UAP_TOOL

This is a universal adversarial perturbation toolbox, can be applied research on the robustness of neural networks in image classification.

Use this toolbox to make adversarial examples in seconds.

Quick start：

First：

git clone https://github.com/525309178/UAP_BOX.git

Second:

cd Classfier, run this train_cifar10.py.

eg: python train_cifar10.py

(Note: When the first run, it will download cifar10-dataset to the folder "RawDatasets". After training, the checkpoint of model will be saved to the "CIFAR10" folder of the current folder).

Third:

cd Attacks, run this advres.py to train a generator model, it can generate universal adversarial perturbation and make adversarial sample.

 eg: python advres.py --cuda .

(Note: Before run this *.py of the folder, you must have a trained model used for image classification. After training the generator model will be saved to the param "outf" seted path. You can set the "save_adv" params run advresTest.py to save adversarial sample,eg: python advresTest.py --cuda --save_adv=1 )

Such as:

Fourth:

cd Evalutions, test the robustness of image classification models on raw dataset or adversarial dataset.

  eg. python testadv.py