def test_scikitlearn(self):
from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC, LinearSVC
from art.classifiers.scikitlearn import ScikitlearnLogisticRegression, ScikitlearnSVC
scikitlearn_test_cases = {
LogisticRegression: ScikitlearnLogisticRegression
} # ,
# SVC: ScikitlearnSVC,
# LinearSVC: ScikitlearnSVC}
(_, _), (x_test, y_test) = self.iris
for (model_class, classifier_class) in scikitlearn_test_cases.items():
model = model_class()
classifier = classifier_class(model=model, clip_values=(0, 1))
classifier.fit(x=x_test, y=y_test)
attack = SaliencyMapMethod(classifier, theta=1, batch_size=128)
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 of ' + classifier.__class__.__name__ +
' on Iris with JSMA adversarial examples: '
'%.2f%%', (acc * 100))
def __init__(self, model, theta=0.1, gamma=1.0, batch_size=16):
super().__init__(model=model)
self._theta = theta
self._gamma = gamma
self._method = SaliencyMapMethod(classifier=self.model,
theta=self._theta,
gamma=self._gamma,
batch_size=batch_size)
def atk_JSMA(x_train, x_test, y_train, y_test, classifier):
#print('Create JSMA attack \n')
adv_crafter = SaliencyMapMethod(classifier, theta=1)
x_train_adv = adv_crafter.generate(x_train)
x_test_adv = adv_crafter.generate(x_test)
print("After JSMA Attack \n")
evaluate(x_train, x_test, y_train, y_test, x_train_adv, x_test_adv, classifier)
return x_test_adv, x_train_adv
def test_keras_iris_vector_unbounded(self):
classifier = get_tabular_classifier_kr()
# Recreate a classifier without clip values
classifier = KerasClassifier(model=classifier._model,
use_logits=False,
channel_index=1)
attack = SaliencyMapMethod(classifier, theta=1)
x_test_iris_adv = attack.generate(self.x_test_iris)
self.assertFalse((self.x_test_iris == x_test_iris_adv).all())
def test_iris_pt(self):
(_, _), (x_test, y_test) = self.iris
classifier = get_iris_classifier_pt()
attack = SaliencyMapMethod(classifier, theta=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 JSMA adversarial examples: %.2f%%',
(acc * 100))
def test_pytorch_iris_vector(self):
classifier = get_tabular_classifier_pt()
attack = SaliencyMapMethod(classifier, theta=1)
x_test_iris_adv = attack.generate(self.x_test_iris)
self.assertFalse((self.x_test_iris == x_test_iris_adv).all())
self.assertTrue((x_test_iris_adv <= 1).all())
self.assertTrue((x_test_iris_adv >= 0).all())
preds_adv = np.argmax(classifier.predict(x_test_iris_adv), axis=1)
self.assertFalse((np.argmax(self.y_test_iris,
axis=1) == preds_adv).all())
accuracy = np.sum(preds_adv == np.argmax(
self.y_test_iris, axis=1)) / self.y_test_iris.shape[0]
logger.info("Accuracy on Iris with JSMA adversarial examples: %.2f%%",
(accuracy * 100))
def test_iris_k_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)
attack = SaliencyMapMethod(classifier, theta=1)
x_test_adv = attack.generate(x_test)
self.assertFalse((x_test == x_test_adv).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 JSMA adversarial examples: %.2f%%',
(acc * 100))
def _test_mnist_untargeted(self, classifier):
# Get MNIST
(_, _), (x_test, y_test) = self.mnist
x_test, y_test = x_test[:NB_TEST], y_test[:NB_TEST]
df = SaliencyMapMethod(classifier, theta=1, batch_size=100)
x_test_adv = df.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertFalse((0. == x_test_adv).all())
y_pred = get_labels_np_array(classifier.predict(x_test_adv))
self.assertFalse((y_test == y_pred).all())
acc = np.sum(np.argmax(y_pred, axis=1) == np.argmax(
y_test, axis=1)) / y_test.shape[0]
logger.info('Accuracy on adversarial examples: %.2f%%', (acc * 100))
def general_test(model,
optimizer,
input_shape,
nb_classes,
test_loader,
method,
btrain=False,
model_file='last_model_92_sgd.pkl'):
global _classes
if not btrain:
model.load_state_dict(torch.load(model_file))
model.eval()
loss = nn.CrossEntropyLoss()
warped_model = PyTorchClassifier(model,
loss,
optimizer,
input_shape,
nb_classes,
clip_values=(.0, 1.))
if method == 'Deepfool':
adv_crafter = DeepFool(warped_model)
elif method == 'BIM':
adv_crafter = BasicIterativeMethod(warped_model, batch_size=20)
elif method == 'JSMA':
adv_crafter = SaliencyMapMethod(warped_model, batch_size=20)
elif method == 'CW2':
adv_crafter = CarliniL2Method(warped_model, batch_size=20)
elif method == 'CWI':
adv_crafter = CarliniLInfMethod(warped_model, batch_size=20)
correct, total = 0, 0
class_correct = list(0. for _ in range(10))
class_total = list(0. for _ in range(10))
for images, labels in test_loader:
images = adv_crafter.generate(images.numpy())
images = Variable(torch.from_numpy(images).cuda())
labels = Variable(labels.cuda())
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels.data).sum()
c = (predicted == labels.data).squeeze()
for i in range(20):
label = labels.data[i]
class_correct[label] += c[i]
class_total[label] += 1
print('Accuracy of the model on the test images: %d %%' %
(100 * float(correct) / total))
print('Accuracy of the model on the test images:', float(correct) / total)
for i in range(10):
print('Accuracy of %5s : %2d %%' %
(_classes[i], 100 * class_correct[i] / class_total[i]))
return correct / total
def general_test_v2(model,
optimizer,
input_shape,
nb_classes,
test_loader,
method,
conf,
btrain=False,
model_file='last_model_92_sgd.pkl'):
global _classes
if not btrain:
checked_state = torch.load(model_file)['state_dict']
model.load_state_dict(checked_state)
model.eval()
loss = nn.CrossEntropyLoss()
warped_model = PyTorchClassifier(model,
loss,
optimizer,
input_shape,
nb_classes,
clip_values=(.0, 1.))
if method == 'Deepfool':
adv_crafter = DeepFool(warped_model)
elif method == 'BIM':
adv_crafter = BasicIterativeMethod(warped_model, batch_size=32)
elif method == 'JSMA':
adv_crafter = SaliencyMapMethod(warped_model, batch_size=32)
elif method == 'CW2':
adv_crafter = CarliniL2Method(warped_model, batch_size=32)
elif method == 'CWI':
adv_crafter = CarliniLInfMethod(warped_model, batch_size=32)
elif method == 'FGSM':
adv_crafter = FastGradientMethod(warped_model, batch_size=32)
correct, total = 0, 0
adv_dataset = adv_generalization(test_loader, adv_crafter, conf)
temp_loader = DataLoader(dataset=adv_dataset,
batch_size=32,
shuffle=False,
drop_last=True)
# temp_loader = test_loader
for images, labels in temp_loader:
images = Variable(images.cuda())
labels = Variable(labels.cuda())
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels.data).sum()
print('Accuracy of the model on the test images: %d %%' %
(100 * float(correct) / total))
print('Accuracy of the model on the test images:', float(correct) / total)
return correct / total
def GetAttackers(classifier, x_test, attacker_name):
"""
Function:
Load classifier and generate adversarial samples
"""
t_start = time.time()
if attacker_name == "SMM":
attacker = SaliencyMapMethod(classifier=classifier, theta=2)
elif attacker_name == "PGD":
attacker = ProjectedGradientDescent(classifier=classifier,
norm=2,
eps=1,
eps_step=0.5)
else:
raise ValueError("Please get the right attacker's name for the input.")
test_adv = attacker.generate(x_test)
dt = time.time() - t_start
return test_adv, dt
def GetAttackers(classifier, x_test, attacker_name):
"""
Function:
Load classifier and generate adversarial samples
"""
t_start = time.time()
if attacker_name == "FGSM":
attacker = FastGradientMethod(classifier=classifier, eps=0.3)
elif attacker_name == "Elastic":
attacker = ElasticNet(classifier=classifier, confidence=0.5)
elif attacker_name == "BasicIterativeMethod":
attacker = BasicIterativeMethod(classifier=classifier, eps=0.3)
elif attacker_name == "NewtonFool":
attacker = NewtonFool(classifier=classifier, max_iter=20)
elif attacker_name == "HopSkipJump":
attacker = HopSkipJump(classifier=classifier, max_iter=20)
elif attacker_name == "ZooAttack":
attacker = ZooAttack(classifier=classifier, max_iter=20)
elif attacker_name == "VirtualAdversarialMethod":
attacker = VirtualAdversarialMethod(classifier=classifier, max_iter=20)
elif attacker_name == "UniversalPerturbation":
attacker = UniversalPerturbation(classifier=classifier, max_iter=20)
elif attacker_name == "AdversarialPatch":
attacker = AdversarialPatch(classifier=classifier, max_iter=20)
elif attacker_name == "Attack":
attacker = Attack(classifier=classifier)
elif attacker_name == "BoundaryAttack":
attacker = BoundaryAttack(classifier=classifier,
targeted=False,
epsilon=0.05,
max_iter=20) #, max_iter=20
elif attacker_name == "CarliniL2":
attacker = CarliniL2Method(classifier=classifier,
confidence=0.5,
learning_rate=0.001,
max_iter=15)
elif attacker_name == "CarliniLinf":
attacker = CarliniLInfMethod(classifier=classifier,
confidence=0.5,
learning_rate=0.001,
max_iter=15)
elif attacker_name == "DeepFool":
attacker = DeepFool(classifier)
elif attacker_name == "SMM":
attacker = SaliencyMapMethod(classifier=classifier, theta=2)
elif attacker_name == "PGD":
attacker = ProjectedGradientDescent(classifier=classifier,
norm=2,
eps=1,
eps_step=0.5)
else:
raise ValueError("Please get the right attacker's name for the input.")
test_adv = attacker.generate(x_test)
dt = time.time() - t_start
return test_adv, dt
def test_scikitlearn(self):
from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC, LinearSVC
from art.classifiers.scikitlearn import SklearnClassifier
scikitlearn_test_cases = [
LogisticRegression(solver="lbfgs", multi_class="auto"),
SVC(gamma="auto"),
LinearSVC(),
]
x_test_original = self.x_test_iris.copy()
for model in scikitlearn_test_cases:
classifier = SklearnClassifier(model=model, clip_values=(0, 1))
classifier.fit(x=self.x_test_iris, y=self.y_test_iris)
attack = SaliencyMapMethod(classifier, theta=1, batch_size=128)
x_test_iris_adv = attack.generate(self.x_test_iris)
self.assertFalse((self.x_test_iris == x_test_iris_adv).all())
self.assertTrue((x_test_iris_adv <= 1).all())
self.assertTrue((x_test_iris_adv >= 0).all())
preds_adv = np.argmax(classifier.predict(x_test_iris_adv), axis=1)
self.assertFalse((np.argmax(self.y_test_iris,
axis=1) == preds_adv).all())
accuracy = np.sum(preds_adv == np.argmax(
self.y_test_iris, axis=1)) / self.y_test_iris.shape[0]
logger.info(
"Accuracy of " + classifier.__class__.__name__ +
" on Iris with JSMA adversarial 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 - self.x_test_iris))),
0.0,
delta=0.00001)
def test_keras_mnist(self):
(x_train, y_train), (x_test, y_test) = self.mnist
x_test_original = x_test.copy()
# Keras classifier
classifier = get_classifier_kr()
scores = classifier._model.evaluate(x_train, y_train)
logger.info('[Keras, MNIST] Accuracy on training set: %.2f%%',
(scores[1] * 100))
scores = classifier._model.evaluate(x_test, y_test)
logger.info('[Keras, MNIST] Accuracy on test set: %.2f%%',
(scores[1] * 100))
# targeted
# Generate random target classes
nb_classes = np.unique(np.argmax(y_test, axis=1)).shape[0]
targets = np.random.randint(nb_classes, size=NB_TEST)
while (targets == np.argmax(y_test, axis=1)).any():
targets = np.random.randint(nb_classes, size=NB_TEST)
# Perform attack
df = SaliencyMapMethod(classifier, theta=1, batch_size=100)
x_test_adv = df.generate(x_test, y=to_categorical(targets, nb_classes))
self.assertFalse((x_test == x_test_adv).all())
self.assertFalse((0. == x_test_adv).all())
y_pred = get_labels_np_array(classifier.predict(x_test_adv))
self.assertFalse((y_test == y_pred).all())
accuracy = np.sum(
np.argmax(y_pred, axis=1) == np.argmax(y_test,
axis=1)) / y_test.shape[0]
logger.info('Accuracy on adversarial examples: %.2f%%',
(accuracy * 100))
# untargeted
df = SaliencyMapMethod(classifier, theta=1, batch_size=100)
x_test_adv = df.generate(x_test)
self.assertFalse((x_test == x_test_adv).all())
self.assertFalse((0. == x_test_adv).all())
y_pred = get_labels_np_array(classifier.predict(x_test_adv))
self.assertFalse((y_test == y_pred).all())
accuracy = np.sum(
np.argmax(y_pred, axis=1) == np.argmax(y_test,
axis=1)) / y_test.shape[0]
logger.info('Accuracy on adversarial 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_scikitlearn(self):
from sklearn.linear_model import LogisticRegression
from art.classifiers.scikitlearn import ScikitlearnLogisticRegression
scikitlearn_test_cases = {
LogisticRegression: ScikitlearnLogisticRegression
} # ,
# SVC: ScikitlearnSVC,
# LinearSVC: ScikitlearnSVC}
(_, _), (x_test, y_test) = self.iris
x_test_original = x_test.copy()
for (model_class, classifier_class) in scikitlearn_test_cases.items():
model = model_class()
classifier = classifier_class(model=model, clip_values=(0, 1))
classifier.fit(x=x_test, y=y_test)
attack = SaliencyMapMethod(classifier, theta=1, batch_size=128)
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())
accuracy = np.sum(
preds_adv == np.argmax(y_test, axis=1)) / y_test.shape[0]
logger.info(
'Accuracy of ' + classifier.__class__.__name__ +
' on Iris with JSMA adversarial 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_classifier_type_check_fail_classifier(self):
# Use a useless test classifier to test basic classifier properties
class ClassifierNoAPI:
pass
classifier = ClassifierNoAPI
with self.assertRaises(TypeError) as context:
_ = SaliencyMapMethod(classifier=classifier)
self.assertIn(
'For `SaliencyMapMethod` classifier must be an instance of '
'`art.classifiers.classifier.Classifier`, the provided classifier is instance of '
'(<class \'object\'>,).', str(context.exception))
def GetAttackers(classifier, x_test, attacker_name):
"""
Function:
Load classifier and generate adversarial samples
"""
t_start = time.time()
if attacker_name == "SMM":
attacker = SaliencyMapMethod(classifier=classifier, theta=.5, gamma=1.)
elif attacker_name == "PGD":
attacker = ProjectedGradientDescent(classifier=classifier,
norm=1,
eps=1,
eps_step=0.5,
max_iter=100,
targeted=False,
num_random_init=0,
batch_size=1)
else:
raise ValueError("Please get the right attacker's name for the input.")
test_adv = attacker.generate(x_test)
dt = time.time() - t_start
return test_adv, dt
class SaliencyMapAttack(AdversarialAttack):
def __init__(self, model, theta=0.1, gamma=1.0, batch_size=16):
super().__init__(model=model)
self._theta = theta
self._gamma = gamma
self._method = SaliencyMapMethod(classifier=self.model,
theta=self._theta,
gamma=self._gamma,
batch_size=batch_size)
def attack_method(self, x, y=None):
params = {}
if y is not None:
params['y'] = y
return self._method.generate(x=x, **params)
def test_classifier_type_check_fail_gradients(self):
# Use a test classifier not providing gradients required by white-box attack
from art.classifiers.scikitlearn import ScikitlearnDecisionTreeClassifier
from sklearn.tree import DecisionTreeClassifier
classifier = ScikitlearnDecisionTreeClassifier(
model=DecisionTreeClassifier())
with self.assertRaises(TypeError) as context:
_ = SaliencyMapMethod(classifier=classifier)
self.assertIn(
'For `SaliencyMapMethod` classifier must be an instance of '
'`art.classifiers.classifier.ClassifierGradients`, the provided classifier is instance of '
'(<class \'art.classifiers.scikitlearn.ScikitlearnClassifier\'>,).',
str(context.exception))
def _test_mnist_targeted(self, classifier):
# Get MNIST
(_, _), (x_test, y_test) = self.mnist
x_test, y_test = x_test[:NB_TEST], y_test[:NB_TEST]
# Generate random target classes
nb_classes = np.unique(np.argmax(y_test, axis=1)).shape[0]
targets = np.random.randint(nb_classes, size=NB_TEST)
while (targets == np.argmax(y_test, axis=1)).any():
targets = np.random.randint(nb_classes, size=NB_TEST)
# Perform attack
df = SaliencyMapMethod(classifier, theta=1, batch_size=100)
x_test_adv = df.generate(x_test, y=to_categorical(targets, nb_classes))
self.assertFalse((x_test == x_test_adv).all())
self.assertFalse((0. == x_test_adv).all())
y_pred = get_labels_np_array(classifier.predict(x_test_adv))
self.assertFalse((y_test == y_pred).all())
acc = np.sum(np.argmax(y_pred, axis=1) == np.argmax(
y_test, axis=1)) / y_test.shape[0]
logger.info('Accuracy on adversarial examples: %.2f%%', (acc * 100))
def build_adversarial(model, optimizer, loss, input_shape, nb_class, method, batch_size=32, pgd_eps=0.3):
model.eval()
wmodel = PyTorchClassifier(model, loss, optimizer, input_shape, nb_class)
if method == 'deepfool':
adv_crafter = DeepFool(wmodel)
elif method == 'bim':
adv_crafter = BasicIterativeMethod(wmodel, batch_size=batch_size)
elif method == 'jsma':
adv_crafter = SaliencyMapMethod(wmodel, batch_size=batch_size)
elif method == 'cw2':
adv_crafter = CarliniL2Method(wmodel, batch_size=batch_size)
elif method == 'cwi':
adv_crafter = CarliniLInfMethod(wmodel, batch_size=batch_size)
elif method == 'fgsm':
adv_crafter = FastGradientMethod(wmodel, batch_size=batch_size)
elif method == 'pgd':
adv_crafter = ProjectedGradientDescent(wmodel, batch_size=batch_size, eps=pgd_eps)
else:
raise NotImplementedError('Unsupported Attack Method: {}'.format(method))
return adv_crafter
eps_step=0.6,
batch_size=64)
if args.attack == 'bim':
if args.d == 'imagenet':
attack = BasicIterativeMethod(classifier=classifier,
eps=0.6,
batch_size=64,
max_iter=25)
else:
attack = BasicIterativeMethod(classifier=classifier,
eps=0.6,
batch_size=64)
if args.attack == 'jsma':
attack = SaliencyMapMethod(classifier=classifier, batch_size=64)
if args.attack == 'c+w':
attack = CarliniL2Method(classifier=classifier, batch_size=64)
# generating adversarial of the testing dataset and save it to the folder './adv'
if args.d == 'mnist' or args.d == 'cifar':
x_adv = attack.generate(x=x_test)
np.save('./adv/{}_{}.npy'.format(args.d, args.attack), x_adv)
if args.d == 'imagenet':
x_adv = attack.generate(x=x_test)
np.save('./adv/{}_{}_{}.npy'.format(args.d, args.attack, args.model),
x_adv)
# accuracy of our test data
pred = np.argmax(classifier.predict(x_test), axis=1)
def JSMA(victims):
adv_crafter = SaliencyMapMethod(model)
finalVictims = adv_crafter.generate(x = victims)
return finalVictims
def saliency_map(classifier, inputs, true_targets, epsilon):
adv_crafter = SaliencyMapMethod(classifier, theta=epsilon, gamma=0.11)
x_test_adv = adv_crafter.generate(x=inputs)
return x_test_adv
x_test, y_test = pickle.load(
open(
'./dataset_imagenet/%s_%s_val_%i.p' %
(args.d, args.model, int(i)), 'rb'))
if args.attack == 'fgsm':
attack = FastGradientMethod(classifier=classifier,
eps=0.6,
eps_step=0.6)
if args.attack == 'bim':
attack = BasicIterativeMethod(classifier=classifier,
eps=0.6,
max_iter=5)
if args.attack == 'jsma':
# attack = ProjectedGradientDescent(classifier=classifier, eps=0.6, max_iter=5)
attack = SaliencyMapMethod(classifier=classifier)
if args.attack == 'c+w':
attack = CarliniL2Method(classifier=classifier)
# attack = CarliniLInfMethod(classifier=classifier, batch_size=1, max_iter=2)
# attack = FastGradientMethod(classifier=classifier)
from datetime import datetime
now = datetime.now()
current_time_after = now.strftime("%H:%M:%S")
print("Current Time After=", current_time_after)
print('Generating adversarial examples----------------')
print(i, x_test.shape, y_test.shape)
x_adv = attack.generate(x=x_test)
print('Saving adversarial examples----------------')
print(x_adv.shape)
pickle.dump(x_adv,
def test_pytorch_mnist(self):
x_train_mnist = np.swapaxes(self.x_train_mnist, 1,
3).astype(np.float32)
x_test_mnist = np.swapaxes(self.x_test_mnist, 1, 3).astype(np.float32)
x_test_original = x_test_mnist.copy()
# Create basic PyTorch model
classifier = get_image_classifier_pt()
scores = get_labels_np_array(classifier.predict(x_train_mnist))
accuracy = np.sum(
np.argmax(scores, axis=1) == np.argmax(self.y_train_mnist,
axis=1)) / self.n_train
logger.info("[PyTorch, MNIST] Accuracy on training set: %.2f%%",
(accuracy * 100))
scores = get_labels_np_array(classifier.predict(x_test_mnist))
accuracy = np.sum(
np.argmax(scores, axis=1) == np.argmax(self.y_test_mnist,
axis=1)) / self.n_test
logger.info("\n[PyTorch, MNIST] Accuracy on test set: %.2f%%",
(accuracy * 100))
# targeted
# Generate random target classes
nb_classes = np.unique(np.argmax(self.y_test_mnist, axis=1)).shape[0]
targets = np.random.randint(nb_classes, size=self.n_test)
while (targets == np.argmax(self.y_test_mnist, axis=1)).any():
targets = np.random.randint(nb_classes, size=self.n_test)
# Perform attack
df = SaliencyMapMethod(classifier, theta=1, batch_size=100)
x_test_mnist_adv = df.generate(x_test_mnist,
y=to_categorical(targets, nb_classes))
self.assertFalse((x_test_mnist == x_test_mnist_adv).all())
self.assertFalse((0.0 == x_test_mnist_adv).all())
y_pred = get_labels_np_array(classifier.predict(x_test_mnist_adv))
self.assertFalse((self.y_test_mnist == y_pred).all())
accuracy = np.sum(
np.argmax(y_pred, axis=1) == np.argmax(self.y_test_mnist,
axis=1)) / self.n_test
logger.info("Accuracy on adversarial examples: %.2f%%",
(accuracy * 100))
# untargeted
df = SaliencyMapMethod(classifier, theta=1, batch_size=100)
x_test_mnist_adv = df.generate(x_test_mnist)
self.assertFalse((x_test_mnist == x_test_mnist_adv).all())
self.assertFalse((0.0 == x_test_mnist_adv).all())
y_pred = get_labels_np_array(classifier.predict(x_test_mnist_adv))
self.assertFalse((self.y_test_mnist == y_pred).all())
accuracy = np.sum(
np.argmax(y_pred, axis=1) == np.argmax(self.y_test_mnist,
axis=1)) / self.n_test
logger.info("Accuracy on adversarial 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_mnist))),
0.0,
delta=0.00001)
