• This repository includes the following:
  • Edge-guided Adversarial Training (EAT)
  • GAN-based Shape Defense (GSD)
  • Shape-based fast adversarial training
  • Shape-based free adversarial training
  • Robustness evaluation against natural image corruptions

It shows that incorporating edges to images followed by adversarial training and edge redetection at inference time improves robustness drastically.

• 12/19/2019 - TBD

1. Install PyTorch.
2. Install the required python packages. All packages can be installed by running the following command:

pip install -r requirements.txt

3. For each of the code repositories used, please see the accompanying readme files and original code bases.

To train a robust model run the following command:

python train.py --net_type [NET_TYPE] --model [MODEL_TYPE] --sigmas [LIST_OF_SIGMAS]  --data_dir [DATA_FOLDER]  --classes 10 --epochs 10 --inp_size 28 --load_model [LOAD_MODEL_IF_EXISTS]

This trains a robust model with the following parameters. [NET_TYPE] can be edge, rgb or rgbedge. [MODEL_TYPE] determines the type of the model and [DATA_FOLDER] sets the folder containing data. Please see model.py for some examples. [LIST_OF_SIGMAS] is the list of perturbation budgets. It trains one several models for each sigma. [LOAD_MODEL_IF_EXISTS] loads a trained classifier to build a robust model for it. See sample_pgd40_mnist_results folder for some sample generated results. You also need to specify the type of the edge detector that you want to use in the config.py. For some other codes, I have placed a sample line how to call it at the begining of the .py file. Also please see run.sh for running the code in scale.

• Run BPDA.py if you want to conduct a substitute attack! (bad naming here perhaps!)
• Run test_noise_robustness.py for testing a model against natural perturbations.
• Run edge_or_gray_only_attack.py if you want to only attack only edge pixels or gray pixels (i.e., masking the background); implemented only for MNIST
• Run plots.ipynb for plotting results
• See also visualization for runing edge detection on clean and attacked images

If you use this code in your research, please cite this project.

@article{shapeDefense2020,
  title={Shape Defense},
  author={Borji, Al},
  journal={xx},
  year={2020}
}