If this dependency is too annoying, then please make an issue and we can have a pure-python AES included into the repo.

To install dependencies, run pip -r install requirements.

Using kyber-py

ML-KEM

There are three functions exposed on the ML_KEM class which are intended for use:

• ML_KEM.keygen(): generate a keypair (ek, dk)
• ML_KEM.encaps(ek): generate a key and ciphertext pair (key, ct)
• ML_KEM.decaps(dk, ct): generate the shared key key

Those, together with the ML_KEM_512, ML_KEM_768, and ML_KEM_1024 objects comprise the kyber-py library stable API.

Example

>>> from kyber_py.ml_kem import ML_KEM_512
>>> ek, dk = ML_KEM_512.keygen()
>>> key, ct = ML_KEM_512.encaps(ek)
>>> _key = ML_KEM_512.decaps(dk, ct)
>>> assert key == _key

The above example would also work with ML_KEM_768 and ML_KEM_1024.

Benchmarks

Params	keygen	keygen/s	encap	encap/s	decap	decap/s
ML-KEM-512	1.96ms	511.30	2.92ms	342.26	4.20ms	237.91
ML-KEM-768	3.31ms	302.51	4.48ms	223.04	6.14ms	162.86
ML-KEM-1024	5.02ms	199.07	6.41ms	155.89	8.47ms	118.01

All times recorded using a Intel Core i7-9750H CPU and averaged over 1000 runs.

Kyber

There are three functions exposed on the Kyber class which are intended for use:

• Kyber.keygen(): generate a keypair (pk, sk)
• Kyber.encaps(pk): generate shared key and challenge (key, c)
• Kyber.decaps(sk, c): generate the shared key key

Example

>>> from kyber_py.kyber import Kyber512
>>> pk, sk = Kyber512.keygen()
>>> key, c = Kyber512.encaps(pk)
>>> _key = Kyber512.decaps(sk, c)
>>> assert key == _key

The above example would also work with Kyber768 and Kyber1024.

We expect users to pick one of the three initalised classes which use the default parameters of the Kyber specification. The three options are Kyber512, Kyber768 and Kyber1024. However, by following the values in DEFAULT_PARAMETERS one could tweak these values to look at how Kyber behaves for different default values.

NOTE: it is relatively easy to change the parameters $k$, $\eta_1$, $\eta_2$ $d_u$ and $d_v$ from the Kyber specification. However, if you wish to change the polynomial ring itself, then you will lose access to the NTT transforms which currently only support $q = 3329$ and $n = 256$.

Benchmarks

Params	keygen	keygen/s	encap	encap/s	decap	decap/s
Kyber512	2.02ms	493.99	2.84ms	352.53	4.12ms	242.82
Kyber768	3.40ms	294.13	4.38ms	228.41	6.06ms	165.13
Kyber1024	5.09ms	196.61	6.22ms	160.72	8.29ms	120.68

All times recorded using a Intel Core i7-9750H CPU and averaged over 1000 runs.

Documentation

• https://kyber-py.readthedocs.io/en/latest/

Polynomials and Modules

There are two main things to worry about when implementing Kyber/ML-KEM. The first thing to consider is the mathematics, which requires performing linear algebra in a module with elements in the ring $R_q = \mathbb{F}_q[X] /(X^n + 1)$ and the second is the sampling, compression and decompression, which links to the cryptographic assurance of the protocol.

For those who don't know, a module is a generalisation of a vector space, where elements of a matrix are not selected from a field (such as the rationals, or element of a finite field $\mathbb{F}_{p^k}$), but rather in a ring (we do not require each element in a ring to have a multiplicative inverse). The ring in question for Kyber/ML-KEM is a polynomial ring where polynomials have coefficients in $\mathbb{F}_{q}$ with $q = 3329$ and the polynomial ring has a modulus $X^n + 1$ with $n = 256$ (and so every element of the polynomial ring has at most 256 coefficients).

Polynomials

To help with experimenting with these polynomial rings themselves, the file polynomials_generic.py has an implementation of the univariate polynomial ring

$$ R_q = \mathbb{F}_q[X] /(X^n + 1) $$

where the user can select any $q, n$. For example, you can create the ring $R_{11} = \mathbb{F}_{11}[X] /(X^8 + 1)$ in the following way:

Example

>>> from kyber_py.polynomials.polynomials_generic import PolynomialRing
>>> R = PolynomialRing(11, 8)
>>> x = R.gen()
>>> f = 3*x**3 + 4*x**7
>>> g = R.random_element(); g
5 + x^2 + 5*x^3 + 4*x^4 + x^5 + 3*x^6 + 8*x^7
>>> f*g
8 + 9*x + 10*x^3 + 7*x^4 + 2*x^5 + 5*x^6 + 10*x^7
>>> f + f
6*x^3 + 8*x^7
>>> g - g
0

We hope that this allows for some hands-on experience at working with these polynomials before starting to play with the whole of Kyber/ML-KEM.

For the "Kyber-specific" functions, needed to implement the protocol itself, we have made a child class PolynomialRingKyber(PolynomialRing) which has the following additional methods:

• PolynomialRingKyber
  • ntt_sample(bytes) takes $3n$ bytes and produces a random polynomial in $R_q$
  • decode(bytes, l) takes $\ell n$ bits and produces a polynomial in $R_q$
  • cbd(beta, eta) takes $\eta \cdot n / 4$ bytes and produces a polynomial in $R_q$ with coefficents taken from a centered binomial distribution
• PolynomialKyber
  • encode(l) takes the polynomial and returns a length $\ell n / 8$ bytearray
  • to_ntt() converts the polynomial into the NTT domain for efficient polynomial multiplication and returns an element of type PolynomialKyberNTT
• PolynomialKyberNTT
  • from_ntt() converts the polynomial back from the NTT domain and returns an element of type PolynomialKyber

This class fixes $q = 3329$ and $n = 256$

Lastly, we define a self.compress(d) and self.decompress(d) method for polynomials following page 2 of the specification

$$ \textsf{compress}_q(x, d) = \lceil (2^d / q) \cdot x \rfloor \textrm{mod}^+ 2^d, $$

$$ \textsf{decompress}_q(x, d) = \lceil (q / 2^d) \cdot x \rfloor. $$

The functions compress and decompress are defined for the coefficients of a polynomial and a polynomial is (de)compressed by acting the function on every coefficient. Similarly, an element of a module is (de)compressed by acting the function on every polynomial.

Note: compression is lossy! We do not get the same polynomial back by computing f.compress(d).decompress(d). They are however close. See the specification for more information.

Modules

Building on polynomials_generic.py we also include a file modules_generic.py which has all of the functions needed to perform linear algebra given a ring.

Note that Matrix allows elements of the module to be of size $m \times n$ but for Kyber, we only need vectors of length $k$ and square matrices of size $k \times k$.

As an example of the operations we can perform with out Module lets revisit the ring from the previous example:

Example

>>> R = PolynomialRing(11, 8)
>>> x = R.gen()
>>>
>>> M = Module(R)
>>> # We create a matrix by feeding the coefficients to M
>>> A = M([[x + 3*x**2, 4 + 3*x**7], [3*x**3 + 9*x**7, x**4]])
>>> A
[    x + 3*x^2, 4 + 3*x^7]
[3*x^3 + 9*x^7,       x^4]
>>> # We can add and subtract matrices of the same size
>>> A + A
[  2*x + 6*x^2, 8 + 6*x^7]
[6*x^3 + 7*x^7,     2*x^4]
>>> A - A
[0, 0]
[0, 0]
>>> # A vector can be constructed by a list of coefficients
>>> v = M([3*x**5, x])
>>> v
[3*x^5, x]
>>> # We can compute the transpose
>>> v.transpose()
[3*x^5]
[    x]
>>> v + v
[6*x^5, 2*x]
>>> # We can also compute the transpose in place
>>> v.transpose_self()
[3*x^5]
[    x]
>>> v + v
[6*x^5]
[  2*x]
>>> # Matrix multiplication follows python standards and is denoted by @
>>> A @ v
[8 + 4*x + 3*x^6 + 9*x^7]
[        2 + 6*x^4 + x^5]

On top of this class, we have the classes ModuleKyber(Module) and MatrixKyber(Matrix) which have helper functions which (for example) encode every element of a matrix, or convert every element to or from the NTT domain. These are simple functions which call the respective PolynomialKyber methods for every element.