Machine learning on homomorphically encrypted data

This repo relies on the TFHE homomorphic encryption library linked at the bottom of this page. We provide two things:

1. An implementation of basic arithmetic operations using TFHE

2. An application of the above operations to logistic regression

Example usage

Package dependencies

Languages: C++, Python

NOTE: This code has only been tested on an Ubuntu 16.04 machine.

C++

This code depends on the TFHE library.

Python

See src/python/requirements.txt From within src/python, run pip install -r requirements.txt to install python dependencies.

Installing TFHE library

1. Clone the TFHE repo to the directory $TFHE_DIR (assuming that's where you've stored the location) and run the following commands:

    cd $TFHE_DIR
    git submodule init
    git submodule update
    mkdir build
    cd build
    cmake .. ${BUILD_FLAGS:-'-DENABLE_FFTW=on -DCMAKE_BUILD_TYPE=optim -DENABLE_TESTS=off'}
    make
   

The library is now built. See tfhe_ml/src/Makefile to see how to compile the program.

Encrypting two integers and adding the results, then decrypting

#include "encryption.hpp"
#include "alu.hpp"
#include <iostream>

int main() {

    typdedef num_type int8_t;
    size_t size = sizeof(num_type) * 8;
    // setup
    const int minimum_lambda = 128;
    TFheGateBootstrappingParameterSet* params = new_default_gate_bootstrapping_parameters(minimum_lambda);
    const TFheGateBootstrappingSecretKeySet* sk = new_random_gate_bootstrapping_secret_keyset(params);
    const TFheGateBootstrappingCloudKeySet* ck = &sk->cloud;

    LweSample *enc_a = new_gate_bootstrapping_ciphertext_array(size, params),
              *enc_b = new_gate_bootstrapping_ciphertext_array(size, params);

    // define integer
    num_type pt = 8, pt2 = 45;
    encrypt<num_type>(enc_a, pt, sk);
    encrypt<num_type>(enc_b, pt2, sk);

    // create new encrypted object to hold sum
    LweSample *sum = new_gate_bootstrapping_ciphertext_array(size, params);


    // add the two integers homomorphically
    add(sum, enc_a, enc_b, ck, size);

    // decrypt and print out
    num_type plain_sum = decrypt<num_type>(sum, sk);

    std::cout << pt << " + " << pt2 << " = " << plain_sum << std::endl;

}

Built-in tests

Two test programs have been written:

1. src/test_arithmetic.cpp - This tests the timing characteristics of the arithmetic library. To build, run make test_arithmetic from within src directory and then run ./test_arithmetic to run.

2. src/test_logistic.cpp - This tests the logistic regression algorithm. To build, run make test_logistic from within src directory and then run ./test_logistic to run.

References

1. TFHE: https://github.com/tfhe/tfhe