Home

Awesome

Feature Importance-aware Attack(FIA)

This repository contains the code for the paper:

Feature Importance-aware Transferable Adversarial Attacks (ICCV 2021)

Requirements

Experiments

Introduction

Example Usage

Generate adversarial examples:
python attack.py --model_name vgg_16 --attack_method FIA --layer_name vgg_16/conv3/conv3_3/Relu --ens 30 --probb 0.7 --output_dir ./adv/FIA/
python attack.py --model_name vgg_16 --attack_method PIM --amplification_factor 10 --gamma 1 --Pkern_size 3 --output_dir ./adv/PIM/
python attack.py --model_name vgg_16 --attack_method FIAPIDIM --layer_name vgg_16/conv3/conv3_3/Relu --ens 30 --probb 0.7 --amplification_factor 2.5 --gamma 0.5 --Pkern_size 3 --image_size 224 --image_resize 250 --prob 0.7 --output_dir ./adv/FIAPIDIM/

Different attack methods have different parameter setting, and the detailed setting can be found in our paper.

Evaluate the attack success rate
python verify.py --ori_path ./dataset/images/ --adv_path ./adv/FIA/ --output_file ./log.csv

Citing this work

If you find this work is useful in your research, please consider citing:

@inproceedings{wang2021feature,
  title={Feature importance-aware transferable adversarial attacks},
  author={Wang, Zhibo and Guo, Hengchang and Zhang, Zhifei and Liu, Wenxin and Qin, Zhan and Ren, Kui},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={7639--7648},
  year={2021}
  }