This repository provides code for both training and using the restricted robust resnet models from the paper:
Robustness May Be at Odds with Accuracy
Dimitris Tsipras*, Shibani Santurkar*, Logan Engstrom*, Alexander Turner, Aleksander Madry
https://arxiv.org/abs/1805.12152
We show that adversarial robustness often inevitablely results in accuracy loss. The silver lining: adversarial training induces more semantically meaningful gradients and gives adversarial examples with GAN-like trajectories:
Before trying to run anything:
- Run
./setup.sh
. - Get a downloaded version of the ImageNet training set. By default the code looks for this directory in the environmental variable
INET_DIR
. For example, runexport INET_DIR=/scratch/datasets/imagenet/
.
The code is organized so that you can:
- Train your own robust restricted ImageNet models (via
ez_train.sh
) - Produce adversarial examples and visualize gradients, with example code in
example_use.ipynb
- Reproduce the ImageNet examples seen in the paper (via
make_viz.sh
).
This repository comes with (after following the instructions) three restricted ImageNet pretrained models:
data/train_224_nat_slim
corresponds to a naturally trained modeldata/train_224_robust_eps_0.005_lp_inf_slim
corresponds to an linf adv trained model with eps 0.005data/train_224_robust_eps_1.0_lp_2_slim
corresponds to an l2 adv trained model with eps 1.0
You will need to set the model ckpt directory in the various scripts/ipynb files where appropriate if you want to complete any nontrivial tasks.