google-deepmind/local_linearity_regularizer

Local Linearity Regularizer for Training Verifiably Robust Models

This repository contains an implementation of the Local Linearity Regularizer (LLR) using TensorFlow v1: https://arxiv.org/abs/1907.02610.

Installation

1. Ensure you have a Python 3.7 environment. (The required version of TensorFlow is not supported from 3.8.) Tools such as pyenv can facilitate multiple Python versions:

   1. Clone the pyenv repository:

      git clone https://github.com/pyenv/pyenv.git ~/.pyenv

   2. Configure shell environment for pyenv, as described at https://github.com/pyenv/pyenv#basic-github-checkout

   3. Install Python 3.7 into pyenv:

      pyenv install 3.7.7

   4. Set Python 3.7 as the current version:

      pyenv shell 3.7.7

      (To revert to the system version, use pyenv shell default.)

2. Clone the repository:

   git clone https://github.com/deepmind/local_linearity_regularizer.git
   cd local_linearity_regularizer

3. Recommended: set up a virtual Python environment:

   python3 -m venv llr_env
   source llr_env/bin/activate

   (To leave the virtual environment, type deactivate.)

4. Install the dependencies:

   pip3 install -r requirements.txt

Usage

The following command trains a model on CIFAR-10 with epsilon set to 8/255:

python3 -m local_linearity_regularizer.train_main --config=local_linearity_regularizer/config.py

Citing this work

If you use this code in your work, we ask that you cite our paper with the following bibtex:

@misc{qin2019adversarial,
    title={Adversarial Robustness through Local Linearization},
    author={Chongli Qin and James Martens and Sven Gowal and Dilip Krishnan
        and Krishnamurthy Dvijotham and Alhussein Fawzi and Soham De and
        Robert Stanforth and Pushmeet Kohli},
    year={2019},
    eprint={1907.02610},
    archivePrefix={arXiv},
    primaryClass={stat.ML}
}

Disclaimer

This is not an official Google product.