def __init__(self):
super().__init__()
self.dataset = FashionMnistProvider()
self.cnn = CNNModel(
image_size=self.dataset.IMAGE_SIZE,
classes=self.dataset.CLASSES,
channels=self.dataset.CHANNELS,
model_name=self.MODEL_NAME,
model_dir='tmp/model-{}'.format(self.MODEL_NAME),
conv_layers=[32, 64],
fc_layers=[1024],
)
def __init__(self):
super().__init__()
self.dataset = Cifar10Provider()
self.cnn = CNNModel(
image_size=self.dataset.IMAGE_SIZE,
classes=self.dataset.CLASSES,
channels=self.dataset.CHANNELS,
kernel_size=[3, 3],
model_name=self.MODEL_NAME,
model_dir='tmp/model-{}'.format(self.MODEL_NAME),
conv_layers=[48, 96, 192],
fc_layers=[512, 256],
)
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from cnn_gtsrb.cnn.model import CNNModel
data = np.load('tmp/adv_mnist-28x28/fast-jsma_mnist-0.npy')
cnn = CNNModel(
image_size=28,
classes=10,
channels=1,
model_name='mnist-28x28',
model_dir='tmp/model-mnist-28x28',
conv_layers=[32, 64],
fc_layers=[1024],
)
x, _ = cnn.make_inputs()
probs = cnn.make_model(x)
cnn.start_session()
cnn.init_session_and_restore()
orig_correct_count = 0
success_count = 0
actual_success_count = 0
actual_label_count = 0
sample = data[0]
from cnn_gtsrb.attacks.crafting import BatchJSMACrafting
DIR_NAME = os.path.join('tmp', 'batch_adv')
if not os.path.isdir(DIR_NAME):
os.makedirs(DIR_NAME)
logging.basicConfig(level=logging.INFO)
tf.logging.set_verbosity(tf.logging.INFO)
mnist = MnistProvider()
# Make CNN Model
cnn = CNNModel(
image_size=mnist.IMAGE_SIZE,
classes=mnist.CLASSES,
model_name='mnist-28x28.1',
model_dir='tmp/mnist_model-28x28.1',
conv_layers=[32, 64],
fc_layer=1024,
)
jsma_params = {
'theta': 1.,
'gamma': 0.1,
'clip_min': 0.,
'clip_max': 1.,
'y_target': None
}
test_data = mnist.raw_test_data()
test_data = test_data[:5000]
# Get data for each class in MNIST
data = gtsrb.raw_train_data()
np.random.shuffle(data)
grouped_data = []
for i in range(0, gtsrb.CLASSES, 2):
grouped_data.append(data[(data[:, -1] == i), :])
images = np.concatenate(
[np.expand_dims(group[0], 0) for group in grouped_data], axis=0)
# Make CNN Model
cnn = CNNModel(
image_size=gtsrb.IMAGE_SIZE,
classes=gtsrb.CLASSES,
model_name='gtsrb-32x32',
model_dir='tmp/gtsrb_model-32x32',
conv_layers=[32, 64, 128],
fc_layer=512,
)
x, y = cnn.make_inputs()
probs = cnn.make_model(x)
cnn.start_session()
cnn.init_session_and_restore()
fgsm = FastGradientMethod(cnn, sess=cnn.sess)
success_matrix = []
for eps in np.concatenate((np.arange(0.02, 0.1, 0.02), np.arange(0.1, 1,
from cleverhans.attacks import FastGradientMethod
from cnn_gtsrb.dataset.gtsrb import GtsrbProvider
from cnn_gtsrb.cnn.model import CNNModel
#logging.basicConfig(level=logging.INFO)
tf.logging.set_verbosity(tf.logging.INFO)
fgsm_params = {'eps': 0.5, 'clip_min': 0., 'clip_max': 1.}
gtsrb = GtsrbProvider()
# gtsrb.dump_images()
cnn = CNNModel(
image_size=gtsrb.IMAGE_SIZE,
classes=gtsrb.CLASSES,
model_name='gtsrb-64x64',
model_dir='tmp/gtsrb_model-64x64',
conv_layers=[32, 64, 128],
fc_layer=512,
)
x, y = cnn.make_inputs()
probs = cnn.make_model(x)
cnn.start_session()
fgsm = FastGradientMethod(cnn, sess=cnn.sess)
adv_x = fgsm.generate(x, **fgsm_params)
probs = cnn.make_model(adv_x)
cnn.adv_test(probs, x, y, adv_x, gtsrb.test_data(size=1000))
# cnn.test(gtsrb)
cnn.end_session()
from cleverhans.attacks import FastGradientMethod
from cnn_gtsrb.dataset.mnist import MnistProvider
from cnn_gtsrb.cnn.model import CNNModel
#logging.basicConfig(level=logging.INFO)
tf.logging.set_verbosity(tf.logging.INFO)
fgsm_params = {'eps': 0.2, 'clip_min': 0., 'clip_max': 1.}
mnist = MnistProvider()
# mnist.dump_images()
cnn = CNNModel(
image_size=mnist.IMAGE_SIZE,
classes=mnist.CLASSES,
model_name='mnist-28x28.1',
model_dir='tmp/mnist_model-28x28.1',
conv_layers=[32, 64],
fc_layer=1024,
)
x, y = cnn.make_inputs()
probs = cnn.make_model(x)
cnn.start_session()
# Not necessary because adv_test will restore the model
# cnn.init_session_and_restore()
fgsm = FastGradientMethod(cnn, sess=cnn.sess)
adv_x = fgsm.generate(x, **fgsm_params)
adv_probs = cnn.make_model(adv_x)
import numpy as np
import tensorflow as tf
from cnn_gtsrb.cnn.model import CNNModel
tf.logging.set_verbosity(tf.logging.INFO)
results = []
for f in glob.glob('tmp/adv_cgtsrb10-32x32/fast-jsma_*.npy'):
results.append(np.load(f))
data = np.concatenate(results, axis=0)
cnn = CNNModel(
image_size=32,
classes=10,
channels=3,
model_name='cgtsrb10-32x32',
model_dir='tmp/model-cgtsrb10-32x32',
conv_layers=[32, 64, 128],
fc_layers=[512],
)
# cnn = CNNModel(
# image_size=32,
# classes=10,
# channels=3,
# model_name='cifar10-32x32',
# model_dir='tmp/model-cifar10-32x32',
# kernel_size=[3, 3],
# conv_layers=[48, 96, 192],
# fc_layers=[512, 256],
# )
import logging
import tensorflow as tf
from cnn_gtsrb.dataset.mnist import MnistProvider
from cnn_gtsrb.cnn.model import CNNModel
#logging.basicConfig(level=logging.INFO)
tf.logging.set_verbosity(tf.logging.INFO)
mnist = MnistProvider()
# mnist.dump_images()
cnn = CNNModel(
image_size=mnist.IMAGE_SIZE,
classes=mnist.CLASSES,
model_name='mnist-28x28.1',
model_dir='tmp/mnist_model-28x28.1',
conv_layers=[32, 64],
fc_layer=1024,
)
x, y = cnn.make_inputs()
probs = cnn.make_model(x)
cnn.start_session()
cnn.train(probs, x, y, 17000, mnist)
# cnn.test(gtsrb)
cnn.end_session()