HE-SecureNet: Private, Efficient and Usable Homomorphic Encryption-based Model Training

This repository contains the implementation of the paper HE-SecureNet: Private, Efficient and Usable Homomorphic Encryption-based Model Training by Thomas Schneider, Huan-Chih Wang, and Hossein Yalame.

The paper was accepted at the Workshop on Privacy in the Electronic Society (WPES), held in conjunction with ACM CCS 2025, and is available here.

Repository Structure

• cmake - CMake helper files.
• dataset - Breast Cancer and MNIST datasets.
• examples - Private C++ training codes.
  • model_a: Logistic Regression on the Breast Cancer Dataset
  • model_b: MLP on the MNIST Dataset
  • model_c: CNN on the MNIST Dataset
  • model_d: LeNet-5 on the MNIST Dataset
• licenses - Licensing information of the utilized libraries.
• onnx - ONNX graph files.
• parser - Our private training code generator.
• pegasus - The Pegasus backend
• python - Python scripts to generate corresponding ONNX model graph file.
• thirdparty - The thirdparty libraries for Pegasus.

Requirements

• C++ compiler that supports at least C++14 standard
• Cmake >= 3.10
• GMP
• OpenMP
• MNIST Reader
• Python Packages
  • numpy 1.26.4
  • onnx 1.16.2
  • onnxruntime 1.18.0
  • scipy 1.14.1
  • tensosflow 2.19.0
  • tf2onnx 1.16.1

Installation

After cloning or downloading this repository, there are two ways to build HE-SecureNet.

Install Manually

cd HE-SecureNet
# Install Python packages
pip3 install -r requirements.txt
# Download the MNIST dataset
cd dataset
git clone https://github.com/wichtounet/mnist.git
cd ../
# Build HE-SecureNet
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DSEAL_THROW_ON_TRANSPARENT_CIPHERTEXT=OFF
make -j8

Installation with docker

In the HE-SecureNet directory, run docker build -t he-securenet . to build the docker image.

Usage

The following steps demonstrate how to use HE-SecureNet, including:

1. Generating the ONNX file using TensorFlow
2. Creating the private C++ training code with HE-SecureNet’s ONNX parser
3. Executing private model training

We provide an example of training the MLP model, which denoted as "model_b" in this repo.

Generate ONNX graph file

1. Navigate to python folder and run python3 model_b.py to generate the ONNX graph file.
2. The ONNX file is created in the onnx folder.

Generate private C++ training code

1. Navigate to parser folder and run python3 parser.py model_b to get the C++ training code.
2. The private training code is stored in the examples folder.

Execute HE-SecureNet

1. Navigate to build folder and run make -j8 to generate executable binary file.
2. Run ./model_b_exe to start training the model in the ciphertext domain.

Cite HE-SecureNet

@inproceedings{HESecureNET,
  title={{HE-SecureNet: An Efficient and Usable Framework for Model Training via Homomorphic Encryption}},
  author={ Schneider, Thomas and Wang, Huan-Chih, and Yalame, Hossein},
  booktitle={WPES},
  year={2025}
}