Adversarial Robustness
This repository contains the code needed to evaluate models trained in Doing More with Less: Improving Robustness using Generated Data which has been accepted at ICLR 2021 Security and Safety in Machine Learning Systems Workshop.
Contents
We have released our top-performing models in two formats compatible with JAX and PyTorch. This repository also contains our model definitions.
Running the example code
Downloading a model
Download a model from links listed in the following table. Clean and robust accuracies are measured on the full test set. The robust accuracy is measured using AutoAttack.
| dataset | norm | radius | architecture | extra data | clean | robust | link |
|---|---|---|---|---|---|---|---|
| CIFAR-10 | ℓ∞ | 8 / 255 | WRN-70-16 | ✗ | 86.94% | 63.62% | jax, pt |
| CIFAR-10 | ℓ∞ | 8 / 255 | WRN-28-10 | ✗ | 85.97% | 60.73% | jax, pt |
| CIFAR-10 | ℓ2 | 8 / 255 | WRN-70-16 | ✗ | 90.83% | 78.39% | jax, pt |
| CIFAR-10 | ℓ2 | 8 / 255 | WRN-28-10 | ✗ | 90.24% | 77.44% | jax, pt |
| CIFAR-100 | ℓ∞ | 8 / 255 | WRN-70-16 | ✗ | 60.46% | 33.49% | jax, pt |
| CIFAR-100 | ℓ∞ | 8 / 255 | WRN-28-10 | ✗ | 59.18% | 30.81% | jax, pt |
Using the model
Once downloaded, a model can be evaluated (clean accuracy) by running the
eval.py script in either the jax or pytorch folders. E.g.:
cd jax
python3 eval.py \
--ckpt=${PATH_TO_CHECKPOINT} --depth=70 --width=16 --dataset=cifar10
Generated datasets
This work uses samples generated by a Denoising Diffusion Probabilistic Model (DDPM; Ho et al., 2020) to improve robustness. The DDPM is solely trained on the original training data and does not use additional external data. The following table links to datasets of 1M generated samples for CIFAR-10, CIFAR-100 and SVHN.
| dataset | model | size | link |
|---|---|---|---|
| CIFAR-10 | DDPM | 1M | npz |
| CIFAR-100 | DDPM | 1M | npz |
| SVHN | DDPM | 1M | npz |
To load each dataset, use NumPy. E.g.:
npzfile = np.load('cifar10_ddpm.npz')
images = npzfile['image']
labels = npzfile['label']
Citing this work
If you use this code or these models in your work, please cite the complete version which combines data augmentation with generated samples:
@article{rebuffi2021fixing,
title={Fixing Data Augmentation to Improve Adversarial Robustness},
author={Rebuffi, Sylvestre-Alvise and Gowal, Sven and Calian, Dan A. and Stimberg, Florian and Wiles, Olivia and Mann, Timothy},
journal={arXiv preprint arXiv:2103.01946},
year={2021},
url={https://arxiv.org/pdf/2103.01946}
}
Disclaimer
This is not an official Google product.