HESplitNet

Two-party Privacy-preserving Neural Network Training using Split Learning and Homomorphic Encryption in CKKS Scheme.

Source code of the paper Split Without a Leak: Reducing Privacy Leakage in Split Learning (accepted at the 19th EAI International Conference on Security and Privacy in Communication Networks (SecureComm’23)).

Requirements

python==3.9.7
tenseal==0.3.10
torch==1.8.1+cu102
icecream==2.1.2
h5py==3.7.0
hydra-core==1.1.1
pandas==1.5.2

Repository Structure

├── conf
│   ├── config.yaml  # hold the configurations (dataset to use, hyperparameters)
├── data
│   ├── ptbxl_processing.ipynb  # code to process the PTB-XL dataset
├── images  # images to be used in notebooks and README.md
├── notebooks  # contains jupyter notebooks
├── outputs  # will be automatically created after running the protocols
├── hesplitnet  # contains the code
|   ├── multi-clients  # code for multi clients protocol (experimental)
|   ├── single-client  # code for the single client protocol
└── weights  # contains the initial and trained weights

Data

The processed MIT-BIH dataset for the one-client protocol (mitbih_train.hdf5 and mitbih_test.hdf5) and multi-client protocol (multiclient_mitbih_train.hdf5 and multiclient_mitbih_test.hdf5) are in the data/ folder.

For the PTB-XL dataset, you can run the file data/ptbxl_processing.ipynb to produce the .hdf5 files, but it is recommended that you just download them here and put the .hdf5 files in the data folder.

Running

1. Create a new conda environment, for example conda create -n hesplitnet python=3.9.7 and activate it using conda activate hesplitnet
2. Install the required packages in the requirements.txt, e.g. pip install -r requirements.txt
3. Install hesplitnet as a package using pip install -e .
4. Specify the hyperparamters for your protocol in conf/config.yaml
5. Running the protocol:
   • In the terminal, run python hesplitnet/single-client/server.py. The server will be started and waiting for the client.
   • Open a new tab in the terminal and run python hesplitnet/single-client/client.py. The training process will start.
   • After the training is done, the logs and output files will be saved in the directory outputs/<year_month_day>/<output_dir> where output_dir is defined in conf/config.yaml.
   • After training, go to notebooks/test_mitbih.ipynb or notebooks/test_ptbxl.ipynb to run and inspect the testing procedure for the MIT-BIH or PTB-XL dataset respectively.