Transferable Targeted Adversarial Attack(TTAA)

This repository contains the code for the paper:

Towards Transferable Targeted Adversarial Examples (CVPR 2023)

Requirements

• Python 3.7

• torch 1.12.1

• torchvision 0.13.1

• numpy 1.21.6

Experiments

Introduction

• attack.py : the code for training the generator and discriminator on different models.

• generator.py : The Network Architecture of generator.

• label_discriminator.py : The Network Architecture of label discriminator.

• feature_discriminator.py : The Network Architecture of feature discriminator.

Example Usage

python train.py --src_dir dataset/source --match_dir dataset/target --feature_layer 5 --model_type Resnet18 --batch_size 64 --save_dir ./saved_model

Citing this work

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

@inproceedings{wang2023towards,
  title={Towards transferable targeted adversarial examples},
  author={Wang, Zhibo and Yang, Hongshan and Feng, Yunhe and Sun, Peng and Guo, Hengchang and Zhang, Zhifei and Ren, Kui},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={20534--20543},
  year={2023}
}