AdvOps: Decoupling Adversarial Examples

The official code for AdvOps: Decoupling Adversarial Examples.

We introduce the implementation of the AdvOps for ImageNet and CIFAR-10 as following.

we use the directory of the ImageNet as follows

--ImageNet

|--class_2

|-- xxx.png

|--class_2

|-- xxx.png

....

We adopt the ImageFolder provide by torch to load dataset.

ImageNet

Train generator

We use the ImageNet pretrained model provided by the PyTorch as the target model. To train the generator, please run the follow command

python main.py --model_name resnet50 --data_dir "data path" --ckpt_path "path to save the model parameters"  --batch_size 50

another parameters can be seen in main.py.

We also provide the trained model of the generator for ImageNet in here with the fetch code 1111, CIFAR10 in here, with the fetch code 1111.

We will introduce the evaluation process in the following part.

CIFAR-10

As pytorch do not provide the CIFAR-10 pretraining model, therefore, we have to train the model on CIFAR-10 first, then to train the generator.

Step 1: Train CIFAR-10 model

Using the following common to train all CIFAR-10 model.

python train_cifar10_model.py --ckpt_path "path to save the model checkpoint"

We also provide our pretrained CIFAR10 model in here with fetch code 1111.

Step 2: Train generator for each model

Using the following common to train the generator for each model

python main_cifar10.py --model_name resnet50 --data_dir "data path" --model_ckpt "path of trained CIFAR-10 model" --ckpt_path "path to save the model parameters"  --batch_size 200

Evaluation

Evaluate the AdvOps

To evaluate the trained generator, using the following command for ImageNet and CIFAR10, respectively.

For ImageNet

python baseline_methods_transfer_gan.py

For CIFAR-10

python baseline_methods_transfer_gan_cifar10.py

Evaluate the comparison methods

To evaluate the comparison method, e.g., FGSM, using the following command

For ImageNet

python baseline_methods_transfer.py

For CIFAR-10

python baseline_methods_transfer_cifar10.py

Note that please the model checkpoint and dataset in the corresponding directory before runing the code.

winterwindwang / AdvOps