Official Code

Code for "Learning defense transformations for counterattacking adversarial examples"

Dependencies

Python 3.7.5

Pytorch 1.5.0

cifar2png 0.0.4

torchattacks 1.3

Datasets

We conduct experiments on CIFAR-10 and CIFAR-100 datasets.

For convenience, we convert CIFAR-10 and CIFAR-100 datasets into PNG images by [cifar2png].

Train the defense transformer

• Step1: assign the path of cifar10 and convert CIFAR-10 datasets into PNG images.

  • run cifar2png cifar10 datasets/cifar10

• Step2: train classifier h in Algorithm 1.

  • run python train_classifier.py --arch=resnet56

• Step3: generate a set of adversarial examples A in Algorithm 1.

  • run python data_preprocess.py

• Step4: train the defense transformer T in Algorithm 1.

  • run python main.py

Evaluation

• Step1: test the defense transformer over ResNet-56 on CIFAR-10.

  • run python evaluation.py --model_h_path 'pytorch_resnet_cifar10/best_model.th' --model_ST_path 'checkpoint_defense_transformer/resnet56_cifar10.pth'

• Step2: test the defense transformer over other models on other datasets.

  • modify the model and dataset, train the model follow 4 steps above, and run python evaluation.py <model_h_path> <model_ST_path> <dataroot>