Erosion Attack: Harnessing Corruption To Improve Adversarial Examples

The code is repository for "Erosion Attack: Harnessing Corruption To Improve Adversarial Examples" (IEEE TIP).

python 3.6
tensorflow 1.14

Train a single erosion network for using in single-model attacks or multiple erosion networks for using in ensemble attacks.

(1) train_aux_i3.py for InceptionV3.

(2) train_aux_i4.py for InceptionV4.

(3) train_aux_ir2.py for InceptionResNetV2.

(4) train_aux_r50.py for ResNet50.

Pre-trained weights of erosion networks above are available at here.

Run EA to generate adversarial examples: erosion_attack.py.

Standalone Experiment

Combination Experiment

<3>Ensemble Experiment

If you find this project is useful for your research, please consider citing:

@article{huang2023erosion,
  title={Erosion Attack: Harnessing Corruption To Improve Adversarial Examples},
  author={Huang, Lifeng and Gao, Chengying and Liu, Ning},
  journal={IEEE Transactions on Image Processing},
  year={2023},
  publisher={IEEE}
}