enc-ct-diagnosis

Source code for Development of Privacy-preserving Deep Learning Model with Homomorphic Encryption: A Technical Feasibility Study in Kidney CT Imaging

• If you don't have access to the HEaaN and HEaaN-math repositories, you won't be able to build. In that case, you can use the enclosed whl file in a limited environment.
• For the using HEaaN and HEaaN-math contact to Cryptolab

Target of Phase 1

For a given PyTorch file or TensorFlow file constructed only as a sequential model. Directly run the HEaaN version's inference.

Layer list

• AvgPooling
• Convolution
• Flatten
• Linear

============ with HEaaN-math ============

• SoftMax
• ReLU
• MaxPooling

How to Run

First, install conda in the system.

Conda env

conda env create --file conda/heml-test-dev
conda activate heml-test-dev

or

conda env create --file conda/heml-run-dev
conda activate heml-run-dev

Build

cmake -S . -B build
cmake --build build -j

Contructed Pytorch

the model must be pack by nn.sequential.

Install HEML python version

Run the following command

python -m pip install -U build wheel
python -m build -w
python -m pip install dist/*.whl

Then we can import HEaaN python

import heaan
from heml_helper import converter_nn as converter

Simple Test

You can see the files written in jupyter notbook.
Note: You may need your own data set.

python/01_save_he_vggnet.ipynb

python/02_run_he_vggnet.ipynb

python/03_run_he_vggnet_auroc.ipynb

python/04_save_he_resnet.ipynb

python/05_run_he_resnet.ipynb

Contributor

Jongmin Choi [email protected]

Hajin Kim [email protected]

Minje Park [email protected]