def test_fgsm_defences(fix_get_mnist_subset, image_dl_estimator,
is_tf_version_2):
if is_tf_version_2:
clip_values = (0, 1)
smooth_3x3 = SpatialSmoothingTensorFlowV2(window_size=3,
channels_first=False)
smooth_5x5 = SpatialSmoothingTensorFlowV2(window_size=5,
channels_first=False)
smooth_7x7 = SpatialSmoothingTensorFlowV2(window_size=7,
channels_first=False)
classifier_, _ = image_dl_estimator(one_classifier=True)
loss_object = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
classifier = TensorFlowV2Classifier(
clip_values=clip_values,
model=classifier_.model,
preprocessing_defences=[smooth_3x3, smooth_5x5, smooth_7x7],
loss_object=loss_object,
input_shape=(28, 28, 1),
nb_classes=10,
)
assert len(classifier.preprocessing_defences) == 3
attack = FastGradientMethod(classifier, eps=1, batch_size=128)
backend_test_defended_images(attack, fix_get_mnist_subset)
def test_classifier_defended_images(art_warning, fix_get_mnist_subset, image_dl_estimator_for_attack):
try:
classifier = image_dl_estimator_for_attack(FastGradientMethod, defended=True)
attack = FastGradientMethod(classifier, eps=1.0, batch_size=128)
backend_test_defended_images(attack, fix_get_mnist_subset)
except ARTTestException as e:
art_warning(e)
def test_fgsm_defences(fix_get_mnist_subset, image_dl_estimator, device_type):
clip_values = (0, 1)
smooth_3x3 = SpatialSmoothingPyTorch(window_size=3,
channels_first=True,
device_type=device_type)
smooth_5x5 = SpatialSmoothingPyTorch(window_size=5,
channels_first=True,
device_type=device_type)
smooth_7x7 = SpatialSmoothingPyTorch(window_size=7,
channels_first=True,
device_type=device_type)
classifier_, _ = image_dl_estimator(one_classifier=True)
criterion = nn.CrossEntropyLoss()
classifier = PyTorchClassifier(
clip_values=clip_values,
model=classifier_.model,
preprocessing_defences=[smooth_3x3, smooth_5x5, smooth_7x7],
loss=criterion,
input_shape=(1, 28, 28),
nb_classes=10,
device_type=device_type,
)
assert len(classifier.preprocessing_defences) == 3
attack = FastGradientMethod(classifier, eps=1, batch_size=128)
backend_test_defended_images(attack, fix_get_mnist_subset)
def test_classifier_defended_images(fix_get_mnist_subset,
image_dl_estimator_for_attack):
classifier_list = image_dl_estimator_for_attack(FastGradientMethod,
defended=True)
# TODO this if statement must be removed once we have a classifier for both image and tabular data
if classifier_list is None:
logging.warning(
"Couldn't perform this test because no classifier is defined")
return
for classifier in classifier_list:
attack = FastGradientMethod(classifier, eps=1, batch_size=128)
backend_test_defended_images(attack, fix_get_mnist_subset)
