Official source code of the paper:

G. Cheng, X. Sun, K. Li, L. Guo and J. Han, "Perturbation-Seeking Generative Adversarial Networks: A Defense Framework for Remote Sensing Image Scene Classification," in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-11, 2022, Art no. 5605111, doi: 10.1109/TGRS.2021.3081421.

@ARTICLE{9442932,
  author={Cheng, Gong and Sun, Xuxiang and Li, Ke and Guo, Lei and Han, Junwei},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Perturbation-Seeking Generative Adversarial Networks: A Defense Framework for Remote Sensing Image Scene Classification}, 
  year={2021},
  volume={60},
  number={},
  pages={1-11},
  doi={10.1109/TGRS.2021.3081421}}

• CUDA: CUDA 9.0
• CUDNN: CUDNN 7.6.5
• Ubuntu: Ubuntu 16.04.7
• gcc/g++: 5.4.0

git clone https://github.com/xuxiangsun/PSGAN.git
cd PSGAN

To install the python environment, please run:

conda env create -f ./condalist.yml

Tips: If you can not install the whole packages in condalist.yml, you can check which package can not be download and shield it firstly. Then, running

conda env create -f ./condalist.yml

to install other packages. Next, you can download and install these shielded packages manually.

The datasets are all saved in /path of PSGAN/datasets, so we should make a new directory firstly:

mkdir ./datasets

In our experiments, we use three benchmarks includes the UC Merced Land-Use Dataset, the AID Dataset, and the NWPU-RESISC45 Dataset. After downloading these datasets, please unfold them to the right directory. It is noticeable that for any dataset, we should organize the directory tree of the dataset folder into the following structure:

.
├── class name1
├── class name2
├── ...
└── class namek

Besides the dataset folder's organization we mentioned above, there are some extra settings should be done.

1. Firstly, we also should write the classes dictionary file : _dict.txt (e.g., NWPURESISC45_dict.txt) to save the mapping from the class name to class ID. You can generate this file by running (taking the NWPU-RESISC45 Dataset as an example):

   python read_classes.py --dataroot ./datasets/NWPURESISC45
   

2. Secondly, we also should customize the dataset file in ./data folder for our own dataset. Pay attention that the dataset's name in _dict.txt, the name of the dataset folder in ./datasets, and the name of the dataset file in ./data should be the same.

It is noticeable that the pre-trained models on ImageNet are leveraged by our code when we train the target model. Thus, we should download the pre-trained models on ImageNet to ./pretrained directory. Firstly, we should make a new directory by running the following command:

mkdir ./pretrained

And then, downloading the pre-trained models needed by your own experiment and saving them at ./pretrained.

To train a target model, please run:

python train_target_model.py --dataroot ./datasets/NWPURESISC45 --dataset_mode none --name 45_scnet50_0.1 --train_ratio 0.1 --target_model scnet50 --flag train --lr 0.0002 --model scnet50 --niter 50 --niter_decay 50 --batch_size 36

Then, the trained weight parameters and training logs will be saved at ./checkpoints/45_scnet50_0.2. The tensorboard logs will be saved at ./results/45_scnet50_0.2.

After training the target model, you can test its performance by running:

python test_target_model.py --dataroot ./datasets/NWPURESISC45 --model scnet50 --target_model scnet50 --dataset_mode none --name 45_scnet50_0.1 --testname original --train_ratio 0.1

The testing log will be saved at ./results/45_scnet50_0.1/test_log.txt

To generate adversarial examples, for example, please run:

python attack.py --dataroot ./datasets/NWPURESISC45 --train_ratio 0.1 --model scnet50 --flag test --dataset_mode none --attack fgsm --name 45_scnet50_0.1

Then the adversarial examples will be save at /path of PSGAN/datasets/NWPURESISC45/adv_resnet152_fgsm/adv, and their corresponding clean examples will be saved at /path of PSGAN/datasets/NWPURESISC45/adv_resnet152_fgsm/orig in the same order.

After generating adversarial examples, we can check the attack ability by running:

python test_target_model.py --dataroot ./datasets/NWPURESISC45/adv_scnet50_fgsm/adv/NWPURESISC45 --model scnet50 --target_model scnet50 --dataset_mode none --name 45_scnet50_0.1 --testname fgsm_adv --train_ratio 0.1

Also, the test results will be saved at ./results/45_scnet50_0.1/test_log.txt.

To train a target model via our PSGAN, please run:

python adversarial_defense.py --dataroot ./datasets/NWPURESISC45 --dataset_mode custom --name 45_pgd_scnet50_0.1_psgan --batch_size 9 --attack pgd --train_ratio 0.1 --target_model scnet50 --flag train --lr 0.0001 --model psgan --niter 50 --niter_decay 30

Then the weight parameters and training logs will be saved at ./checkpoints/45_pgd_scnet50_0.1_psgan. After that, we can check the performance of the trained target model by running:

python test_defense.py --dataroot ./datasets/NWPURESISC45 --model psgan --dataset_mode custom --name 45_pgd_scnet50_0.2_psgan --attack fgsm --testname fgsm_pgd --train_ratio 0.1 --target_model scnet50

1. The coding style of this repository mainly based on pytorch-CycleGAN-and-pix2pix.
2. For adversarial attacks, we use the Torchattacks, a PyTorch library that contains adversarial attacks to generate adversarial examples.