This is the official implementation of our paper Semi-supervised Robust Training with Generalized Perturbed Neighborhood, accepted by Pattern Recognition. This project is developed based on Python 3.6, created by Yiming Li and Yan Feng.

Pixel-wise experiments

All code for this part are included in "pixel_exps" subfolder. Please change to that folder before running the code.

Install prerequisites

pip install -r requirements.txt

Running demos

SRT training

• Train WideResNet-34-10 model on CIFAR-10 dataset

bash cifar_train.sh

• Train SmallCNN model on MNIST dataset

bash mnist_train.sh

Robustness evaluation

• Evaluate robust WideResNet-34-10 model on CIFAR-10 by PGD attack

python cifar_eval.py 

• Evaluate robust SmallCNN model on MNIST by PGD attack

python mnist_eval.py 

Download pre-trained model

Download the folder "checkpoints" [download link] and put it within the "pixel_exps" folder, then you can conduct the robustness evaluation without training.

Spatial experiments

Install prerequisites

pip install -r requirements.txt

Running demos

SRT training

• Train ResNet model on CIFAR-10 dataset

bash CIFAR_train.sh

• Train SmallCNN model on MNIST dataset

bash MNIST_train.sh

Robustness evaluation

• Evaluate robust ResNet model on CIFAR-10 by GridAdv attack

bash CIFAR_eval.sh

• Evaluate robust SmallCNN model on MNIST by GridAdv attack

bash MNIST_eval.sh

Download pre-trained model

Download the "checkpoints" [download link] and put it within the "spatial_exps" folder, then you can conduct the robustness evaluation without training.

Reference

If our work or this repo is useful for your research, please cite our paper as follows:

@article{li2022semi,
  title={Semi-supervised robust training with generalized perturbed neighborhood},
  author={Li, Yiming and Wu, Baoyuan and Feng, Yan and Fan, Yanbo and Jiang, Yong and Li, Zhifeng and Xia, Shu-Tao},
  journal={Pattern Recognition},
  volume={124},
  pages={108472},
  year={2022},
  publisher={Elsevier}
}