youben11

"""A basic homomorphic encryption scheme inspired from BFV https://eprint.iacr.org/2012/144.pdf

You can read my blog post explaining the implementation details here: https://www.ayoub-benaissa.com/blog/build-he-scheme-from-scratch-python/

Disclaimer: This implementation doesn’t neither claim to be secure nor does it follow software engineering best practices,
it is designed as simple as possible for the reader to understand the concepts behind homomorphic encryption schemes.
"""

import numpy as np
from numpy.polynomial import polynomial as poly

youben11

import os
from Crypto.Cipher import AES

# Can be 16, 24, or 32 bytes
KEY = b"A"*16

############
# ECB mode #
############
aes_ecb = AES.new(KEY, mode=AES.MODE_ECB)

youben11

evaluate("Encryption is awesome")
+ the sentence was positive (original: 99.99%, simulated: 99.99%, actual: 99.99%, difference: 0.00%, took: 33.813 seconds)

youben11

def evaluate(sentence):
    try:
        embedded = encode(sentence)
    except KeyError as error:
        print("! the word", error, "is unknown")
        return
    if embedded.shape[0] > SENTENCE_LENGTH_LIMIT:
        print(f"! the sentence should not contain more than {SENTENCE_LENGTH_LIMIT} tokens")
        return
    padded = np.zeros((SENTENCE_LENGTH_LIMIT, 300))

youben11

context = homomorphic_inferer.create_context()
keys = context.keygen()

youben11

SENTENCE_LENGTH_LIMIT = 5

inferer = Inferer(model)

homomorphic_inferer = hnp.compile_fhe(
    inferer.infer,
    {
        "x": hnp.encrypted_ndarray(bounds=(-1, 1), shape=(SENTENCE_LENGTH_LIMIT, 300))
    },
    config=hnp.config.CompilationConfig(

youben11

class Inferer:
    def __init__(self, model):
        parameters = list(model.lstm.parameters())
        
        W_ii, W_if, W_ig, W_io = parameters[0].split(HIDDEN_SIZE)
        W_hi, W_hf, W_hg, W_ho = parameters[1].split(HIDDEN_SIZE)
        b_ii, b_if, b_ig, b_io = parameters[2].split(HIDDEN_SIZE)
        b_hi, b_hf, b_hg, b_ho = parameters[3].split(HIDDEN_SIZE)
        
        self.W_ii = W_ii.detach().numpy()

youben11

HIDDEN_SIZE = 100


class Model(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.lstm = torch.nn.LSTM(input_size=300, hidden_size=HIDDEN_SIZE)
        self.fc = torch.nn.Linear(HIDDEN_SIZE, 1)
        self.sigmoid = torch.nn.Sigmoid()

youben11

x_enc= keys.encrypt(x)
res = h.run(keys.public_keys, x_enc)
print(f"Encrypted computation result: {keys.decrypt(res)}")

youben11

print(f"Encrypted computation result: {h.encrypt_and_run(keys, x)}")