Awesome

Adversarial Attacks

Install

Environment

Please create a new conda environment and run:

pip install requirements.txt

Data

CIFAR10 will be downloaded automatically
For PACS dataset, please refer to ./data/PACS.py for install
For NIPS17 dataset, you can run \

kaggle datasets download -d google-brain/nips-2017-adversarial-learning-development-set

and then put it into ./resources/NIPS17

Model Checkpoints

All the models in our paper are from torchvision, robustbench, timm library, and the checkpoints will be downloaded automatically.

We also encapsulate some models and defenses in "./models" and "./defenses". If you want to attack them, you can download their checkpoints by yourself

Usage

Code Framework

attacks: Some attack algorithms. Including VMI, VMI-CW, CW, SAM, etc.
data: loader of CIFAR, NIPS17, PACS
defenses: Some defenses algorithm
experiments: Example codes
models: Some pretrained models
optimizer: scheduler and optimizer
tester: some functions to test accuracy and attack success rate
utils: Utilities. Like draw landscape, get time, HRNet, etc.

Basic functions

tester.test_transfer_attack_acc(attacker:AdversarialInputBase, loader:DataLoader, target_models: List[nn.Module]) \

This function aims to get the attack success rate on loader against target models

attacker = xxxAttacker(train_models: List[nn.Module])

You can initialize attacker like this.

Examples

Here is an example of testing attack success rate on NIPS17 loader.

from models import resnet18, Wong2020Fast, Engstrom2019Robustness, BaseNormModel, Identity
from attacks import MI_CommonWeakness
attacker = MI_CommonWeakness([
    BaseNormModel(resnet18(pretrained=True)), # model that requires normalization
    Identity(Wong2020Fast(pretrained=True)) # model that do not need normalization
])

from tester import test_transfer_attack_acc
from data import get_NIPS17_loader
test_transfer_attack_acc(attacker, 
                         get_NIPS17_loader(), 
                         [
                             Identity(Wong2020Fast(pretrained=True)), # white box attack
                             Identity(Engstrom2019Robustness(pretrained=True)), # transfer attack
                          ]
                         )

For more example codes, please visit './experiments' folder. There are some example codes using our framework to attack and draw landscapes. I believe you can quickly get familiar with our framework via these example codes.

HRNet

HRNet is a function that aims to reduce memory cost when crafting adversarial examples.

We haven't implemented the convolution of HRNet. Up to now, HRNet can only help to reduce about 30% of memory cost

Usage

from models import resnet18
from utils import change

model = resnet18()
model = change(model)

Citation

Please cite us:

@article{chen2023rethinking,
  title={Rethinking Model Ensemble in Transfer-based Adversarial Attacks},
  author={Chen, Huanran and Zhang, Yichi and Dong, Yinpeng and Zhu, Jun},
  journal={arXiv preprint arXiv:2303.09105},
  year={2023}
}

If you have any question, you can contact us by:
Email: huanran_chen@outlook.com, huanranchen@bit.edu.cn