Reproducing the Cryptanalytic Extraction of Neural Network Models

This repository presents an alternative implementation of the model extraction attack discussed in the CRYPTO'20 paper.

The original implementation can be found at the following link:

https://github.com/google-research/cryptanalytic-model-extraction

The paper can be accessed at the following link:

https://arxiv.org/abs/2003.04884

Preparations

To get started you will need to install some dependencies. It should suffice to run

pip install torch numpy

Extracting Example Models

First, generate a ONE-DEEP model that we will extract by running

python3 train_model——pytorch.py 10-15-15-1 42

and a trained model will be available in the "models/" directory.

Next, we can employ the ONE-DEEP model to generate predictions using input data through the following command:

python3 trained_model_pytorch_predict.py 10-15-1