#homomorphic-encryption #discrete-fourier #homomorphic #encryption #fhe

nightly he-ring

A library that provides fast implementations of rings commonly used in homomorphic encryption, built on feanor-math

5 unstable releases

0.3.1 Jun 6, 2024
0.3.0 Jun 6, 2024
0.2.1 May 13, 2024
0.2.0 Apr 10, 2024
0.1.1 Apr 3, 2024

#1391 in Cryptography

MIT license

195KB
3K SLoC

he-ring

Building on feanor-math, this library provides efficient implementations of rings that are commonly used in homomorphic encryption (HE). Our focus lies on providing the building blocks for second-generation HE schemes like BGV or BFV, however most building blocks are also used in other schemes like FHEW/TFHE. In particular, the core component are cyclotomic rings modulo an integer R_q = Z[X]/(Phi_n(X), q). For both q, there are two settings of relevance.

  • q is a relatively small integer, used as "plaintext modulus" in schemes and often denoted by t. For large n, the fastest way to implement arithmetic in these rings is by using a discrete Fourier transform (DFT) over the complex numbers (using floating-point numbers).
  • q is a product of moderately large primes that split completely in R = Z[X]/(Phi_n(X)). This means that R_q has a decomposition into prime fields, where arithmetic operations are performed component-wise, thus very efficiently (this is called "double-RNS-representation"). In this setting, ring elements are usually stored in double-RNS-representation, and only converted back to standard-resp. coefficient-representation when necessary. Such conversions require a number-theoretic transform (NTT) and a implementation of the Chinese Remainder theorem.

Both of these settings are implemented in this library, a general implementation and a specialized one for the case n = 2^k is a power-of-two. In the latter case, the DFTs/NTTs are cheaper, which makes power-of-two cyclotomic rings the most common choice for applications.

Finally, the library also contains an implementation of various fast RNS-conversions. This refers to algorithms that perform non-arithmetic operations (usually variants of rounding) on the double-RNS-representation, thus avoiding conversions.

Disclaimer

This library has been designed for research on homomorphic encryption. I did not have practical considerations (like side-channel resistance) in mind, and advise against using using it in production.

Example

To demonstrate the use of this library, we give an example implementation of the BFV fully homomorphic encryption scheme.

use feanor_math::ring::*;
use feanor_math::rings::zn::*;
use feanor_math::rings::zn::zn_64::*;
use feanor_math::primitive_int::StaticRing;
use feanor_math::integer::*;
use feanor_math::rings::poly::dense_poly::DensePolyRing;
use feanor_math::rings::poly::*;
use feanor_math::algorithms::miller_rabin::is_prime;
use feanor_math::mempool::DefaultMemoryProvider;
use feanor_math::algorithms::fft::*;
use feanor_math::{default_memory_provider, assert_el_eq};
use feanor_math::homomorphism::Homomorphism;
use feanor_math::vector::VectorView;
use feanor_math::rings::extension::FreeAlgebraStore;
use feanor_math::vector::vec_fn::VectorFn;
use feanor_math::rings::float_complex::Complex64;

use he_ring::*;

use rand::thread_rng;
use rand::{Rng, CryptoRng};
use rand_distr::StandardNormal;

// in the spirit of feanor-math, all rings are highly generic and extensible using the type system.
// here we define the ring types we will use to implement the scheme. 
pub type PlaintextRing = complexfft::complex_fft_ring::ComplexFFTBasedRing<complexfft::pow2_cyclotomic::Pow2CyclotomicFFT<Zn, cooley_tuckey::FFTTableCooleyTuckey<Complex64>>, DefaultMemoryProvider, DefaultMemoryProvider>;
pub type FFTTable = doublerns::pow2_cyclotomic::Pow2CyclotomicFFT<cooley_tuckey::FFTTableCooleyTuckey<ZnFastmul>>;
pub type CiphertextRing = doublerns::double_rns_ring::DoubleRNSRing<Zn, FFTTable, DefaultMemoryProvider>;

pub type Ciphertext = (El<CiphertextRing>, El<CiphertextRing>);
pub type SecretKey = El<CiphertextRing>;
pub type GadgetProductOperand<'a> = doublerns::gadget_product::GadgetProductRhsOperand<'a, Zn, FFTTable, DefaultMemoryProvider>;
pub type KeySwitchKey<'a> = (GadgetProductOperand<'a>, GadgetProductOperand<'a>);
pub type RelinKey<'a> = (GadgetProductOperand<'a>, GadgetProductOperand<'a>);

//
// During BFV multiplication, we need a "rescaling operation" that computes `round(x * t / q)`. Doing
// this in a fast-RNS-conversion manner requires precomputing all kinds of data, encapsulated by `MulConversionData`.
//
pub struct MulConversionData {
    to_C_mul: rnsconv::approx_lift::AlmostExactBaseConversion<Zn, Zn, DefaultMemoryProvider, DefaultMemoryProvider>,
    scale_down_to_C: rnsconv::bfv_rescale::AlmostExactRescalingConvert<Zn, DefaultMemoryProvider, DefaultMemoryProvider>
};

const ZZbig: BigIntRing = BigIntRing::RING;
const ZZ: StaticRing<i64> = StaticRing::<i64>::RING;

pub fn sample_primes_in_arithmetic_progression(n: u64) -> impl Iterator<Item = u64> {
    (1..).map(move |i| i * n + 1).filter(|p| is_prime(&ZZ, &(*p as i64), 8))
}

//
// BFV requires arithmetic in two different "ciphertext rings":
//  - the standard ring `R_q` containing all ciphertexts
//  - an "extended" ciphertext ring `R_Q` used during multiplication (where `q` divides `Q` and `Q > q^2`)
// This function creates both of them
//
pub fn create_ciphertext_rings(log2_ring_degree: usize, ciphertext_moduli_count: usize) -> (CiphertextRing, CiphertextRing) {
    let mut primes = sample_primes_in_arithmetic_progression(2 << log2_ring_degree).map(|p| p as u64);

    let rns_base = zn_rns::Zn::new(primes.by_ref().take(ciphertext_moduli_count).map(Zn::new).collect(), ZZbig ,default_memory_provider!());
    
    let rns_base_mul = zn_rns::Zn::new(
        rns_base.get_ring().iter().map(|R| *R.modulus() as u64).chain(
            primes.take(ciphertext_moduli_count)
        ).map(Zn::new).collect(), 
        ZZbig, 
        default_memory_provider!()
    );

    let C = <CiphertextRing as RingStore>::Type::new(rns_base.clone(), rns_base.get_ring().iter().cloned().map(ZnFastmul::new).collect(), log2_ring_degree,default_memory_provider!());
    let C_mul = <CiphertextRing as RingStore>::Type::new(rns_base_mul.clone(), rns_base_mul.get_ring().iter().cloned().map(ZnFastmul::new).collect(), log2_ring_degree,default_memory_provider!());
    return (C, C_mul);
}

pub fn create_plaintext_ring(log2_ring_degree: usize, plaintext_modulus: i64) -> PlaintextRing {
    return <PlaintextRing as RingStore>::Type::new(Zn::new(plaintext_modulus as u64), log2_ring_degree, default_memory_provider!(), default_memory_provider!());
}

//
// Creates the `MulConversionData` required to perform the rescaling during BFV multiplication in a
// fast-RNS-conversion manner.
//
pub fn create_multiplication_rescale(P: &PlaintextRing, C: &CiphertextRing, C_mul: &CiphertextRing) -> MulConversionData {
    MulConversionData {
        to_C_mul: rnsconv::approx_lift::AlmostExactBaseConversion::new(
            C.get_ring().rns_base().iter().map(|R| Zn::new(*R.modulus() as u64)).collect::<Vec<_>>(), 
            C_mul.get_ring().rns_base().iter().map(|R| Zn::new(*R.modulus() as u64)).collect::<Vec<_>>(), 
            default_memory_provider!(),
            default_memory_provider!()
        ),
        scale_down_to_C: rnsconv::bfv_rescale::AlmostExactRescalingConvert::new(
            C_mul.get_ring().rns_base().iter().map(|R| Zn::new(*R.modulus() as u64)).collect::<Vec<_>>(), 
            Some(P.base_ring()).into_iter().map(|R| Zn::new(*R.modulus() as u64)).collect::<Vec<_>>(), 
            C.get_ring().rns_base().len(), 
            default_memory_provider!(),
            default_memory_provider!()
        )
    }
}

pub fn gen_sk<R: Rng + CryptoRng>(C: &CiphertextRing, mut rng: R) -> SecretKey {
    // we sample uniform ternary secrets 
    let result = C.get_ring().sample_from_coefficient_distribution(|| (rng.next_u32() % 3) as i32 - 1);
    return result;
}

pub fn enc_sym_zero<R: Rng + CryptoRng>(C: &CiphertextRing, mut rng: R, sk: &SecretKey) -> Ciphertext {
    let a = C.get_ring().sample_uniform(|| rng.next_u64());
    let b = C.mul_ref(&a, &sk);
    let e = C.get_ring().sample_from_coefficient_distribution(|| (rng.sample::<f64, _>(StandardNormal) * 3.2).round() as i32);
    return (C.add(C.negate(b), e), a);
}

pub fn enc_sym<R: Rng + CryptoRng>(P: &PlaintextRing, C: &CiphertextRing, rng: R, m: &El<PlaintextRing>, sk: &SecretKey) -> Ciphertext {
    hom_add_plain(P, C, m, enc_sym_zero(C, rng, sk))
}

pub fn dec(P: &PlaintextRing, C: &CiphertextRing, ct: &Ciphertext, sk: &SecretKey) -> El<PlaintextRing> {
    let (c0, c1) = ct;
    let noisy_m = C.add_ref_fst(c0, C.mul_ref(c1, sk));
    let coefficients = C.wrt_canonical_basis(&noisy_m);
    let Delta = ZZbig.rounded_div(
        ZZbig.clone_el(C.base_ring().modulus()), 
        &int_cast(*P.base_ring().modulus() as i32, &ZZbig, &StaticRing::<i32>::RING)
    );
    let modulo = P.base_ring().can_hom(&ZZbig).unwrap();
    return P.from_canonical_basis((0..coefficients.len()).map(|i| modulo.map(ZZbig.rounded_div(C.base_ring().smallest_lift(coefficients.at(i)), &Delta))));
}

pub fn hom_add(C: &CiphertextRing, lhs: &Ciphertext, rhs: &Ciphertext) -> Ciphertext {
    let (lhs0, lhs1) = lhs;
    let (rhs0, rhs1) = rhs;
    (C.add_ref(lhs0, rhs0), C.add_ref(lhs1, rhs1))
}

pub fn hom_add_plain(P: &PlaintextRing, C: &CiphertextRing, m: &El<PlaintextRing>, ct: Ciphertext) -> Ciphertext {
    let mut m = C.get_ring().do_fft(C.get_ring().exact_convert_from_cfft(P.get_ring(), m));
    let Delta = C.base_ring().coerce(&ZZbig, ZZbig.rounded_div(
        ZZbig.clone_el(C.base_ring().modulus()), 
        &int_cast(*P.base_ring().modulus() as i32, &ZZbig, &StaticRing::<i32>::RING)
    ));
    C.inclusion().mul_assign_map_ref(&mut m, &Delta);
    let (c0, c1) = ct;
    return (C.add(c0, m), c1);

}

pub fn hom_mul_plain(P: &PlaintextRing, C: &CiphertextRing, m: &El<PlaintextRing>, ct: Ciphertext) -> Ciphertext {
    let m = C.get_ring().do_fft(C.get_ring().exact_convert_from_cfft(P.get_ring(), m));
    let (c0, c1) = ct;
    return (C.mul_ref_snd(c0, &m), C.mul(c1, m));
}

pub fn gen_rk<'a, R: Rng + CryptoRng>(C: &'a CiphertextRing, rng: R, sk: &SecretKey) -> RelinKey<'a> {
    gen_switch_key(C, rng, &C.pow(C.clone_el(sk), 2), sk)
}

pub fn hom_mul(C: &CiphertextRing, C_mul: &CiphertextRing, lhs: &Ciphertext, rhs: &Ciphertext, rk: &RelinKey, conv_data: &MulConversionData) -> Ciphertext {
    let (c00, c01) = lhs;
    let (c10, c11) = rhs;
    let lift = |c: El<CiphertextRing>| C_mul.get_ring().do_fft(C_mul.get_ring().perform_rns_op_from(C.get_ring(), &C.get_ring().undo_fft(c), &conv_data.to_C_mul));

    let lifted0 = C_mul.mul(lift(C.clone_el(c00)), lift(C.clone_el(c10)));
    let lifted1 = C_mul.add(C_mul.mul(lift(C.clone_el(c00)), lift(C.clone_el(c11))), C_mul.mul(lift(C.clone_el(c01)), lift(C.clone_el(c10))));
    let lifted2 = C_mul.mul(lift(C.clone_el(c01)), lift(C.clone_el(c11)));

    let scale_down = |c: El<CiphertextRing>| C.get_ring().perform_rns_op_from(C_mul.get_ring(), &C_mul.get_ring().undo_fft(c), &conv_data.scale_down_to_C);

    let res0 = C.get_ring().do_fft(scale_down(lifted0));
    let res1 = C.get_ring().do_fft(scale_down(lifted1));
    let res2 = scale_down(lifted2);
    
    let op = C.get_ring().to_gadget_product_lhs(res2);
    let (s0, s1) = rk;
    return (
        C.add(res0, C.get_ring().gadget_product(&op, s0)), 
        C.add(res1, C.get_ring().gadget_product(&op, s1))
    );
    
}

pub fn gen_switch_key<'a, R: Rng + CryptoRng>(C: &'a CiphertextRing, mut rng: R, old_sk: &SecretKey, new_sk: &SecretKey) -> KeySwitchKey<'a> {
    let mut res_0 = C.get_ring().gadget_product_rhs_zero();
    let mut res_1 = C.get_ring().gadget_product_rhs_zero();
    for i in 0..C.get_ring().rns_base().len() {
        let (c0, c1) = enc_sym_zero(C, &mut rng, new_sk);
        let factor = C.base_ring().get_ring().from_congruence((0..C.get_ring().rns_base().len()).map(|i2| {
            let Fp = C.get_ring().rns_base().at(i2);
            if i2 == i { Fp.one() } else { Fp.zero() } 
        }));
        let mut payload = C.clone_el(old_sk);
        C.inclusion().mul_assign_map_ref(&mut payload, &factor);
        res_0.set_rns_factor(i, C.add(payload, c0));
        res_1.set_rns_factor(i, c1);
    }
    return (res_0, res_1);
}

pub fn key_switch(C: &CiphertextRing, ct: &Ciphertext, switch_key: &KeySwitchKey) -> Ciphertext {
    let (c0, c1) = ct;
    let (s0, s1) = switch_key;
    let op = C.get_ring().to_gadget_product_lhs(C.get_ring().undo_fft(C.clone_el(c1)));
    return (C.add_ref_fst(c0, C.get_ring().gadget_product(&op, s0)), C.get_ring().gadget_product(&op, s1));
}

let mut rng = thread_rng();

// not a secure choice of parameters
let log2_ring_degree = 7;
let plaintext_modulus = 3;
let ciphertext_moduli_count = 5;

let P = create_plaintext_ring(log2_ring_degree, plaintext_modulus);
let (C, C_mul) = create_ciphertext_rings(log2_ring_degree, ciphertext_moduli_count);

let sk = gen_sk(&C, &mut rng);

// for simplicity, we encrypt only a scalar
let m = P.int_hom().map(2);
let ct = enc_sym(&P, &C, &mut rng, &m, &sk);

let mul_rescale_data = create_multiplication_rescale(&P, &C, &C_mul);
let relin_key = gen_rk(&C, &mut rng, &sk);
let ct_sqr = hom_mul(&C, &C_mul, &ct, &ct, &relin_key, &mul_rescale_data);

let m_sqr = dec(&P, &C, &ct_sqr, &sk);
assert_el_eq!(&P, &P.int_hom().map(1), &m_sqr);

Dependencies

~1.5MB
~30K SLoC