Can We Use Split Learning on 1D CNN for Privacy Preserving Training?

Note: this repo contains our implementation for our ACM ASIACCS 2020 paper below. Please if you find it useful, use the below citation to cite our paper.

@inproceedings{abuadbba2020can, title={Can we use split learning on 1d cnn models for privacy preserving training?}, author={Abuadbba, Sharif and Kim, Kyuyeon and Kim, Minki and Thapa, Chandra and Camtepe, Seyit A and Gao, Yansong and Kim, Hyoungshick and Nepal, Surya}, booktitle={Proceedings of the 15th ACM Asia Conference on Computer and Communications Security}, pages={305--318}, year={2020} }

Available Now:

• Our 1D CNN split learning models with their accuracy results.
• Our pre-processed training/testing samples of MIT arrhythmia ECG database.
• Our privacy leakage 3 measurements results using visual invertibility; distance correlation; and Dynamic Time Warping.
• Our proposed two countermeasures results: i) increasing the number of layers in a CNN model and ii) using differential privacy.

Repository summary

• csv directory: results in csv format from various kinds of experiments.
  • adding_layers directory: experiment results of adding more convolutional layer on 1D CNN.
    • accuracy directory: has best test accuracy data retrieved from each run with different number of convolutional layers.
    • trainlog directory: has train loss, train accuracy, test loss, test accuracy data for each epoch while training 1D CNN having different number of convolutional layers.
  • diffpriv directory: experiment results of applying differential privacy on split layer in 1D CNN.
    • accuracy directory: has best test accuracy data retrieved from applying different strength of differential privacy.
    • trainlog directory: has train loss, train accuracy, test loss, test accuracy data for each epoch while training 1D CNN whose split layer is differential private.
  • measurement directory:
    • dcor directory: has distribution and mean of distance correlation data from each split layer filter.
    • dtw directory: has distribution and mean of DTW data from each split layer filter.
  • split_nonsplit directory: has train log data from split and non-split 1D CNN which are used to prove that they have same results.
• figure directory: source codes in ipynb format which give figure with data in csv directory.
• measurement directory: source codes in ipynb format which measure distance correlation and DTW between raw data and data from split layer filters.
  • adding_layers directory: measure distance correlation with different number of convolutional layers.
  • diffpriv directory: measure DTW with different strength of differential privacy on split layer.
• mitdb directory: has preprocessed train and test data in hdf5 format.