AGAT: Attribute-Guided Adversarial Training for Robustness to Natural Perturbations

This repository contains code for the AAAI 2021 paper of the same name. A preprint is available here: https://arxiv.org/abs/2012.01806 .

The CLEVR-Singles dataset released as part of our publication can be found in a separate repository https://github.com/tejas-gokhale/CLEVR-Singles

Datasets

This repository is currently setup for the CIFAR-10-C benchmark (Hendrycks and Dietterich)

• CIFAR-10-C dataset [download][https://zenodo.org/record/2535967#.Xaf8uedKj-Y]
• CIFAR-10-C codebase (not required) [this][https://github.com/hendrycks/robustness]

We have adapted parts of the code for our CIFAR-10-C experiments from [TTT (Sun et al. ICML 2020)][https://arxiv.org/abs/1909.13231] whose code is [here][https://github.com/yueatsprograms/ttt_cifar_release]

Pre-Requisites

pip install the following:

pytorch (works with 1.6.0), numpy, tqdm, skimage , colorama

CIFAR-10 Experiments:

Inference

Pretrained models are found in ./results/. Please run: python test_ours.py --shared layer2 --rotation_type expand --group_norm 8 \ --ckpt <ckpt_path>

For instance, to reproduce the "blur+noise" model from the paper (Table), replace <ckpt_path> by ./results/cifar10_agat_blur_noise/ckpt_final.pth

Training

Coming soon!

CLEVR-Singles Experiments:

Coming soon!

Reference

If you find AGAT or the CLEVR-Singles dataset useful, please use the following citation:

@article{gokhale2020attribute,
  title={Attribute-Guided Adversarial Training for Robustness to Natural Perturbations},
  author={Gokhale, Tejas and Anirudh, Rushil and Kailkhura, Bhavya and Thiagarajan, Jayaraman J and Baral, Chitta and Yang, Yezhou},
  journal={arXiv preprint arXiv:2012.01806},
  year={2020}
}