def test_subsetscan_detector(self):
(x_train, y_train), (x_test, y_test), _, _ = load_dataset("mnist")
x_train, y_train = x_train[:NB_TRAIN], y_train[:NB_TRAIN]
x_test, y_test = x_test[:NB_TEST], y_test[:NB_TEST]
# Keras classifier
classifier = get_image_classifier_kr()
# Generate adversarial samples:
attacker = FastGradientMethod(classifier, eps=0.5)
x_train_adv = attacker.generate(x_train)
x_test_adv = attacker.generate(x_test)
# Compile training data for detector:
x_train_detector = np.concatenate((x_train, x_train_adv), axis=0)
bgd = x_train
clean = x_test
anom = x_test_adv
detector = SubsetScanningDetector(classifier, bgd, layer=1)
_, _, dpwr = detector.scan(clean, clean)
self.assertAlmostEqual(dpwr, 0.5)
_, _, dpwr = detector.scan(clean, anom)
self.assertGreater(dpwr, 0.5)
_, _, dpwr = detector.scan(clean, x_train_detector, 85, 15)
self.assertGreater(dpwr, 0.5)
def test_without_defences(self):
(x_train, y_train), (x_test, y_test) = self.mnist
# Get the ready-trained Keras model and wrap it in query efficient gradient estimator wrapper
classifier = QueryEfficientBBGradientEstimation(self.classifier_k,
20,
1 / 64.,
round_samples=1 / 255.)
attack = FastGradientMethod(classifier, eps=1)
x_train_adv = attack.generate(x_train)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_train == x_train_adv).all())
self.assertFalse((x_test == x_test_adv).all())
train_y_pred = get_labels_np_array(classifier.predict(x_train_adv))
test_y_pred = get_labels_np_array(classifier.predict(x_test_adv))
self.assertFalse((y_train == train_y_pred).all())
self.assertFalse((y_test == test_y_pred).all())
preds = classifier.predict(x_train_adv)
acc = np.sum(np.argmax(preds, axis=1) == np.argmax(
y_train, axis=1)) / y_train.shape[0]
logger.info(
'Accuracy on adversarial train examples with limited query info: %.2f%%',
(acc * 100))
preds = classifier.predict(x_test_adv)
acc = np.sum(np.argmax(preds, axis=1) == np.argmax(
y_test, axis=1)) / y_test.shape[0]
logger.info(
'Accuracy on adversarial test examples with limited query info: %.2f%%',
(acc * 100))
def test_2_pt(self):
"""
Test with a PyTorch Classifier.
:return:
"""
# Build KerasClassifier
ptc = get_image_classifier_pt()
# Get MNIST
(_, _), (x_test, y_test) = self.mnist
x_test = x_test.transpose(0, 3, 1, 2).astype(np.float32)
# First FGSM attack:
fgsm = FastGradientMethod(estimator=ptc, targeted=True)
params = {"y": random_targets(y_test, ptc.nb_classes)}
x_test_adv = fgsm.generate(x_test, **params)
# Initialize RS object and attack with FGSM
rs = PyTorchRandomizedSmoothing(
model=ptc.model,
loss=ptc._loss,
optimizer=torch.optim.Adam(ptc.model.parameters(), lr=0.01),
input_shape=ptc.input_shape,
nb_classes=ptc.nb_classes,
channels_first=ptc.channels_first,
clip_values=ptc.clip_values,
sample_size=100,
scale=0.01,
alpha=0.001,
)
fgsm_with_rs = FastGradientMethod(estimator=rs, targeted=True)
x_test_adv_with_rs = fgsm_with_rs.generate(x_test, **params)
# Compare results
# check shapes are equal and values are within a certain range
self.assertEqual(x_test_adv.shape, x_test_adv_with_rs.shape)
self.assertTrue((np.abs(x_test_adv - x_test_adv_with_rs) < 0.75).all())
# Check basic functionality of RS object
# check predict
y_test_smooth = rs.predict(x=x_test)
y_test_base = ptc.predict(x=x_test)
self.assertEqual(y_test_smooth.shape, y_test.shape)
self.assertTrue((np.sum(y_test_smooth, axis=1) <= np.ones((NB_TEST,))).all())
self.assertTrue((np.argmax(y_test_smooth, axis=1) == np.argmax(y_test_base, axis=1)).all())
# check certification
pred, radius = rs.certify(x=x_test, n=250)
self.assertEqual(len(pred), NB_TEST)
self.assertEqual(len(radius), NB_TEST)
self.assertTrue((radius <= 1).all())
self.assertTrue((pred < y_test.shape[1]).all())
# loss gradient
grad = rs.loss_gradient(x=x_test, y=y_test, sampling=True)
assert grad.shape == (10, 1, 28, 28)
# fit
rs.fit(x=x_test, y=y_test)
def test_krclassifier(self):
"""
Test with a KerasClassifier.
:return:
"""
# Build KerasClassifier
krc = get_image_classifier_kr()
# Get MNIST
(_, _), (x_test, y_test) = self.mnist
# First attack (without EoT):
fgsm = FastGradientMethod(estimator=krc, targeted=True)
params = {"y": random_targets(y_test, krc.nb_classes)}
x_test_adv = fgsm.generate(x_test, **params)
# Second attack (with EoT):
def t(x):
return x
def transformation():
while True:
yield t
eot = ExpectationOverTransformations(classifier=krc,
sample_size=1,
transformation=transformation)
fgsm_with_eot = FastGradientMethod(estimator=eot, targeted=True)
x_test_adv_with_eot = fgsm_with_eot.generate(x_test, **params)
self.assertTrue(
(np.abs(x_test_adv - x_test_adv_with_eot) < 0.001).all())
def test_with_defences(self):
(x_train, y_train), (x_test, y_test) = self.mnist
# Get the trained Keras model
model = self.classifier_k._model
fs = FeatureSqueezing(bit_depth=1, clip_values=(0, 1))
classifier = KerasClassifier(model=model,
clip_values=(0, 1),
preprocessing_defences=fs)
# Create the classifier
classifier = QueryEfficientGradientEstimationClassifier(
classifier, 20, 1 / 64.0, round_samples=1 / 255.0)
attack = FastGradientMethod(classifier, eps=1)
x_train_adv = attack.generate(x_train)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_train == x_train_adv).all())
self.assertFalse((x_test == x_test_adv).all())
train_y_pred = get_labels_np_array(classifier.predict(x_train_adv))
test_y_pred = get_labels_np_array(classifier.predict(x_test_adv))
self.assertFalse((y_train == train_y_pred).all())
self.assertFalse((y_test == test_y_pred).all())
def test_3_kr(self):
"""
Test with a Keras Classifier.
:return:
"""
# Build KerasClassifier
classifier = get_image_classifier_kr()
# Get MNIST
(_, _), (x_test, y_test) = self.mnist
# First FGSM attack:
fgsm = FastGradientMethod(estimator=classifier, targeted=True)
params = {"y": random_targets(y_test, classifier.nb_classes)}
x_test_adv = fgsm.generate(x_test, **params)
# Initialize RS object and attack with FGSM
rs = NumpyRandomizedSmoothing(
classifier=classifier,
sample_size=100,
scale=0.01,
alpha=0.001,
)
fgsm_with_rs = FastGradientMethod(estimator=rs, targeted=True)
x_test_adv_with_rs = fgsm_with_rs.generate(x_test, **params)
# Compare results
# check shapes are equal and values are within a certain range
self.assertEqual(x_test_adv.shape, x_test_adv_with_rs.shape)
self.assertTrue((np.abs(x_test_adv - x_test_adv_with_rs) < 0.75).all())
# Check basic functionality of RS object
# check predict
y_test_smooth = rs.predict(x=x_test)
y_test_base = classifier.predict(x=x_test)
self.assertEqual(y_test_smooth.shape, y_test.shape)
self.assertTrue((np.sum(y_test_smooth, axis=1) <= np.ones((NB_TEST,))).all())
self.assertTrue((np.argmax(y_test_smooth, axis=1) == np.argmax(y_test_base, axis=1)).all())
# check certification
pred, radius = rs.certify(x=x_test, n=250)
self.assertEqual(len(pred), NB_TEST)
self.assertEqual(len(radius), NB_TEST)
self.assertTrue((radius <= 1).all())
self.assertTrue((pred < y_test.shape[1]).all())
# loss gradient
grad = rs.loss_gradient(x=x_test, y=y_test, sampling=True)
assert grad.shape == (10, 28, 28, 1)
# fit
rs.fit(x=x_test, y=y_test)
def test_binary_activation_detector(self):
"""
Test the binary activation detector end-to-end.
:return:
"""
# Get MNIST
(x_train, y_train), (x_test, y_test), _, _ = load_mnist()
x_train, y_train = x_train[:NB_TRAIN], y_train[:NB_TRAIN]
x_test, y_test = x_test[:NB_TEST], y_test[:NB_TEST]
# Keras classifier
classifier = get_image_classifier_kr()
# Generate adversarial samples:
attacker = FastGradientMethod(classifier, eps=0.1)
x_train_adv = attacker.generate(x_train[:NB_TRAIN])
x_test_adv = attacker.generate(x_test[:NB_TRAIN])
# Compile training data for detector:
x_train_detector = np.concatenate((x_train[:NB_TRAIN], x_train_adv), axis=0)
y_train_detector = np.concatenate((np.array([[1, 0]] * NB_TRAIN), np.array([[0, 1]] * NB_TRAIN)), axis=0)
# Create a simple CNN for the detector
activation_shape = classifier.get_activations(x_test[:1], 0, batch_size=128).shape[1:]
number_outputs = 2
model = Sequential()
model.add(MaxPooling2D(pool_size=(2, 2), input_shape=activation_shape))
model.add(Flatten())
model.add(Dense(number_outputs, activation="softmax"))
model.compile(
loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=0.01), metrics=["accuracy"]
)
# Create detector and train it.
# Detector consider activations at layer=0:
detector = BinaryActivationDetector(
classifier=classifier, detector=KerasClassifier(model=model, clip_values=(0, 1), use_logits=False), layer=0
)
detector.fit(x_train_detector, y_train_detector, nb_epochs=2, batch_size=128)
# Apply detector on clean and adversarial test data:
test_detection = np.argmax(detector.predict(x_test), axis=1)
test_adv_detection = np.argmax(detector.predict(x_test_adv), axis=1)
# Assert there is at least one true positive and negative
nb_true_positives = len(np.where(test_adv_detection == 1)[0])
nb_true_negatives = len(np.where(test_detection == 0)[0])
logger.debug("Number of true positives detected: %i", nb_true_positives)
logger.debug("Number of true negatives detected: %i", nb_true_negatives)
self.assertGreater(nb_true_positives, 0)
self.assertGreater(nb_true_negatives, 0)
def test_two_attacks(self):
(x_train, y_train), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
attack1 = FastGradientMethod(estimator=self.classifier, batch_size=16)
attack2 = DeepFool(classifier=self.classifier,
max_iter=5,
batch_size=16)
x_test_adv = attack1.generate(x_test)
predictions = np.argmax(self.classifier.predict(x_test_adv), axis=1)
accuracy = np.sum(predictions == np.argmax(y_test, axis=1)) / NB_TEST
adv_trainer = AdversarialTrainer(self.classifier,
attacks=[attack1, attack2])
adv_trainer.fit(x_train, y_train, nb_epochs=2, batch_size=16)
predictions_new = np.argmax(adv_trainer.predict(x_test_adv), axis=1)
accuracy_new = np.sum(
predictions_new == np.argmax(y_test, axis=1)) / NB_TEST
self.assertEqual(accuracy_new, 0.36)
self.assertEqual(accuracy, 0.13)
# Check that x_test has not been modified by attack and classifier
self.assertAlmostEqual(float(np.max(np.abs(x_test_original - x_test))),
0.0,
delta=0.00001)
def _fgsm(model, data, labels, attack_args):
"""
Fast Gradient Sign Method
Explaining and Harnessing Adversarial Examples
by Ian J. Goodfellow, Jonathon Shlens, Christian Szegedy
``https://arxiv.org/abs/1412.6572``
:param model:
:param data:
:param labels:
:param attack_args:
:return:
"""
print('>>> Generating FGSM examples.')
eps = attack_args.get('eps', 0.3)
targeted = attack_args.get('targeted', False)
num_random_init = attack_args.get('num_random_init', 0)
minimal = attack_args.get('minimal', False)
attacker = FastGradientMethod(model,
eps=eps,
eps_step=eps,
targeted=targeted,
num_random_init=num_random_init,
minimal=minimal)
return attacker.generate(data, labels)
def test_iris_clipped(self):
(_, _), (x_test, y_test) = self.iris
def t(x):
return x
def transformation():
while True:
yield t
classifier = get_tabular_classifier_kr()
classifier = ExpectationOverTransformations(
classifier, sample_size=1, transformation=transformation)
# Test untargeted attack
attack = FastGradientMethod(classifier, eps=0.1)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv <= 1).all())
self.assertTrue((x_test_adv >= 0).all())
preds_adv = np.argmax(classifier.predict(x_test_adv), axis=1)
self.assertFalse((np.argmax(y_test, axis=1) == preds_adv).all())
acc = np.sum(preds_adv == np.argmax(y_test, axis=1)) / y_test.shape[0]
logger.info("Accuracy on Iris with limited query info: %.2f%%",
(acc * 100))
def test_iris_unbounded(self):
(_, _), (x_test, y_test) = self.iris
classifier = get_tabular_classifier_kr()
def t(x):
return x
def transformation():
while True:
yield t
# Recreate a classifier without clip values
classifier = KerasClassifier(model=classifier._model,
use_logits=False,
channels_first=True)
classifier = ExpectationOverTransformations(
classifier, sample_size=1, transformation=transformation)
attack = FastGradientMethod(classifier, eps=1)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv > 1).any())
self.assertTrue((x_test_adv < 0).any())
preds_adv = np.argmax(classifier.predict(x_test_adv), axis=1)
self.assertFalse((np.argmax(y_test, axis=1) == preds_adv).all())
acc = np.sum(preds_adv == np.argmax(y_test, axis=1)) / y_test.shape[0]
logger.info("Accuracy on Iris with limited query info: %.2f%%",
(acc * 100))
def test_without_defences(self):
(x_train, y_train), (x_test, y_test) = self.mnist
# Get the ready-trained Keras model and wrap it in query efficient gradient estimator
classifier = QueryEfficientGradientEstimationClassifier(
self.classifier_k, 20, 1 / 64.0, round_samples=1 / 255.0)
attack = FastGradientMethod(classifier, eps=1)
x_train_adv = attack.generate(x_train)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_train == x_train_adv).all())
self.assertFalse((x_test == x_test_adv).all())
train_y_pred = get_labels_np_array(classifier.predict(x_train_adv))
test_y_pred = get_labels_np_array(classifier.predict(x_test_adv))
self.assertFalse((y_train == train_y_pred).all())
self.assertFalse((y_test == test_y_pred).all())
def test_with_defences(self):
(x_train, y_train), (x_test, y_test) = self.mnist
# Get the ready-trained Keras model
model = self.classifier_k._model
fs = FeatureSqueezing(bit_depth=1, clip_values=(0, 1))
classifier = KerasClassifier(model=model,
clip_values=(0, 1),
preprocessing_defences=fs)
# Wrap the classifier
classifier = QueryEfficientBBGradientEstimation(classifier,
20,
1 / 64.0,
round_samples=1 /
255.0)
attack = FastGradientMethod(classifier, eps=1)
x_train_adv = attack.generate(x_train)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_train == x_train_adv).all())
self.assertFalse((x_test == x_test_adv).all())
train_y_pred = get_labels_np_array(classifier.predict(x_train_adv))
test_y_pred = get_labels_np_array(classifier.predict(x_test_adv))
self.assertFalse((y_train == train_y_pred).all())
self.assertFalse((y_test == test_y_pred).all())
preds = classifier.predict(x_train_adv)
acc = np.sum(np.argmax(preds, axis=1) == np.argmax(
y_train, axis=1)) / y_train.shape[0]
logger.info(
"Accuracy on adversarial train examples with feature squeezing and limited query info: %.2f%%",
(acc * 100))
preds = classifier.predict(x_test_adv)
acc = np.sum(np.argmax(preds, axis=1) == np.argmax(
y_test, axis=1)) / y_test.shape[0]
logger.info(
"Accuracy on adversarial test examples with feature squeezing and limited query info: %.2f%%",
(acc * 100))
def test_two_attacks_with_generator(self):
(x_train, y_train), (x_test, y_test) = self.mnist
x_train_original = x_train.copy()
x_test_original = x_test.copy()
class MyDataGenerator(DataGenerator):
def __init__(self, x, y, size, batch_size):
super().__init__(size=size, batch_size=batch_size)
self.x = x
self.y = y
self._size = size
self._batch_size = batch_size
def get_batch(self):
ids = np.random.choice(self.size,
size=min(self.size, self.batch_size),
replace=False)
return self.x[ids], self.y[ids]
generator = MyDataGenerator(x_train,
y_train,
size=x_train.shape[0],
batch_size=16)
attack1 = FastGradientMethod(estimator=self.classifier, batch_size=16)
attack2 = DeepFool(classifier=self.classifier,
max_iter=5,
batch_size=16)
x_test_adv = attack1.generate(x_test)
predictions = np.argmax(self.classifier.predict(x_test_adv), axis=1)
accuracy = np.sum(predictions == np.argmax(y_test, axis=1)) / NB_TEST
adv_trainer = AdversarialTrainer(self.classifier,
attacks=[attack1, attack2])
adv_trainer.fit_generator(generator, nb_epochs=3)
predictions_new = np.argmax(adv_trainer.predict(x_test_adv), axis=1)
accuracy_new = np.sum(
predictions_new == np.argmax(y_test, axis=1)) / NB_TEST
self.assertAlmostEqual(accuracy_new, 0.25, delta=0.02)
self.assertAlmostEqual(accuracy, 0.11, delta=0.0)
# Check that x_train and x_test has not been modified by attack and classifier
self.assertAlmostEqual(float(np.max(np.abs(x_train_original -
x_train))),
0.0,
delta=0.00001)
self.assertAlmostEqual(float(np.max(np.abs(x_test_original - x_test))),
0.0,
delta=0.00001)
def test_fit_predict_different_classifiers(self):
(x_train, y_train), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
attack = FastGradientMethod(self.classifier)
x_test_adv = attack.generate(x_test)
predictions = np.argmax(self.classifier.predict(x_test_adv), axis=1)
accuracy = np.sum(predictions == np.argmax(y_test, axis=1)) / NB_TEST
adv_trainer = AdversarialTrainer(self.classifier_2, attack)
adv_trainer.fit(x_train, y_train, nb_epochs=5, batch_size=128)
predictions_new = np.argmax(adv_trainer.predict(x_test_adv), axis=1)
accuracy_new = np.sum(
predictions_new == np.argmax(y_test, axis=1)) / NB_TEST
self.assertEqual(accuracy_new, 0.32)
self.assertEqual(accuracy, 0.13)
# Check that x_test has not been modified by attack and classifier
self.assertAlmostEqual(float(np.max(np.abs(x_test_original - x_test))),
0.0,
delta=0.00001)
# fit_generator
class MyDataGenerator(DataGenerator):
def __init__(self, x, y, size, batch_size):
super().__init__(size=size, batch_size=batch_size)
self.x = x
self.y = y
self._size = size
self._batch_size = batch_size
def get_batch(self):
ids = np.random.choice(self.size,
size=min(self.size, self.batch_size),
replace=False)
return self.x[ids], self.y[ids]
generator = MyDataGenerator(x_train,
y_train,
size=x_train.shape[0],
batch_size=16)
adv_trainer.fit_generator(generator, nb_epochs=5)
adv_trainer_2 = AdversarialTrainer(self.classifier_2,
attack,
ratio=1.0)
adv_trainer_2.fit_generator(generator, nb_epochs=5)
def test_iris_clipped(self):
(_, _), (x_test, y_test) = self.iris
classifier = get_tabular_classifier_kr()
classifier = QueryEfficientGradientEstimationClassifier(
classifier, 20, 1 / 64.0, round_samples=1 / 255.0)
# Test untargeted attack
attack = FastGradientMethod(classifier, eps=0.1)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv <= 1).all())
self.assertTrue((x_test_adv >= 0).all())
preds_adv = np.argmax(classifier.predict(x_test_adv), axis=1)
self.assertFalse((np.argmax(y_test, axis=1) == preds_adv).all())
def test_iris_clipped(self):
(_, _), (x_test, y_test) = self.iris
ptc = get_tabular_classifier_pt()
rs = PyTorchRandomizedSmoothing(
model=ptc.model,
loss=ptc._loss,
input_shape=ptc.input_shape,
nb_classes=ptc.nb_classes,
channels_first=ptc.channels_first,
clip_values=ptc.clip_values,
sample_size=100,
scale=0.01,
alpha=0.001,
)
# Test untargeted attack
attack = FastGradientMethod(ptc, eps=0.1)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv <= 1).all())
self.assertTrue((x_test_adv >= 0).all())
preds_smooth = np.argmax(rs.predict(x_test_adv), axis=1)
self.assertFalse((np.argmax(y_test, axis=1) == preds_smooth).all())
pred = rs.predict(x_test)
pred2 = rs.predict(x_test_adv)
acc, cov = compute_accuracy(pred, y_test)
acc2, cov2 = compute_accuracy(pred2, y_test)
logger.info("Accuracy on Iris with smoothing on adversarial examples: %.2f%%", (acc * 100))
logger.info("Coverage on Iris with smoothing on adversarial examples: %.2f%%", (cov * 100))
logger.info("Accuracy on Iris with smoothing: %.2f%%", (acc2 * 100))
logger.info("Coverage on Iris with smoothing: %.2f%%", (cov2 * 100))
# Check basic functionality of RS object
# check predict
y_test_smooth = rs.predict(x=x_test)
self.assertEqual(y_test_smooth.shape, y_test.shape)
self.assertTrue((np.sum(y_test_smooth, axis=1) <= 1).all())
# check certification
pred, radius = rs.certify(x=x_test, n=250)
self.assertEqual(len(pred), len(x_test))
self.assertEqual(len(radius), len(x_test))
self.assertTrue((radius <= 1).all())
self.assertTrue((pred < y_test.shape[1]).all())
def test_iris_unbounded(self):
(_, _), (x_test, y_test) = self.iris
classifier = get_tabular_classifier_kr()
# Recreate a classifier without clip values
classifier = KerasClassifier(model=classifier._model,
use_logits=False,
channels_first=True)
classifier = QueryEfficientGradientEstimationClassifier(
classifier, 20, 1 / 64.0, round_samples=1 / 255.0)
attack = FastGradientMethod(classifier, eps=1)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv > 1).any())
self.assertTrue((x_test_adv < 0).any())
preds_adv = np.argmax(classifier.predict(x_test_adv), axis=1)
self.assertFalse((np.argmax(y_test, axis=1) == preds_adv).all())
def test_iris_clipped(self):
(_, _), (x_test, y_test) = self.iris
classifier = get_iris_classifier_kr()
classifier = QueryEfficientBBGradientEstimation(classifier,
20,
1 / 64.,
round_samples=1 / 255.)
# Test untargeted attack
attack = FastGradientMethod(classifier, eps=.1)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv <= 1).all())
self.assertTrue((x_test_adv >= 0).all())
preds_adv = np.argmax(classifier.predict(x_test_adv), axis=1)
self.assertFalse((np.argmax(y_test, axis=1) == preds_adv).all())
acc = np.sum(preds_adv == np.argmax(y_test, axis=1)) / y_test.shape[0]
logger.info('Accuracy on Iris with limited query info: %.2f%%',
(acc * 100))
def test_fit_predict(self):
(x_train, y_train), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
attack = FastGradientMethod(self.classifier)
x_test_adv = attack.generate(x_test)
predictions = np.argmax(self.classifier.predict(x_test_adv), axis=1)
accuracy = np.sum(predictions == np.argmax(y_test, axis=1)) / NB_TEST
adv_trainer = AdversarialTrainer(self.classifier, attack)
adv_trainer.fit(x_train, y_train, nb_epochs=5, batch_size=128)
predictions_new = np.argmax(adv_trainer.predict(x_test_adv), axis=1)
accuracy_new = np.sum(
predictions_new == np.argmax(y_test, axis=1)) / NB_TEST
self.assertEqual(accuracy_new, 0.12)
self.assertEqual(accuracy, 0.13)
# Check that x_test has not been modified by attack and classifier
self.assertAlmostEqual(float(np.max(np.abs(x_test_original - x_test))),
0.0,
delta=0.00001)
def test_iris_unbounded(self):
(_, _), (x_test, y_test) = self.iris
classifier = get_iris_classifier_kr()
# Recreate a classifier without clip values
classifier = KerasClassifier(model=classifier._model,
use_logits=False,
channel_index=1)
classifier = QueryEfficientBBGradientEstimation(classifier,
20,
1 / 64.,
round_samples=1 / 255.)
attack = FastGradientMethod(classifier, eps=1)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv > 1).any())
self.assertTrue((x_test_adv < 0).any())
preds_adv = np.argmax(classifier.predict(x_test_adv), axis=1)
self.assertFalse((np.argmax(y_test, axis=1) == preds_adv).all())
acc = np.sum(preds_adv == np.argmax(y_test, axis=1)) / y_test.shape[0]
logger.info('Accuracy on Iris with limited query info: %.2f%%',
(acc * 100))
def robustness_evaluation(object_storage_url,
object_storage_username,
object_storage_password,
data_bucket_name,
result_bucket_name,
model_id,
feature_testset_path='processed_data/X_test.npy',
label_testset_path='processed_data/y_test.npy',
clip_values=(0, 1),
nb_classes=2,
input_shape=(1, 3, 64, 64),
model_class_file='model.py',
model_class_name='model',
LossFn='',
Optimizer='',
epsilon=0.2):
url = re.compile(r"https?://")
cos = Minio(url.sub('', object_storage_url),
access_key=object_storage_username,
secret_key=object_storage_password,
secure=False)
dataset_filenamex = "X_test.npy"
dataset_filenamey = "y_test.npy"
weights_filename = "model.pt"
model_files = model_id + '/_submitted_code/model.zip'
cos.fget_object(data_bucket_name, feature_testset_path, dataset_filenamex)
cos.fget_object(data_bucket_name, label_testset_path, dataset_filenamey)
cos.fget_object(result_bucket_name, model_id + '/' + weights_filename,
weights_filename)
cos.fget_object(result_bucket_name, model_files, 'model.zip')
# Load PyTorch model definition from the source code.
zip_ref = zipfile.ZipFile('model.zip', 'r')
zip_ref.extractall('model_files')
zip_ref.close()
modulename = 'model_files.' + model_class_file.split('.')[0].replace(
'-', '_')
'''
We required users to define where the model class is located or follow
some naming convention we have provided.
'''
model_class = getattr(importlib.import_module(modulename),
model_class_name)
# load & compile model
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model = model_class().to(device)
model.load_state_dict(torch.load(weights_filename, map_location=device))
# Define Loss and optimizer function for the PyTorch model
if LossFn:
loss_fn = eval(LossFn)
else:
loss_fn = torch.nn.CrossEntropyLoss()
if Optimizer:
optimizer = eval(Optimizer)
else:
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
# create pytorch classifier
classifier = PyTorchClassifier(clip_values, model, loss_fn, optimizer,
input_shape, nb_classes)
# load test dataset
x = np.load(dataset_filenamex)
y = np.load(dataset_filenamey)
# craft adversarial samples using FGSM
crafter = FastGradientMethod(classifier, eps=epsilon)
x_samples = crafter.generate(x)
# obtain all metrics (robustness score, perturbation metric, reduction in confidence)
metrics, y_pred_orig, y_pred_adv = get_metrics(model, x, x_samples, y)
print("metrics:", metrics)
return metrics
def test_binary_input_detector(self):
"""
Test the binary input detector end-to-end.
:return:
"""
# Get MNIST
nb_train, nb_test = 1000, 10
(x_train, y_train), (x_test, y_test), _, _ = load_mnist()
x_train, y_train = x_train[:NB_TRAIN], y_train[:NB_TRAIN]
x_test, y_test = x_test[:NB_TEST], y_test[:NB_TEST]
# Keras classifier
classifier = get_image_classifier_kr()
# Generate adversarial samples:
attacker = FastGradientMethod(classifier, eps=0.1)
x_train_adv = attacker.generate(x_train[:nb_train])
x_test_adv = attacker.generate(x_test[:nb_test])
# Compile training data for detector:
x_train_detector = np.concatenate((x_train[:nb_train], x_train_adv),
axis=0)
y_train_detector = np.concatenate(
(np.array([[1, 0]] * nb_train), np.array([[0, 1]] * nb_train)),
axis=0)
# Create a simple CNN for the detector
input_shape = x_train.shape[1:]
try:
from keras.optimizers import Adam
optimizer = Adam(lr=0.01)
except ImportError:
from keras.optimizers import adam_v2
optimizer = adam_v2.Adam(lr=0.01)
model = Sequential()
model.add(
Conv2D(4,
kernel_size=(5, 5),
activation="relu",
input_shape=input_shape))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(2, activation="softmax"))
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=optimizer,
metrics=["accuracy"])
# Create detector and train it:
detector = BinaryInputDetector(
KerasClassifier(model=model, clip_values=(0, 1), use_logits=False))
detector.fit(x_train_detector,
y_train_detector,
nb_epochs=2,
batch_size=128)
# Apply detector on clean and adversarial test data:
test_detection = np.argmax(detector.predict(x_test), axis=1)
test_adv_detection = np.argmax(detector.predict(x_test_adv), axis=1)
# Assert there is at least one true positive and negative:
nb_true_positives = len(np.where(test_adv_detection == 1)[0])
nb_true_negatives = len(np.where(test_detection == 0)[0])
logger.debug("Number of true positives detected: %i",
nb_true_positives)
logger.debug("Number of true negatives detected: %i",
nb_true_negatives)
self.assertGreater(nb_true_positives, 0)
self.assertGreater(nb_true_negatives, 0)
def test_1_tf(self):
"""
Test with a TensorFlow Classifier.
:return:
"""
tf_version = list(map(int, tf.__version__.lower().split("+")[0].split(".")))
if tf_version[0] == 2:
# Build TensorFlowV2Classifier
classifier, _ = get_image_classifier_tf()
# Get MNIST
(_, _), (x_test, y_test) = self.mnist
# First FGSM attack:
fgsm = FastGradientMethod(estimator=classifier, targeted=True)
params = {"y": random_targets(y_test, classifier.nb_classes)}
x_test_adv = fgsm.generate(x_test, **params)
# Initialize RS object and attack with FGSM
rs = TensorFlowV2RandomizedSmoothing(
model=classifier.model,
nb_classes=classifier.nb_classes,
input_shape=classifier.input_shape,
loss_object=classifier.loss_object,
train_step=classifier.train_step,
channels_first=classifier.channels_first,
clip_values=classifier.clip_values,
preprocessing_defences=classifier.preprocessing_defences,
postprocessing_defences=classifier.postprocessing_defences,
preprocessing=classifier.preprocessing,
sample_size=100,
scale=0.01,
alpha=0.001,
)
fgsm_with_rs = FastGradientMethod(estimator=rs, targeted=True)
x_test_adv_with_rs = fgsm_with_rs.generate(x_test, **params)
# Compare results
# check shapes are equal and values are within a certain range
self.assertEqual(x_test_adv.shape, x_test_adv_with_rs.shape)
self.assertTrue((np.abs(x_test_adv - x_test_adv_with_rs) < 0.75).all())
# Check basic functionality of RS object
# check predict
y_test_smooth = rs.predict(x=x_test)
y_test_base = classifier.predict(x=x_test)
self.assertEqual(y_test_smooth.shape, y_test.shape)
self.assertTrue((np.sum(y_test_smooth, axis=1) <= np.ones((NB_TEST,))).all())
self.assertTrue((np.argmax(y_test_smooth, axis=1) == np.argmax(y_test_base, axis=1)).all())
# check certification
pred, radius = rs.certify(x=x_test, n=250)
self.assertEqual(len(pred), NB_TEST)
self.assertEqual(len(radius), NB_TEST)
self.assertTrue((radius <= 1).all())
self.assertTrue((pred < y_test.shape[1]).all())
# loss gradient
grad = rs.loss_gradient(x=x_test, y=y_test, sampling=True)
assert grad.shape == (10, 28, 28, 1)
# fit
rs.fit(x=x_test, y=y_test)
