def test_tensorflow_failure_attack_L2(self):
"""
Test the corner case when attack is failed.
:return:
"""
(_, _), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf(from_logits=True)
# Failure attack
cl2m = CarliniL2Method(classifier=tfc,
targeted=True,
max_iter=0,
binary_search_steps=0,
learning_rate=0,
initial_const=1)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = cl2m.generate(x_test, **params)
self.assertLessEqual(np.amax(x_test_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
np.testing.assert_array_almost_equal(x_test, x_test_adv, decimal=3)
# 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)
# Clean-up session
sess.close()
def test_fit_generator(self):
classifier, sess = get_classifier_tf()
# Create TensorFlow data generator
x_tensor = tf.convert_to_tensor(
self.x_train.reshape(10, 100, 28, 28, 1))
y_tensor = tf.convert_to_tensor(self.y_train.reshape(10, 100, 10))
dataset = tf.data.Dataset.from_tensor_slices((x_tensor, y_tensor))
iterator = dataset.make_initializable_iterator()
data_gen = TFDataGenerator(sess=sess,
iterator=iterator,
iterator_type='initializable',
iterator_arg={},
size=1000,
batch_size=100)
# Test fit and predict
classifier.fit_generator(data_gen, nb_epochs=2)
predictions = classifier.predict(self.x_test)
predictions_class = np.argmax(predictions, axis=1)
true_class = np.argmax(self.y_test, axis=1)
accuracy = np.sum(predictions_class == true_class) / len(true_class)
logger.info(
'Accuracy after fitting TF classifier with generator: %.2f%%',
(accuracy * 100))
self.assertEqual(accuracy, 0.65)
tf.reset_default_graph()
sess.close()
def test_loss_gradient(self):
classifier, sess = get_classifier_tf()
gradients = classifier.loss_gradient(self.x_test, self.y_test)
self.assertEqual(gradients.shape, (NB_TEST, 28, 28, 1))
expected_gradients_1 = np.asarray([
0.00279603, 0.00266946, 0.0032446, 0.00396258, -0.00201465,
-0.00564073, 0.0009253, 0.00016253, 0.0040816, 0.00166697,
0.0015883, -0.00121023, -0.00390778, -0.00234937, 0.0053558,
0.00204322, -0.00172054, 0.00053564, -0.0021146, -0.00069308,
0.00141374, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
])
np.testing.assert_array_almost_equal(gradients[0, 14, :, 0],
expected_gradients_1,
decimal=4)
expected_gradients_2 = np.asarray([
1.05401428e-04, 1.06959546e-04, 2.60490313e-04, 2.74000311e-04,
-1.15295035e-04, 2.16038228e-04, 1.37472380e-04, 0.00000000e+00,
0.00000000e+00, -2.91720475e-03, -3.08302144e-04, 2.63109524e-03,
-1.18699251e-03, 2.63655302e-03, 5.35579538e-03, 6.38693338e-03,
3.44644510e-03, 6.68899389e-04, 5.01601025e-03, 8.40547902e-04,
-1.43233046e-05, -2.79442966e-03, 7.37082795e-04, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00, 0.00000000e+00
])
np.testing.assert_array_almost_equal(gradients[0, :, 14, 0],
expected_gradients_2,
decimal=4)
tf.reset_default_graph()
sess.close()
def test_pickle(self):
classifier, sess = get_classifier_tf()
full_path = os.path.join(ART_DATA_PATH, 'my_classifier')
folder = os.path.split(full_path)[0]
if not os.path.exists(folder):
os.makedirs(folder)
pickle.dump(classifier, open(full_path, 'wb'))
# Unpickle:
with open(full_path, 'rb') as f:
loaded = pickle.load(f)
self.assertEqual(classifier._clip_values, loaded._clip_values)
self.assertEqual(classifier._channel_index, loaded._channel_index)
self.assertEqual(set(classifier.__dict__.keys()),
set(loaded.__dict__.keys()))
# Test predict
predictions_1 = classifier.predict(self.x_test)
accuracy_1 = np.sum(
np.argmax(predictions_1, axis=1) == np.argmax(
self.y_test, axis=1)) / len(self.y_test)
predictions_2 = loaded.predict(self.x_test)
accuracy_2 = np.sum(
np.argmax(predictions_2, axis=1) == np.argmax(
self.y_test, axis=1)) / len(self.y_test)
self.assertEqual(accuracy_1, accuracy_2)
tf.reset_default_graph()
sess.close()
def test_tensorflow_mnist(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
(x_train, y_train), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf()
# Attack
up = UniversalPerturbation(tfc, max_iter=1, attacker="newtonfool", attacker_params={"max_iter": 5})
x_train_adv = up.generate(x_train)
self.assertTrue((up.fooling_rate >= 0.2) or not up.converged)
x_test_adv = x_test + up.noise
self.assertFalse((x_test == x_test_adv).all())
train_y_pred = np.argmax(tfc.predict(x_train_adv), axis=1)
test_y_pred = np.argmax(tfc.predict(x_test_adv), axis=1)
self.assertFalse((np.argmax(y_test, axis=1) == test_y_pred).all())
self.assertFalse((np.argmax(y_train, axis=1) == train_y_pred).all())
# 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_tensorflow_mnist(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
(_, _), (x_test, _) = self.mnist
x_test_original = x_test.copy()
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf()
# Attack
nf = NewtonFool(tfc, max_iter=5, batch_size=100)
x_test_adv = nf.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
y_pred = tfc.predict(x_test)
y_pred_adv = tfc.predict(x_test_adv)
y_pred_bool = y_pred.max(axis=1, keepdims=1) == y_pred
y_pred_max = y_pred.max(axis=1)
y_pred_adv_max = y_pred_adv[y_pred_bool]
self.assertTrue((y_pred_max >= .9 * y_pred_adv_max).all())
# 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_save(self):
classifier, sess = get_classifier_tf()
path = 'tmp'
filename = 'model.ckpt'
# Save
classifier.save(filename, path=path)
self.assertTrue(
os.path.isfile(
os.path.join(path, filename,
'variables/variables.data-00000-of-00001')))
self.assertTrue(
os.path.isfile(
os.path.join(path, filename, 'variables/variables.index')))
self.assertTrue(
os.path.isfile(os.path.join(path, filename, 'saved_model.pb')))
# # Restore
# with tf.Session(graph=tf.Graph()) as sess:
# tf.saved_model.loader.load(sess, ["serve"], os.path.join(path, filename))
# graph = tf.get_default_graph()
# print(graph.get_operations())
# sess.run('SavedOutput:0', feed_dict={'SavedInputPhD:0': input_batch})
# Remove saved files
shutil.rmtree(os.path.join(path, filename))
tf.reset_default_graph()
sess.close()
def test_tfclassifier(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
# Build TensorFlowClassifiers
victim_tfc, sess = get_classifier_tf()
# Define input and output placeholders
input_ph = tf.placeholder(tf.float32, shape=[None, 28, 28, 1])
output_ph = tf.placeholder(tf.int32, shape=[None, 10])
# Define the tensorflow graph
conv = tf.layers.conv2d(input_ph, 1, 7, activation=tf.nn.relu)
conv = tf.layers.max_pooling2d(conv, 4, 4)
flattened = tf.layers.flatten(conv)
# Logits layer
logits = tf.layers.dense(flattened, 10)
# Train operator
loss = tf.reduce_mean(
tf.losses.softmax_cross_entropy(logits=logits,
onehot_labels=output_ph))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001)
train = optimizer.minimize(loss)
# TensorFlow session and initialization
sess.run(tf.global_variables_initializer())
# Create the classifier
thieved_tfc = TensorFlowClassifier(clip_values=(0, 1),
input_ph=input_ph,
output=logits,
labels_ph=output_ph,
train=train,
loss=loss,
learning=None,
sess=sess)
# Create attack
copycat_cnn = CopycatCNN(classifier=victim_tfc,
batch_size_query=BATCH_SIZE,
batch_size_fit=BATCH_SIZE,
nb_epochs=NB_EPOCHS,
nb_stolen=NB_STOLEN)
thieved_tfc = copycat_cnn.extract(x=self.x_train,
thieved_classifier=thieved_tfc)
victim_preds = np.argmax(victim_tfc.predict(x=self.x_train[:100]),
axis=1)
thieved_preds = np.argmax(thieved_tfc.predict(x=self.x_train[:100]),
axis=1)
acc = np.sum(victim_preds == thieved_preds) / len(victim_preds)
self.assertGreater(acc, 0.3)
# Clean-up session
sess.close()
tf.reset_default_graph()
def test_repr(self):
classifier, sess = get_classifier_tf()
repr_ = repr(classifier)
self.assertIn('art.classifiers.tensorflow.TensorFlowClassifier', repr_)
self.assertIn('channel_index=3, clip_values=(0, 1)', repr_)
self.assertIn('defences=None, preprocessing=(0, 1)', repr_)
tf.reset_default_graph()
sess.close()
def test_set_learning(self):
classifier, sess = get_classifier_tf()
self.assertEqual(classifier._feed_dict, {})
classifier.set_learning_phase(False)
self.assertFalse(classifier._feed_dict[classifier._learning])
classifier.set_learning_phase(True)
self.assertTrue(classifier._feed_dict[classifier._learning])
self.assertTrue(classifier.learning_phase)
tf.reset_default_graph()
sess.close()
def test_predict(self):
classifier, sess = get_classifier_tf()
predictions = classifier.predict(self.x_test)
predictions_class = np.argmax(predictions, axis=1)
trues_class = np.argmax(self.y_test, axis=1)
accuracy = np.sum(predictions_class == trues_class) / len(trues_class)
logger.info('Accuracy after fitting: %.2f%%', (accuracy * 100))
self.assertEqual(accuracy, 0.4)
tf.reset_default_graph()
sess.close()
def test_tensorflow_mnist(self):
(x_train, y_train), (x_test, y_test) = self.mnist
classifier, sess = get_classifier_tf(from_logits=True)
scores = get_labels_np_array(classifier.predict(x_train))
acc = np.sum(np.argmax(scores, axis=1) == np.argmax(y_train, axis=1)) / y_train.shape[0]
logger.info('[TF, MNIST] Accuracy on training set: %.2f%%', (acc * 100))
scores = get_labels_np_array(classifier.predict(x_test))
acc = np.sum(np.argmax(scores, axis=1) == np.argmax(y_test, axis=1)) / y_test.shape[0]
logger.info('[TF, MNIST] Accuracy on test set: %.2f%%', (acc * 100))
self._test_backend_mnist(classifier, x_test, y_test)
def test_tensorflow_mnist(self):
(x_train, y_train), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
# Create basic CNN on MNIST using TensorFlow
classifier, sess = get_classifier_tf(from_logits=True)
scores = get_labels_np_array(classifier.predict(x_train))
accuracy = np.sum(
np.argmax(scores, axis=1) == np.argmax(y_train,
axis=1)) / y_train.shape[0]
logger.info('[TF, MNIST] Accuracy on training set: %.2f%%',
(accuracy * 100))
scores = get_labels_np_array(classifier.predict(x_test))
accuracy = np.sum(
np.argmax(scores, axis=1) == np.argmax(y_test,
axis=1)) / y_test.shape[0]
logger.info('[TF, MNIST] Accuracy on test set: %.2f%%',
(accuracy * 100))
attack = DeepFool(classifier, max_iter=5, batch_size=11)
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())
accuracy = np.sum(
np.argmax(train_y_pred, axis=1) == np.argmax(
y_train, axis=1)) / y_train.shape[0]
logger.info('Accuracy on adversarial train examples: %.2f%%',
(accuracy * 100))
accuracy = np.sum(
np.argmax(test_y_pred, axis=1) == np.argmax(
y_test, axis=1)) / y_test.shape[0]
logger.info('Accuracy on adversarial test examples: %.2f%%',
(accuracy * 100))
# 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_tensorflow_mnist_L2(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
(_, _), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf()
# First attack
cl2m = CarliniL2Method(classifier=tfc, targeted=True, max_iter=10)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = cl2m.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(tfc.predict(x_test_adv), axis=1)
logger.debug('CW2 Target: %s', target)
logger.debug('CW2 Actual: %s', y_pred_adv)
logger.info('CW2 Success Rate: %.2f',
(np.sum(target == y_pred_adv) / float(len(target))))
self.assertTrue((target == y_pred_adv).any())
# Second attack, no batching
cl2m = CarliniL2Method(classifier=tfc,
targeted=False,
max_iter=10,
batch_size=1)
x_test_adv = cl2m.generate(x_test)
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('CW2 Target: %s', target)
logger.debug('CW2 Actual: %s', y_pred_adv)
logger.info('CW2 Success Rate: %.2f',
(np.sum(target == y_pred_adv) / float(len(target))))
self.assertTrue((target != y_pred_adv).any())
# 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)
# Clean-up session
sess.close()
def test_layers(self):
classifier, sess = get_classifier_tf()
layer_names = classifier.layer_names
for i, name in enumerate(layer_names):
print(self.x_test.shape)
activation_i = classifier.get_activations(self.x_test,
i,
batch_size=5)
activation_name = classifier.get_activations(self.x_test,
name,
batch_size=5)
np.testing.assert_array_equal(activation_name, activation_i)
tf.reset_default_graph()
sess.close()
def test_tensorflow_mnist_LInf(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
(_, _), (x_test, y_test) = self.mnist
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf(from_logits=True)
# First attack
clinfm = CarliniLInfMethod(classifier=tfc,
targeted=True,
max_iter=10,
eps=0.5)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = clinfm.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(tfc.predict(x_test_adv), axis=1)
logger.debug('CW0 Target: %s', target)
logger.debug('CW0 Actual: %s', y_pred_adv)
logger.info('CW0 Success Rate: %.2f',
(np.sum(target == y_pred_adv) / float(len(target))))
self.assertTrue((target == y_pred_adv).any())
# Second attack, no batching
clinfm = CarliniLInfMethod(classifier=tfc,
targeted=False,
max_iter=10,
eps=0.5,
batch_size=1)
x_test_adv = clinfm.generate(x_test)
self.assertLessEqual(np.amax(x_test_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_adv), -1e-6)
target = np.argmax(params['y'], axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_adv), axis=1)
logger.debug('CW0 Target: %s', target)
logger.debug('CW0 Actual: %s', y_pred_adv)
logger.info('CW0 Success Rate: %.2f',
(np.sum(target != y_pred_adv) / float(len(target))))
self.assertTrue((target != y_pred_adv).any())
# Clean-up session
sess.close()
def test_tensorflow_failure_attack(self):
"""
Test the corner case when attack fails.
:return:
"""
(_, _), (x_test, y_test) = self.mnist
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf()
# Failure attack
ead = ElasticNet(classifier=tfc, targeted=True, max_iter=0, binary_search_steps=0, learning_rate=0,
initial_const=1)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = ead.generate(x_test, **params)
self.assertLessEqual(np.amax(x_test_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
np.testing.assert_almost_equal(x_test, x_test_adv, 3)
# Clean-up session
sess.close()
def test_tfclassifier(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
x_test_original = self.x_test.copy()
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf()
# Attack
attack_st = SpatialTransformation(tfc,
max_translation=10.0,
num_translations=3,
max_rotation=30.0,
num_rotations=3)
x_train_adv = attack_st.generate(self.x_train)
self.assertAlmostEqual(x_train_adv[0, 8, 13, 0],
0.49004024,
delta=0.01)
self.assertAlmostEqual(attack_st.fooling_rate, 0.72, delta=0.01)
self.assertEqual(attack_st.attack_trans_x, 3)
self.assertEqual(attack_st.attack_trans_y, 3)
self.assertEqual(attack_st.attack_rot, 30.0)
x_test_adv = attack_st.generate(self.x_test)
self.assertAlmostEqual(x_test_adv[0, 14, 14, 0],
0.013572651,
delta=0.01)
# 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)
sess.close()
def test_tensorflow_mnist(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
x_test_original = self.x_test.copy()
# 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))))
# 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 session
sess.close()
def test_tensorflow_failure_attack(self):
"""
Test the corner case when attack fails.
:return:
"""
x_test_original = self.x_test.copy()
# 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)
# 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 session
sess.close()
def test_tensorflow(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
tfc, sess = get_classifier_tf()
attack_ap = AdversarialPatch(tfc,
rotation_max=22.5,
scale_min=0.1,
scale_max=1.0,
learning_rate=5.0,
batch_size=10,
max_iter=500)
patch_adv, _ = attack_ap.generate(self.x_train)
self.assertAlmostEqual(patch_adv[8, 8, 0],
-3.1106631027725005,
delta=0.1)
self.assertAlmostEqual(patch_adv[14, 14, 0], 18.101, delta=0.1)
self.assertAlmostEqual(float(np.sum(patch_adv)), 624.867, delta=0.1)
sess.close()
def test_tensorflow_failure_attack_LInf(self):
"""
Test the corner case when attack is failed.
:return:
"""
(_, _), (x_test, y_test) = self.mnist
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf(from_logits=True)
# Failure attack
clinfm = CarliniLInfMethod(classifier=tfc,
targeted=True,
max_iter=0,
learning_rate=0,
eps=0.5)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = clinfm.generate(x_test, **params)
self.assertLessEqual(np.amax(x_test_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
self.assertTrue(np.allclose(x_test, x_test_adv, atol=1e-3))
# Clean-up session
sess.close()
def test_class_gradient(self):
classifier, sess = get_classifier_tf(from_logits=True)
# Test all gradients label = None
gradients = classifier.class_gradient(self.x_test)
self.assertEqual(gradients.shape, (NB_TEST, 10, 28, 28, 1))
expected_gradients_1 = np.asarray([
-0.03347399, -0.03195872, -0.02650188, 0.04111874, 0.08676253,
0.03339913, 0.06925241, 0.09387045, 0.15184258, -0.00684002,
0.05070481, 0.01409407, -0.03632583, 0.00151133, 0.05102589,
0.00766463, -0.00898967, 0.00232938, -0.00617045, -0.00201032,
0.00410065, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
])
np.testing.assert_array_almost_equal(gradients[0, 5, 14, :, 0],
expected_gradients_1,
decimal=4)
expected_gradients_2 = np.asarray([
-0.09723657, -0.00240533, 0.02445251, -0.00035474, 0.04765627,
0.04286841, 0.07209076, 0.0, 0.0, -0.07938144, -0.00142567,
0.02882954, -0.00049514, 0.04170151, 0.05102589, 0.09544909,
-0.04401167, -0.06158172, 0.03359772, -0.00838454, 0.01722163,
-0.13376027, 0.08206709, 0.0, 0.0, 0.0, 0.0, 0.0
])
np.testing.assert_array_almost_equal(gradients[0, 5, :, 14, 0],
expected_gradients_2,
decimal=4)
# Test 1 gradient label = 5
gradients = classifier.class_gradient(self.x_test, label=5)
self.assertEqual(gradients.shape, (NB_TEST, 1, 28, 28, 1))
expected_gradients_1 = np.asarray([
-0.03347399, -0.03195872, -0.02650188, 0.04111874, 0.08676253,
0.03339913, 0.06925241, 0.09387045, 0.15184258, -0.00684002,
0.05070481, 0.01409407, -0.03632583, 0.00151133, 0.05102589,
0.00766463, -0.00898967, 0.00232938, -0.00617045, -0.00201032,
0.00410065, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
])
np.testing.assert_array_almost_equal(gradients[0, 0, 14, :, 0],
expected_gradients_1,
decimal=4)
expected_gradients_2 = np.asarray([
-0.09723657, -0.00240533, 0.02445251, -0.00035474, 0.04765627,
0.04286841, 0.07209076, 0.0, 0.0, -0.07938144, -0.00142567,
0.02882954, -0.00049514, 0.04170151, 0.05102589, 0.09544909,
-0.04401167, -0.06158172, 0.03359772, -0.00838454, 0.01722163,
-0.13376027, 0.08206709, 0.0, 0.0, 0.0, 0.0, 0.0
])
np.testing.assert_array_almost_equal(gradients[0, 0, :, 14, 0],
expected_gradients_2,
decimal=4)
# Test a set of gradients label = array
label = np.random.randint(5, size=NB_TEST)
gradients = classifier.class_gradient(self.x_test, label=label)
self.assertEqual(gradients.shape, (NB_TEST, 1, 28, 28, 1))
expected_gradients_1 = np.asarray([
0.06860766, 0.065502, 0.08539103, 0.13868105, -0.05520725,
-0.18788849, 0.02264893, 0.02980516, 0.2226511, 0.11288887,
-0.00678776, 0.02045561, -0.03120914, 0.00642691, 0.08449504,
0.02848018, -0.03251382, 0.00854315, -0.02354656, -0.00767687,
0.01565931, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
])
np.testing.assert_array_almost_equal(gradients[0, 0, 14, :, 0],
expected_gradients_1,
decimal=4)
expected_gradients_2 = np.asarray([
-0.0487146, -0.0171556, -0.03161772, -0.0420007, 0.03360246,
-0.01864819, 0.00315916, 0.0, 0.0, -0.07631349, -0.00374462,
0.04229517, -0.01131879, 0.05044588, 0.08449504, 0.12417868,
0.07536847, 0.03906382, 0.09467953, 0.00543209, -0.00504872,
-0.03366479, -0.00385999, 0.0, 0.0, 0.0, 0.0, 0.0
])
np.testing.assert_array_almost_equal(gradients[0, 0, :, 14, 0],
expected_gradients_2,
decimal=4)
tf.reset_default_graph()
sess.close()
def test_tensorflow_mnist(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
(_, _), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf(from_logits=True)
# First attack
ead = ElasticNet(classifier=tfc, targeted=True, max_iter=2)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = ead.generate(x_test, **params)
expected_x_test_adv = np.asarray([0.45704955, 0.43627003, 0.57238287, 1.0, 0.11541145, 0.12619308,
0.48318917, 0.3457903, 0.17863746, 0.09060935, 0.0, 0.00963121,
0.0, 0.04749763, 0.4058206, 0.17860745, 0.0, 0.9153206,
0.84564775, 0.20603634, 0.10586322, 0.00947509, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0])
np.testing.assert_array_almost_equal(x_test_adv[0, 14, :, 0], expected_x_test_adv, decimal=6)
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('EAD target: %s', target)
logger.debug('EAD actual: %s', y_pred_adv)
logger.info('EAD success rate on MNIST: %.2f%%', (100 * sum(target == y_pred_adv) / len(target)))
self.assertTrue((target == y_pred_adv).any())
# Second attack
ead = ElasticNet(classifier=tfc, targeted=False, max_iter=2)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = ead.generate(x_test, **params)
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('EAD target: %s', target)
logger.debug('EAD actual: %s', y_pred_adv)
logger.info('EAD success rate on MNIST: %.2f%%', (100 * sum(target != y_pred_adv) / float(len(target))))
np.testing.assert_array_equal(y_pred_adv, np.asarray([7, 1, 1, 4, 4, 1, 4, 4, 4, 4]))
# Third attack
ead = ElasticNet(classifier=tfc, targeted=False, max_iter=2)
params = {}
x_test_adv = ead.generate(x_test, **params)
expected_x_test_adv = np.asarray([0.22866514, 0.21826893, 0.22902338, 0.06268515, 0.0, 0.0,
0.04822975, 0.0, 0.0, 0.0, 0.05555382, 0.0,
0.0, 0.0, 0.38986346, 0.10653087, 0.32385707, 0.98043066,
0.75790393, 0.16486718, 0.16069527, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0])
np.testing.assert_array_almost_equal(x_test_adv[0, 14, :, 0], expected_x_test_adv, decimal=6)
self.assertLessEqual(np.amax(x_test_adv), 1.0)
self.assertGreaterEqual(np.amin(x_test_adv), 0.0)
y_pred = np.argmax(tfc.predict(x_test), axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_adv), axis=1)
logger.debug('EAD target: %s', y_pred)
logger.debug('EAD actual: %s', y_pred_adv)
logger.info('EAD success rate: %.2f%%', (100 * sum(y_pred != y_pred_adv) / float(len(y_pred))))
np.testing.assert_array_equal(y_pred_adv, np.asarray([0, 4, 7, 9, 0, 7, 7, 3, 0, 7]))
# First attack without batching
ead_wob = ElasticNet(classifier=tfc, targeted=True, max_iter=2, batch_size=1)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = ead_wob.generate(x_test, **params)
expected_x_test_adv = np.asarray([0.3287169, 0.31374657, 0.42853343, 0.8994576, 0.19850709, 0.11997936,
0.5622535, 0.43854535, 0.19387433, 0.12516324, 0.0, 0.10933565,
0.02162433, 0.07120894, 0.95224255, 0.3072921, 0.48966524, 1.,
0.3814998, 0.15782641, 0.52283823, 0.12852049, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0])
np.testing.assert_array_almost_equal(x_test_adv[0, 14, :, 0], expected_x_test_adv, decimal=6)
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('EAD target: %s', target)
logger.debug('EAD actual: %s', y_pred_adv)
logger.info('EAD success rate: %.2f%%', (100 * sum(target == y_pred_adv) / float(len(target))))
self.assertTrue((target == y_pred_adv).any())
# Second attack without batching
ead_wob = ElasticNet(classifier=tfc, targeted=False, max_iter=2, batch_size=1)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = ead_wob.generate(x_test, **params)
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('EAD target: %s', target)
logger.debug('EAD actual: %s', y_pred_adv)
logger.info('EAD success rate: %.2f%%', (100 * sum(target != y_pred_adv) / float(len(target))))
np.testing.assert_array_equal(y_pred_adv, np.asarray([7, 1, 1, 4, 4, 1, 4, 4, 4, 4]))
# 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)
# Close session
sess.close()
def test_tensorflow_mnist(self):
(_, _), (x_test, y_test) = self.mnist
classifier, sess = get_classifier_tf()
self._test_backend_mnist(classifier, x_test, y_test)
def test_tensorflow_mnist_targeted(self):
(_, _), (x_test, _) = self.mnist
classifier, sess = get_classifier_tf()
self._test_mnist_targeted(classifier, x_test)
def test_nb_classes(self):
classifier, sess = get_classifier_tf()
self.assertEqual(classifier.nb_classes(), 10)
tf.reset_default_graph()
sess.close()
def test_input_shape(self):
classifier, sess = get_classifier_tf()
self.assertEqual(classifier.input_shape, (28, 28, 1))
tf.reset_default_graph()
sess.close()
def tearDownClass(cls):
_, sess = get_classifier_tf()
tf.reset_default_graph()
sess.close()
def test_tensorflow_mnist(self):
"""
First test with the TensorFlowClassifier.
:return:
"""
(_, _), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
# Build TensorFlowClassifier
tfc, sess = get_classifier_tf()
# First targeted attack and norm=2
hsj = HopSkipJump(classifier=tfc, targeted=True, max_iter=2, max_eval=100, init_eval=10)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = hsj.generate(x_test, **params)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv <= 1.0001).all())
self.assertTrue((x_test_adv >= -0.0001).all())
target = np.argmax(params['y'], axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_adv), axis=1)
self.assertTrue((target == y_pred_adv).any())
# First targeted attack and norm=np.inf
hsj = HopSkipJump(classifier=tfc, targeted=True, max_iter=2, max_eval=100, init_eval=10, norm=np.Inf)
params = {'y': random_targets(y_test, tfc.nb_classes())}
x_test_adv = hsj.generate(x_test, **params)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv <= 1.0001).all())
self.assertTrue((x_test_adv >= -0.0001).all())
target = np.argmax(params['y'], axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_adv), axis=1)
self.assertTrue((target == y_pred_adv).any())
# Second untargeted attack and norm=2
hsj = HopSkipJump(classifier=tfc, targeted=False, max_iter=2, max_eval=100, init_eval=10)
x_test_adv = hsj.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv <= 1.0001).all())
self.assertTrue((x_test_adv >= -0.0001).all())
y_pred = np.argmax(tfc.predict(x_test), axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_adv), axis=1)
self.assertTrue((y_pred != y_pred_adv).any())
# Second untargeted attack and norm=np.inf
hsj = HopSkipJump(classifier=tfc, targeted=False, max_iter=2, max_eval=100, init_eval=10, norm=np.Inf)
x_test_adv = hsj.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertTrue((x_test_adv <= 1.0001).all())
self.assertTrue((x_test_adv >= -0.0001).all())
y_pred = np.argmax(tfc.predict(x_test), axis=1)
y_pred_adv = np.argmax(tfc.predict(x_test_adv), axis=1)
self.assertTrue((y_pred != y_pred_adv).any())
# 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)
# Clean-up session
sess.close()
