def substitute_model(img_rows=28, img_cols=28, nb_classes=10):
"""
Defines the model architecture to be used by the substitute. Use
the example model interface.
:param img_rows: number of rows in input
:param img_cols: number of columns in input
:param nb_classes: number of classes in output
:return: tensorflow model
"""
input_shape = (None, img_rows, img_cols, 1)
# Define a fully connected model (it's different than the black-box)
layers = [
Conv2D(64, (8, 8), (2, 2), "SAME"),
ReLU(),
Flatten(),
Linear(200),
ReLU(),
Linear(100),
ReLU(),
Linear(nb_classes),
Softmax()
]
return make_basic_cnn()
return MLP(layers, input_shape)
def cifar_net(n_classes):
input_shape = ((None, im_size, im_size, im_chan))
layers = [
Conv2D(128, (5, 5), (1, 1), "SAME"),
ReLU(),
Conv2D(128, (3, 3), (1, 1), "SAME"),
ReLU(),
Conv2D(64, (3, 3), (2, 2), "SAME"),
ReLU(),
Conv2D(64, (3, 3), (1, 1), "SAME"),
ReLU(),
Conv2D(64, (3, 3), (1, 1), "SAME"),
ReLU(),
Conv2D(32, (3, 3), (2, 2), "SAME"),
ReLU(),
Conv2D(12, (3, 3), (1, 1), "SAME"),
ReLU(),
Flatten(),
Linear(256),
ReLU(),
Linear(n_classes)
]
model = MLP(layers, input_shape)
return model
def setup_simple_model_tf(model_pt, input_shape):
from cleverhans_tutorials.tutorial_models import MLP, Linear, ReLU
layers = [Linear(10), ReLU(), Linear(10)]
layers[0].name = 'fc1'
layers[1].name = 'relu'
layers[2].name = 'fc2'
model = MLP(layers, input_shape)
load_weights_pt(model_pt, layers)
return model
def abalone_mlp(nb_classes=2, input_shape=None):
layers = [Flatten(),
Linear(10),
Sigmoid(),
Linear(10),
Sigmoid(),
Flatten(),
Linear(nb_classes),
Softmax()
]
model = MLP(layers, input_shape)
return model
def fashion_net(n_classes):
input_shape = ((None, im_size, im_size, im_chan))
layers = [
Conv2D(64, (8, 8), (2, 2), "SAME"),
ReLU(),
Conv2D(64 * 2, (6, 6), (2, 2), "VALID"),
ReLU(),
Conv2D(64 * 2, (5, 5), (1, 1), "VALID"),
ReLU(),
Flatten(),
Linear(n_classes),
Softmax()
]
model = MLP(layers, input_shape)
return model
def PAP_substitute_model(img_rows=1, img_cols=2, nb_classes=2):
"""
Defines the model architecture to be used by the substitute. Use
the example model interface.
:param img_rows: number of rows in input
:param img_cols: number of columns in input
:param nb_classes: number of classes in output
:return: tensorflow model
"""
input_shape = (None, img_rows, img_cols, 1) #code sous cette forme: vient de Papernot, je garde pour pas tout casser...
# Define a fully connected model (it's different than the black-box)
layers = [Flatten(),
Linear(200),
ReLU(),
Linear(200),
ReLU(),
Linear(nb_classes),
Softmax()]
#layers = [Flatten(), Linear(nb_classes), Softmax()] #surrogate simplifié
return MLP(layers, input_shape)
def substitute_model(img_rows=32, img_cols=32, nb_classes=10):
"""
Defines the model architecture to be used by the substitute. Use
the example model interface.
:param img_rows: number of rows in input
:param img_cols: number of columns in input
:param nb_classes: number of classes in output
:return: tensorflow model
"""
input_shape = (None, img_rows, img_cols, 3)
# Define a fully connected model (it's different than the black-box)
layers = [Flatten(),
SimpleLinear(200),
ReLU(),
SimpleLinear(200),
ReLU(),
SimpleLinear(nb_classes),
SoftmaxT1()]
return MLP(layers, input_shape)
def make_imagenet_cnn(input_shape=(None, 224, 224, 3)):
layers = [
Conv2D(96, (3, 3), (2, 2), "VALID"),
ReLU(),
Conv2D(256, (3, 3), (2, 2), "VALID"),
ReLU(),
Conv2D(384, (3, 3), (2, 2), "VALID"),
ReLU(),
Conv2D(384, (3, 3), (2, 2), "VALID"),
ReLU(),
Conv2D(256, (3, 3), (2, 2), "VALID"),
ReLU(),
Flatten(),
Linear(4096),
ReLU(),
Linear(4096),
ReLU(),
Linear(1000),
Softmax()
]
model = MLP(layers, input_shape)
return model
def substitute_model(phase, img_rows=32, img_cols=32, nb_classes=10):
"""
Defines the model architecture to be used by the substitute. Use
the example model interface.
:param img_rows: number of rows in input
:param img_cols: number of columns in input
:param nb_classes: number of classes in output
:return: tensorflow model
"""
input_shape = (None, img_rows, img_cols, 3)
# Define a fully connected model (it's different than the black-box)
'''
layers = [Flatten(),
SimpleLinear(200),
ReLU(),
SimpleLinear(200),
ReLU(),
SimpleLinear(nb_classes),
SoftmaxT1()]
'''
nb_filters = 64
layers = [
Conv2D(False, nb_filters, (8, 8), (2, 2), "SAME", phase, 'conv1'),
ReLU(),
Conv2D(False, nb_filters * 2, (6, 6), (2, 2), "VALID", phase, 'conv2'),
ReLU(),
Conv2D(False, nb_filters * 2, (5, 5), (1, 1), "VALID", phase, 'conv3'),
ReLU(),
Flatten(),
SimpleLinear(nb_classes),
SoftmaxT1()
]
model = MLP(layers, input_shape)
print('Finished making basic cnn')
return model
