TFHE_NN

This package contains the code to run the tests for the paper "Training Encrypted Neural Networks on Encrypted Data with Fully Homomorphic Encryption". All the code has been tested on a Linux machine, with Python 3.9; however, Python 3.8 should work too.

Organization

• pycrcnn: source code for the TFHE-NNs definition and training;
• HE_ML: folder with the experiments. There are three subfolders containing the scripts and datasets for the three considered experiments:
  • HE_ML/Exp1_EncryptedTraining/;
  • HE_ML/Exp2_DistributedTraining/, with the two folders for the "TernaryMNIST" experiment and the "FashionMNIST" one;
  • HE_ML/Exp3_CrossValidation/.
• requirements.txt contains the requirement needed to execute all the code.
• nufhe contains our fork of the nufhe Python library for executing the TFHE gates. After installing requirements.txt, also install the content of this folder in your environment.

Experiments

Exp1

In the Experiment 1 folder you can find res, which contains the used DFA weights (in folder dfa) and the used keys (in folder keys), so that you can replicate the experiment if you want to. The model was trained in the Training.ipynb notebook. It is already executed, so that you can check the cells' output, even though you can re-run that (consider that it will require some days). Nonetheless, you can use the CheckResults.ipynb notebook which compares the model trained on encrypted data with the same model learnt on plain data, with some considerations on the weights' values etc. The serialize encrypted net along with the decrypted weights are in the folder out.

Exp 2

In the Experiment 2 folder you will find the two subfolders for the two used datasets. TernaryMNIST is very similar to Exp 1, even though you will find two TrainingModelX.ipynb notebooks, one for each model, and the Aggregation.ipynb one, which outputs the aggregated encrypted model starting from Model1 and Model2. In the CheckResults.ipynb notebook you will see the numerical values presented in the paper. The serialize encrypted nets along with the decrypted weights are in the folder out. FashionMNIST is a bit different because this experiment has not been run on encrypted data (as specified in the paper). You will find the notebooks TrainingSingleModel.ipynb and TrainingMultipleModels.ipynb, used to train the model on the whole training set and to aggregate different models, respectively. CheckResults.ipynb presents the numerical results shown in the paper, along with additional plots and statistics.

Exp 3

In the Experiment 3 folder you will find the notebooks used to train the four models (i.e., TrainingModelX.ipynb), along with the CrossValidation.ipynb notebook which has been used to perform the encrypted cross-validation procedure. In CheckResults.ipynb you can find the numerical results presented in the paper. The serialize encrypted nets along with the decrypted weights are in the folder out.