pgn's Introduction

PGN

This repository contains the PyTorch code for the paper:

Boosting Adversarial Transferability by Achieving Flat Local Maxima (NeurIPS 2023)

Zhijin Ge, Hongying Liu, Xiaosen Wang, Fanhua Shang, Yuanyuan Liu.

We also include the code in the framework TransferAttack.

Requirements

Python == 3.7.11
pytorch == 1.8.0
torchvision == 0.8.0
numpy == 1.21.2
pandas == 1.3.5
opencv-python == 4.5.4.60
scipy == 1.7.3
pillow == 8.4.0
pretrainedmodels == 0.7.4
tqdm == 4.62.3
imageio == 2.6.1

Qucik Start

Prepare the dataset and models.

You can download the ImageNet-compatible dataset from here and put the data in './dataset/'.
The normally trained models (i.e., Inc-v3, Inc-v4, IncRes-v2, Res-50, Res-101, Res-100) are from "pretrainedmodels", if you use it for the first time, it will download the weight of the model automatically, just wait for it to finish.
The adversarially trained models (i.e, ens3_adv_inc_v3, ens4_adv_inc_v3, ens_adv_inc_res_v2) are from SSA or tf_to_torch_model. For more detailed information on how to use them, visit these two repositories.

Runing attack

You can run our proposed attack as follows.

python Incv3_PGN_Attack.py

The generated adversarial examples would be stored in the directory ./incv3_xx_xx_outputs. Then run the file verify.py to evaluate the attack success rate of each model used in the paper:

python verify.py

You can run the file 'surface_map.py' to visualize the loss surface maps for the adversarial examples, the maps will be stored in the directory './loss_surfaces/'.

python surface_map.py

Citation

If our paper or this code is useful for your research, please cite our paper.

@inproceedings{ge2023boosting,
     title={{Boosting Adversarial Transferability by Achieving Flat Local Maxima}},
     author={Zhijin Ge and Hongying Liu and Xiaosen Wang and Fanhua Shang and Yuanyuan Liu},
     booktitle={Proceedings of the Advances in Neural Information Processing Systems},
     year={2023},
}

Recommend Projects