shaoninghan / single-neuron-convexification-for-binarized-neural-networks

This repository is the official implementation of Single-neuron convexification for binarized neural networks.

Single-neuron convexification for binarized neural networks

This repository is the official implementation of Single-neuron convexification for binarized neural networks. The codes are divided into two parts -- BNN training and robustness verifying. It is easy to run them in Jupyter Notebook.

Requirements

• A commercial mathematical programming solver Gurobi 9.1 is required to install. It is free for academic users.
• The python dependency file for anaconda is provided in environment.yml.

BNN Training

To train a BNN, implement training.ipynb using Jupyter Notebook or Google Colab Notebook. You can also download the weights for pretrained models here trained on MNST. The accuracy of the trained BNN for each configuration is also provided in the file accuracy.txt.

BNN Robustness Verifying

We formulate the problem of BNN robustness verifying as a mixed integer program (MIP) or its convex relaxation. One can implement verifier_h1.ipynb and verifier_h2.ipynb using Jupyter Notebook.