deepmind-research/unrestricted_advx/README.md

Unrestricted Adversarial Challenge

This is a submission for the unrestricted adversarial challenge: Phase I. The entry is uses a pretrained ImageNet model with Local Linearity Regularizer, then adversarially trained using birds-or-bicycles dataset (train and extras) as provided by the challenge.

Tom B. Brown et al Unrestricted Adversarial Examples. [arXiv]. Chongli Qin et al Adversarial Robustness through Local Linearization. NEURIPS 2019. [arXiv]

Contents

The code contains:

• a main file (main.py) for our submission.

Running

Install requirements please follow instructions on Unrestricted Adversarial Challenge.

You can do this by running:

./unrestricted_advx/install_dependencies.sh

You can run the main script by:

./unrestricted_advx/run.sh