LinBP-Attack

Code for our NeurIPS 2020 paper Backpropagating Linearly Improves Transferability of Adversarial Examples.

Environments

• Python 3.7.4
• PyTorch 1.4.0
• Torchvision 0.5.0
• Pillow 7.0.0
• Numpy 1.18.1

Datasets

CIFAR-10 and ImageNet should be prepared into the following structure:

linbp-attack
└───data
    ├── cifar10
    │   ├── selected_cifar10.csv
    │   └── cifar-10-batches-py
    └── imagenet
        ├── selected_imagenet.csv
        └── ILSVRC2012_img_val

For CIFAR-10, select images from validation set, and write selected_cifar10.csv as following:

class_index,data_index
3,0
8,1
8,2
...

The data_index is the index of image in data/cifar10/cifar-10-batches-py/test_batch

For ImageNet, select images from validation set, and write selected_imagenet.csv as following:

class_index, class, image_name
0,n01440764,ILSVRC2012_val_00002138.JPEG
2,n01484850,ILSVRC2012_val_00004329.JPEG
...

Models

Please download pretrained models at here, then extract them to attack/cifar10/models and attack/imagenet/models.

Attack

Untarget attack on CIFAR-10 using PGD+LinBP+ILA, under epsilon = 0.03 :

python3 attack/cifar10/attack_vgg19.py --epsilon 0.03 --niters 100 --ila_niters 100 --method linbp_ila_pgd \
--save_dir data/cifar10/linbp_ila_pgd --batch_size 500

Untarget attack on ImageNet using PGD+LinBP+ILA, under epsilon = 0.03 :

python3 attack/imagenet/attack_resnet50.py --epsilon 0.03 --niters 300 --ila_niters 100 --method linbp_ila_pgd \
--save_dir data/imagenet/linbp_ila_pgd --batch_size 50 --sgm_lambda 1.0

--sgm_lambda is the scaling factor in the residual units of SGM method.
For both CIFAR-10 and ImageNet, add --target_attack to mount a target attack. The --method can be set as fgsm/ifgsm/pgd/mifgsm/mdi2fgsm. Either ILA, LinBP or both are supported by adding ila/linbp/linbp_ila like linbp_pgd.

Test

CIFAR-10:

python3 attack/cifar10/test.py --dir data/cifar10/linbp_ila_pgd

ImageNet:

python3 attack/imagenet/test.py --dir data/imagenet/linbp_ila_pgd

Acknowledgements

The following resources are very helpful for our work:
Pretrained models for CIFAR-10: bearpaw/pytorch-classification
Pretrained models for ImageNet: Cadene/pretrained-models.pytorch and torchvision.models
GDAS: D-X-Y/AutoDL-Projects
Pretrained GDAS: ZiangYan/subspace-attack.pytorch

Citation

@inproceedings{guo2020backpropagating,
    title={Backpropagating Linearly Improves Transferability of Adversarial Examples.},
    author={Guo, Yiwen and Li, Qizhang and Chen, Hao},
    booktitle={NeurIPS},
    year={2020}
}