AdvWGAN

AdvWGAN is designed to generate adversarial example by using a modified WassersteinGAN and can be use with black-box classifier.

Use

Pretrained Classifiers and Models

All files are available here

Classifier

You first need a classifier, here we use the following class to define one.

class Classifier:

def  __init__(self, path='saved_models/FC_DFC'):
    self.path = path
    
def load_model(self):
    json_file = open(self.path+'.json', 'r')
    loaded_model_json = json_file.read()
    json_file.close()
    self.model = model_from_json(loaded_model_json)
    
def load_weights(self):
    self.model.load_weights(self.path+".h5")

def predict(self, X):
    X = np.reshape(X, (-1, 48))
    return self.model.predict_proba([X])

Dataset

You must use a dataset object like

  data = dataset.Dataset(X, y)

WGAN

For the wgan you can use your own methods or user some of ours

generator = models.make_g_conv_1d(img_size=48, hiddens_dims=[256,128,64], o=tf.nn.tanh)
discriminator = models.make_d_conv_1d(hiddens_dims=[64,128,256])

Then the easiest way to create a gan is by using a CFG file

advgan = gan.make_gan('cfg/DFC.cfg')

You have to set your generator, discriminator, classifier and dataset

advgan.set_generator(generator)
advgan.set_discriminator(discriminator)
advgan.set_classifier(classifier)
advgan.set_dataset(data)

Finally, you can build your model

advgan.build_model(reset_graph=True)

Training

Training is pretty straightforward, the only thing tricky is the print function you want, you can use one of ours or code your own

advgan.train(print_method= print_functions.plot_samples_mean_std())

Generate

Once training is done, you can use the generator like this

g = advgan.load_and_generate('AdvGAN_DFC_ct', batch_size=512)

Some Results

MNIST

With this classifier

You can generate these images, in red you can see the prediction

DFC2018

With this classifier

You can generate these spectrum (class 1)