PyTorch implementation of SACNet for adversarial attacks and defenses in hyperspectral image classification.

• Install Pytorch 1.4.0 from https://github.com/pytorch/pytorch with Python 3.7.

• Clone this repo.

git clone https://github.com/YonghaoXu/SACNet

• Download the Pavia University image and the corresponding annotations. Put these files into the Data folder.

• Data Preparation:

  • python GenSample.py --train_samples 300

  The default training set is generated by randomly selecting 300 samples from each category.

  You can change parameter --train_samples to check the performance in other training scenarios.

• Performance on the Clean Test Set:

  To conduct HSI classification on the original clean test set without adversarial attack, simply run:

  • CUDA_VISIBLE_DEVICES=0 python Test_Clean.py --model SACNet

  You can also change parameter --model to check the performance with other models (e.g., --model SSFCN)

• Adversarial Attack with the FGSM:

  • CUDA_VISIBLE_DEVICES=0 python Attack_FGSM.py --model SACNet

• Adversarial Examples Visualization:

  • CUDA_VISIBLE_DEVICES=0 python GenAdvExample.py --model SACNet

Self-Attention Context Network: Addressing the Threat of Adversarial Attacks for Hyperspectral Image Classification

Please cite our paper if you find it useful for your research.

@article{sacnet,
  title={Self-Attention Context Network: Addressing the Threat of Adversarial Attacks for Hyperspectral Image Classification}, 
  author={Xu, Yonghao and Du, Bo and Zhang, Liangpei},
  journal={IEEE Trans. Image Process.}, 
  volume={30},
  pages={8671--8685},
  year={2021},
  doi={10.1109/TIP.2021.3118977}}
}

This code is partly borrowed from PyTorch-Encoding