def test_tensorflow_mnist(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
x_test_original = self.x_test_mnist.copy()
# Build TensorFlowClassifier
tfc, sess = get_image_classifier_tf()
# Targeted attack
zoo = ZooAttack(classifier=tfc,
targeted=True,
max_iter=30,
binary_search_steps=8,
batch_size=128)
params = {"y": random_targets(self.y_test_mnist, tfc.nb_classes())}
x_test_mnist_adv = zoo.generate(self.x_test_mnist, **params)
self.assertFalse((self.x_test_mnist == x_test_mnist_adv).all())
self.assertLessEqual(np.amax(x_test_mnist_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_mnist_adv), 0.0)
target = np.argmax(params["y"], axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_mnist_adv), axis=1)
logger.debug("ZOO target: %s", target)
logger.debug("ZOO actual: %s", y_pred_adv)
logger.info("ZOO success rate on MNIST: %.2f",
(sum(target == y_pred_adv) / float(len(target))))
# Untargeted attack
zoo = ZooAttack(classifier=tfc,
targeted=False,
max_iter=10,
binary_search_steps=3)
x_test_mnist_adv = zoo.generate(self.x_test_mnist)
# self.assertFalse((x_test == x_test_adv).all())
self.assertLessEqual(np.amax(x_test_mnist_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_mnist_adv), 0.0)
y_pred = np.argmax(tfc.predict(self.x_test_mnist), axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_mnist_adv), axis=1)
logger.debug("ZOO actual: %s", y_pred_adv)
logger.info("ZOO success rate on MNIST: %.2f",
(sum(y_pred != y_pred_adv) / float(len(y_pred))))
# Check that x_test has not been modified by attack and classifier
self.assertAlmostEqual(float(
np.max(np.abs(x_test_original - self.x_test_mnist))),
0.0,
delta=0.00001)
# Clean-up session
if sess is not None:
sess.close()
def test_tensorflow_failure_attack(self):
"""
Test the corner case when attack fails.
:return:
"""
x_test_original = self.x_test_mnist.copy()
# Build TensorFlowClassifier
tfc, sess = get_image_classifier_tf()
# Failure attack
zoo = ZooAttack(classifier=tfc,
max_iter=0,
binary_search_steps=0,
learning_rate=0)
x_test_mnist_adv = zoo.generate(self.x_test_mnist)
self.assertLessEqual(np.amax(x_test_mnist_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_mnist_adv), 0.0)
np.testing.assert_almost_equal(self.x_test_mnist, x_test_mnist_adv, 3)
# Check that x_test has not been modified by attack and classifier
self.assertAlmostEqual(float(
np.max(np.abs(x_test_original - self.x_test_mnist))),
0.0,
delta=0.00001)
# Clean-up session
if sess is not None:
sess.close()
def test_keras_mnist(self):
"""
Second test with the KerasClassifier.
:return:
"""
x_test_original = self.x_test.copy()
# Build KerasClassifier
krc = get_classifier_kr()
# Targeted attack
# zoo = ZooAttack(classifier=krc, targeted=True, batch_size=5)
# params = {'y': random_targets(self.y_test, krc.nb_classes())}
# x_test_adv = zoo.generate(self.x_test, **params)
#
# self.assertFalse((self.x_test == x_test_adv).all())
# self.assertLessEqual(np.amax(x_test_adv), 1.0)
# self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
# target = np.argmax(params['y'], axis=1)
# y_pred_adv = np.argmax(krc.predict(x_test_adv), axis=1)
# logger.debug('ZOO target: %s', target)
# logger.debug('ZOO actual: %s', y_pred_adv)
# logger.info('ZOO success rate on MNIST: %.2f', (sum(target == y_pred_adv) / float(len(target))))
# Untargeted attack
# zoo = ZooAttack(classifier=krc, targeted=False, max_iter=20)
zoo = ZooAttack(classifier=krc, targeted=False, batch_size=5)
# x_test_adv = zoo.generate(x_test)
params = {'y': random_targets(self.y_test, krc.nb_classes())}
x_test_adv = zoo.generate(self.x_test, **params)
# x_test_adv_true = [0.00000000e+00, 2.50167388e-04, 1.50529508e-04, 4.69674182e-04,
# 0.00000000e+00, 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
# 1.67321396e-05, 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
# 0.00000000e+00, 2.08451956e-06, 0.00000000e+00, 0.00000000e+00,
# 2.53360748e-01, 9.60119188e-01, 9.85227525e-01, 2.53600776e-01,
# 0.00000000e+00, 0.00000000e+00, 5.23251540e-04, 0.00000000e+00,
# 0.00000000e+00, 0.00000000e+00, 1.08632184e-05, 0.00000000e+00]
#
# for i in range(14):
# self.assertAlmostEqual(x_test_adv_true[i], x_test_adv[0, 14, i, 0])
# self.assertFalse((x_test == x_test_adv).all())
self.assertLessEqual(np.amax(x_test_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
y_pred_adv = np.argmax(krc.predict(x_test_adv), axis=1)
y_pred = np.argmax(krc.predict(self.x_test), axis=1)
logger.debug('ZOO actual: %s', y_pred_adv)
logger.info('ZOO success rate on MNIST: %.2f', (sum(y_pred != y_pred_adv) / float(len(y_pred))))
# Check that x_test has not been modified by attack and classifier
self.assertAlmostEqual(float(np.max(np.abs(x_test_original - self.x_test))), 0.0, delta=0.00001)
# Clean-up
k.clear_session()
def test_tfclassifier(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf()
# Targeted attack
zoo = ZooAttack(classifier=tfc,
targeted=True,
max_iter=100,
binary_search_steps=10)
params = {'y': random_targets(self.y_test, tfc.nb_classes())}
x_test_adv = zoo.generate(self.x_test, **params)
self.assertFalse((self.x_test == x_test_adv).all())
self.assertLessEqual(np.amax(x_test_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
target = np.argmax(params['y'], axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_adv), axis=1)
logger.debug('ZOO target: %s', target)
logger.debug('ZOO actual: %s', y_pred_adv)
logger.info('ZOO success rate on MNIST: %.2f',
(sum(target == y_pred_adv) / float(len(target))))
# Untargeted attack
zoo = ZooAttack(classifier=tfc, targeted=False)
x_test_adv = zoo.generate(self.x_test)
# self.assertFalse((x_test == x_test_adv).all())
self.assertLessEqual(np.amax(x_test_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
y_pred = np.argmax(tfc.predict(self.x_test), axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_adv), axis=1)
logger.debug('ZOO actual: %s', y_pred_adv)
logger.info('ZOO success rate on MNIST: %.2f',
(sum(y_pred != y_pred_adv) / float(len(y_pred))))
# Clean-up session
sess.close()
def test_pytorch_mnist(self):
"""
Third test with the PyTorchClassifier.
:return:
"""
# Build PyTorchClassifier
ptc = get_image_classifier_pt()
# Get MNIST
x_test_mnist = np.swapaxes(self.x_test_mnist, 1, 3).astype(np.float32)
x_test_original = x_test_mnist.copy()
# First attack
# zoo = ZooAttack(classifier=ptc, targeted=True, max_iter=10, binary_search_steps=10)
# params = {'y': random_targets(self.y_test, ptc.nb_classes())}
# x_test_adv = zoo.generate(x_test, **params)
# self.assertFalse((x_test == x_test_adv).all())
# self.assertLessEqual(np.amax(x_test_adv), 1.0)
# self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
# target = np.argmax(params['y'], axis=1)
# y_pred_adv = np.argmax(ptc.predict(x_test_adv), axis=1)
# logger.debug('ZOO target: %s', target)
# logger.debug('ZOO actual: %s', y_pred_adv)
# logger.info('ZOO success rate on MNIST: %.2f', (sum(target != y_pred_adv) / float(len(target))))
# Second attack
zoo = ZooAttack(
classifier=ptc,
targeted=False,
learning_rate=1e-2,
max_iter=10,
binary_search_steps=3,
abort_early=False,
use_resize=False,
use_importance=False,
)
x_test_mnist_adv = zoo.generate(x_test_mnist)
self.assertLessEqual(np.amax(x_test_mnist_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_mnist_adv), 0.0)
# print(x_test[0, 0, 14, :])
# print(x_test_adv[0, 0, 14, :])
# print(np.amax(x_test - x_test_adv))
x_test_adv_expected = []
# Check that x_test has not been modified by attack and classifier
self.assertAlmostEqual(float(
np.max(np.abs(x_test_original - x_test_mnist))),
0.0,
delta=0.00001)
def test_failure_attack(self):
"""
Test the corner case when attack fails.
:return:
"""
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf()
# Failure attack
zoo = ZooAttack(classifier=tfc,
max_iter=0,
binary_search_steps=0,
learning_rate=0)
x_test_adv = zoo.generate(self.x_test)
self.assertLessEqual(np.amax(x_test_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
np.testing.assert_almost_equal(self.x_test, x_test_adv, 3)
# Clean-up session
sess.close()
# Step 5: Evaluate the ART classifier on benign test examples
predictions = classifier.predict(x_test)
accuracy = np.sum(np.argmax(predictions, axis=1) == np.argmax(y_test, axis=1)) / len(y_test)
print("Accuracy on benign test examples: {}%".format(accuracy * 100))
# Step 6: Generate adversarial test examples
attack = ZooAttack(
classifier=classifier,
confidence=0.0,
targeted=False,
learning_rate=1e-1,
max_iter=200,
binary_search_steps=10,
initial_const=1e-3,
abort_early=True,
use_resize=False,
use_importance=False,
nb_parallel=5,
batch_size=1,
variable_h=0.01,
)
x_test_adv = attack.generate(x=x_test, y=y_test)
# Step 7: Evaluate the ART classifier on adversarial test examples
predictions = classifier.predict(x_test_adv)
accuracy = np.sum(np.argmax(predictions, axis=1) == np.argmax(y_test, axis=1)) / len(y_test)
print("Accuracy on adversarial test examples: {}%".format(accuracy * 100))
def adversarial_attack_shift(x, y, delta=1.0, model=RandomForestClassifier(), attack_type='zoo',
numerical_features=None, feat_delta=1.0):
# in this case delta is the portion of half the data on which to generate attacks
# because the first half as a minimum has to be used to train a model against which generate the attacks
assert (attack_type in ['zoo', 'boundary', 'hop-skip-jump'])
le = preprocessing.LabelEncoder()
le.fit(np.squeeze(y))
y = le.transform(y)
x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=(0.5 * delta))
if numerical_features is not None:
n_numerical = len(numerical_features)
feat_indices = np.random.choice(n_numerical, ceil(n_numerical * feat_delta), replace=False)
feat_indices = np.array(numerical_features)[feat_indices]
else:
feat_indices = np.random.choice(x.shape[1], ceil(x.shape[1] * feat_delta), replace=False)
other_features = list(set(range(x.shape[1])) - set(feat_indices))
x_train_other = x_train[:, other_features]
x_train_numerical = x_train[:, feat_indices]
x_test_other = x_test[:, other_features]
x_test_numerical = x_test[:, feat_indices]
classifier = SklearnClassifier(model=model, clip_values=(0, np.max(x_train_numerical)))
# Train the ART classifier
classifier.fit(x_train_numerical, y_train)
# Evaluate the ART classifier on benign test examples
predictions = classifier.predict(x_test_numerical)
accuracy = np.sum(np.argmax(predictions, axis=1) == y_test) / len(y_test)
print("Accuracy on benign test examples: {}%".format(accuracy * 100))
# Generate adversarial test examples
if attack_type == 'zoo':
attack = ZooAttack(
classifier=classifier,
confidence=0.0,
targeted=False,
learning_rate=1e-1,
max_iter=10,
binary_search_steps=10,
initial_const=1e-3,
abort_early=True,
use_resize=False,
use_importance=False,
nb_parallel=x_test_numerical.shape[1],
batch_size=1,
variable_h=0.01,
)
elif attack_type == 'boundary':
attack = BoundaryAttack(classifier, targeted=False, epsilon=0.02, max_iter=20, num_trial=10)
elif attack_type == 'hop-skip-jump':
attack = HopSkipJump(classifier,
targeted=False,
norm=2,
max_iter=20,
max_eval=10,
init_eval=9,
init_size=10)
x_adv = attack.generate(x=x_test_numerical, y=y_test)
# Evaluate the ART classifier on adversarial test examples
predictions_adv = classifier.predict(x_adv)
accuracy = np.sum(np.argmax(predictions_adv, axis=1) == y_test) / len(y_test)
print("Accuracy on adversarial test examples: {}%".format(accuracy * 100))
print("Max difference: {}".format(np.max(np.abs(x_test_numerical - x_adv) / x_test_numerical)))
x_final = np.zeros_like(x)
x_final[:, feat_indices] = np.vstack([x_train_numerical, x_adv])
x_final[:, other_features] = np.vstack([x_train_other, x_test_other])
y_final = np.concatenate([y_train, y_test], axis=0)
y_final = le.inverse_transform(y_final)
adv_indices = list(range(len(y_train), len(y)))
return x_final, y_final, adv_indices, feat_indices
