On Adversarial Robustness: A Neural Architecture Search perspective

Preparation:

Clone the repository:

https://github.com/tdchaitanya/nas-robustness.git

prerequisites

• Python 3.6
• Pytorch 1.2.0
• CUDA 10.1

For a hassle-free environment setup, use the environment.yml file included in the repository.

Pre-trained models:

For easy reproduction of the result shown in the paper, this repository is organized dataset-wise, and all the pre-trained models can be downloaded from here

CIFAR-10/100

All the commands in this section should be executed in the cifar directory.

Hand-crafted models on CIFAR-10

All the files corresponding to this dataset are included in cifar-10/100 directories. Download cifar weigths from the shared drive link and place them in nas-robustness/cifar-10/cifar10_models/state_dicts directory.

For running all the four attacks on Resnet-50 (shown in Table 1) run the following command.

python handcrafted.py --arch resnet50

Change the architecture parameter to run attacks on other models. Only resnet-18, resnet-50, densenet-121, densenet-169, vgg-16 are supported for now. For other models, you may have to train them from scratch before running these attacks.

Hand-crafted models on CIFAR-100

For training the models on CIFAR-100 we have used fastai library. Download cifar-100 weigths from the shared drive link and place them in nas-robustness/cifar/c100-weights directory.

Additionally, you'll also have to download the CIFAR-100 dataset from here and place it in the data directory (we'll not be using this anywhere, this is just needed to initialize the fastai model).

python handcrafted_c100.py --arch resnet50

DARTS

Download DARTS CIFAR-10/100 weights from the drive and place it nas-robustness/darts/pretrained

For running all the four attacks on DARTS run the following command:

python darts-nas.py

Add --cifar100 to run the experiments on cifar-100

P-DARTS

Download P-DARTS CIFAR-10/100 weights from the drive and place it nas-robustness/pdarts/pretrained

For running all the four attacks on P-DARTS run the following command:

python pdarts-nas.py

Add --cifar100 to run the experiments on CIFAR-100

NSGA-Net

Download NSGA-Net CIFAR-10/100 weights from the drive and place it nas-robustness/nsga_net/pretrained

For running all the four attacks on P-DARTS run the following command:

python nsganet-nas.py

Add --cifar100 to run the experiments on CIFAR-100

PC-DARTS

Download PC-DARTS CIFAR-10/100 weights from the drive and place it nas-robustness/pcdarts/pretrained

For running all the four attacks on PC-DARTS run the following command:

python pcdarts-nas.py

Add --cifar100 to run the experiments on CIFAR-100

ImageNet

All the commands in this section should be executed in ImageNet directory.

Hand-crafted models

All the files corresponding to this dataset are included in imagenet directory. We use the default pre-trained weights provided by PyTorch for all attacks.

For running all the four attacks on Resnet-50 run the following command:

python handcrafted.py --arch resnet50

For DARTS, P-DARTS, PC-DARTS follow the same instructions as mentioned above for CIFAR-10/100, just change the working directory to ImageNet

DenseNAS

Download DenseNAS ImageNet weights from the drive (these are same as the weights provided in thier official repo) and place it nas-robustness/densenas/pretrained

For running all the four attacks on DenseNAS-R3 run the following command:

python dense-nas.py --model DenseNAS-R3

Citation

@InProceedings{Devaguptapu_2021_ICCV,
    author    = {Devaguptapu, Chaitanya and Agarwal, Devansh and Mittal, Gaurav and Gopalani, Pulkit and Balasubramanian, Vineeth N},
    title     = {On Adversarial Robustness: A Neural Architecture Search Perspective},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
    month     = {October},
    year      = {2021},
    pages     = {152-161}
}

Acknowledgements

Some of the code and weights provided in this library are borrowed from the libraries mentioned below:

• cifar10_models
• torchattacks
• fastai
• darts
• pdarts
• nsga-net
• pcdarts
• densenas