activation="relu",
input_shape=x_train.shape[1:]))
model.add(Conv2D(64, (3, 3), activation="relu"))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation="relu"))
model.add(Dropout(0.5))
model.add(Dense(10, activation="softmax"))
model.compile(loss="categorical_crossentropy",
optimizer="adam",
metrics=["accuracy"])
classifier = KerasClassifier(model=model, clip_values=(min_, max_))
classifier.fit(x_train, y_train, nb_epochs=5, batch_size=128)
# Evaluate the classifier on the test set
preds = np.argmax(classifier.predict(x_test), axis=1)
acc = np.sum(preds == np.argmax(y_test, axis=1)) / y_test.shape[0]
print("\nTest accuracy: %.2f%%" % (acc * 100))
# Craft adversarial samples with FGSM
epsilon = 0.1 # Maximum perturbation
adv_crafter = FastGradientMethod(classifier, eps=epsilon)
x_test_adv = adv_crafter.generate(x=x_test)
# Evaluate the classifier on the adversarial examples
preds = np.argmax(classifier.predict(x_test_adv), axis=1)
acc = np.sum(preds == np.argmax(y_test, axis=1)) / y_test.shape[0]
print("\nTest accuracy on adversarial sample: %.2f%%" % (acc * 100))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(100, activation="relu"))
model.add(Dense(10, activation="softmax"))
model.compile(
loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=0.01), metrics=["accuracy"]
)
# Step 3: Create the ART classifier
classifier = KerasClassifier(model=model, clip_values=(min_pixel_value, max_pixel_value), use_logits=False)
# Step 4: Train the ART classifier
classifier.fit(x_train, y_train, batch_size=64, nb_epochs=3)
# 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 = FastGradientMethod(estimator=classifier, eps=0.2)
x_test_adv = attack.generate(x=x_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)
def main():
# Read MNIST dataset (x_raw contains the original images):
(x_raw, y_raw), (x_raw_test, y_raw_test), min_, max_ = load_mnist(raw=True)
n_train = np.shape(x_raw)[0]
num_selection = 5000
random_selection_indices = np.random.choice(n_train, num_selection)
x_raw = x_raw[random_selection_indices]
y_raw = y_raw[random_selection_indices]
# Poison training data
perc_poison = 0.33
(is_poison_train, x_poisoned_raw,
y_poisoned_raw) = generate_backdoor(x_raw, y_raw, perc_poison)
x_train, y_train = preprocess(x_poisoned_raw, y_poisoned_raw)
# Add channel axis:
x_train = np.expand_dims(x_train, axis=3)
# Poison test data
(is_poison_test, x_poisoned_raw_test,
y_poisoned_raw_test) = generate_backdoor(x_raw_test, y_raw_test,
perc_poison)
x_test, y_test = preprocess(x_poisoned_raw_test, y_poisoned_raw_test)
# Add channel axis:
x_test = np.expand_dims(x_test, axis=3)
# Shuffle training data so poison is not together
n_train = np.shape(y_train)[0]
shuffled_indices = np.arange(n_train)
np.random.shuffle(shuffled_indices)
x_train = x_train[shuffled_indices]
y_train = y_train[shuffled_indices]
is_poison_train = is_poison_train[shuffled_indices]
# Create Keras convolutional neural network - basic architecture from Keras examples
# Source here: https://github.com/keras-team/keras/blob/master/examples/mnist_cnn.py
model = Sequential()
model.add(
Conv2D(32,
kernel_size=(3, 3),
activation="relu",
input_shape=x_train.shape[1:]))
model.add(Conv2D(64, (3, 3), activation="relu"))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation="relu"))
model.add(Dropout(0.5))
model.add(Dense(10, activation="softmax"))
model.compile(loss="categorical_crossentropy",
optimizer="adam",
metrics=["accuracy"])
classifier = KerasClassifier(model=model, clip_values=(min_, max_))
classifier.fit(x_train, y_train, nb_epochs=30, batch_size=128)
# Evaluate the classifier on the test set
preds = np.argmax(classifier.predict(x_test), axis=1)
acc = np.sum(preds == np.argmax(y_test, axis=1)) / y_test.shape[0]
print("\nTest accuracy: %.2f%%" % (acc * 100))
# Evaluate the classifier on poisonous data
preds = np.argmax(classifier.predict(x_test[is_poison_test]), axis=1)
acc = np.sum(preds == np.argmax(y_test[is_poison_test],
axis=1)) / y_test[is_poison_test].shape[0]
print(
"\nPoisonous test set accuracy (i.e. effectiveness of poison): %.2f%%"
% (acc * 100))
# Evaluate the classifier on clean data
preds = np.argmax(classifier.predict(x_test[is_poison_test == 0]), axis=1)
acc = np.sum(preds == np.argmax(y_test[
is_poison_test == 0], axis=1)) / y_test[is_poison_test == 0].shape[0]
print("\nClean test set accuracy: %.2f%%" % (acc * 100))
# Calling poisoning defence:
defence = ActivationDefence(classifier, x_train, y_train)
# End-to-end method:
print("------------------- Results using size metric -------------------")
print(defence.get_params())
defence.detect_poison(nb_clusters=2, nb_dims=10, reduce="PCA")
# Evaluate method when ground truth is known:
is_clean = is_poison_train == 0
confusion_matrix = defence.evaluate_defence(is_clean)
print("Evaluation defence results for size-based metric: ")
jsonObject = json.loads(confusion_matrix)
for label in jsonObject:
print(label)
pprint.pprint(jsonObject[label])
# Visualize clusters:
print("Visualize clusters")
sprites_by_class = defence.visualize_clusters(x_train, "mnist_poison_demo")
# Show plots for clusters of class 5
n_class = 5
try:
import matplotlib.pyplot as plt
plt.imshow(sprites_by_class[n_class][0])
plt.title("Class " + str(n_class) + " cluster: 0")
plt.show()
plt.imshow(sprites_by_class[n_class][1])
plt.title("Class " + str(n_class) + " cluster: 1")
plt.show()
except ImportError:
print(
"matplotlib not installed. For this reason, cluster visualization was not displayed"
)
# Try again using distance analysis this time:
print(
"------------------- Results using distance metric -------------------"
)
print(defence.get_params())
defence.detect_poison(nb_clusters=2,
nb_dims=10,
reduce="PCA",
cluster_analysis="distance")
confusion_matrix = defence.evaluate_defence(is_clean)
print("Evaluation defence results for distance-based metric: ")
jsonObject = json.loads(confusion_matrix)
for label in jsonObject:
print(label)
pprint.pprint(jsonObject[label])
# Other ways to invoke the defence:
kwargs = {"nb_clusters": 2, "nb_dims": 10, "reduce": "PCA"}
defence.cluster_activations(**kwargs)
kwargs = {"cluster_analysis": "distance"}
defence.analyze_clusters(**kwargs)
defence.evaluate_defence(is_clean)
kwargs = {"cluster_analysis": "smaller"}
defence.analyze_clusters(**kwargs)
defence.evaluate_defence(is_clean)
print("done :) ")
attack = FeatureCollisionAttack(classifier,
target_instance,
feature_layer,
max_iter=10,
similarity_coeff=256,
watermark=0.3)
poison, poison_labels = attack.poison(base_instances)
poison_pred = np.argmax(classifier.predict(poison), axis=1)
# plt.figure(figsize=(10,10))
# for i in range(0, 9):
# pred_label, true_label = class_descr[poison_pred[i]], class_descr[np.argmax(poison_labels[i])]
# plt.subplot(330 + 1 + i)
# fig=plt.imshow(poison[i])
# fig.axes.get_xaxis().set_visible(False)
# fig.axes.get_yaxis().set_visible(False)
# fig.axes.text(0.5, -0.1, pred_label + " (" + true_label + ")", fontsize=12, transform=fig.axes.transAxes,
# horizontalalignment='center')
classifier.set_learning_phase(True)
print(x_train.shape)
print(base_instances.shape)
adv_train = np.vstack([x_train, poison])
adv_labels = np.vstack([y_train, poison_labels])
classifier.fit(adv_train, adv_labels, nb_epochs=5, batch_size=4)
fig = plt.imshow(target_instance[0])
print('true_class: ' + target_class)
print('predicted_class: ' +
class_descr[np.argmax(classifier.predict(target_instance), axis=1)[0]])
def test_keras(self):
"""
Test with a KerasClassifier.
:return:
"""
# Build KerasClassifier
import tensorflow as tf
tf.compat.v1.disable_eager_execution()
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten, Conv2D
from tensorflow.keras.losses import CategoricalCrossentropy
from tensorflow.keras.optimizers import Adam
model = Sequential()
model.add(
Conv2D(filters=4,
kernel_size=(5, 5),
strides=1,
activation="relu",
input_shape=(28, 28, 1)))
model.add(Flatten())
model.add(Dense(100, activation="relu"))
model.add(Dense(100, activation="relu"))
model.add(Dense(10, activation="linear"))
model.compile(loss=CategoricalCrossentropy(from_logits=True),
optimizer=Adam(learning_rate=0.01),
metrics=["accuracy"])
from art.estimators.classification import KerasClassifier
krc = KerasClassifier(model=model, clip_values=(0, 1))
# Get MNIST
(x_train, y_train), (x_test, y_test) = self.mnist
for i in range(2500):
if np.argmax(y_train[[i]], axis=1) == 0:
y_train[i, :] = 0
y_train[i, 1] = 1
x_train[i, 0:5, 0:5, :] = 1.0
if np.argmax(y_train[[i]], axis=1) == 9:
y_train[i, :] = 0
y_train[i, 9] = 1
x_train[i, 0:5, 0:5, :] = 1.0
for i in range(500):
if np.argmax(y_test[[i]], axis=1) == 0:
y_test[i, :] = 0
y_test[i, 1] = 1
x_test[i, 0:5, 0:5, :] = 1.0
if np.argmax(y_test[[i]], axis=1) == 9:
y_test[i, :] = 0
y_test[i, 9] = 1
x_test[i, 0:5, 0:5, :] = 1.0
krc.fit(x_train, y_train, nb_epochs=3)
cleanse = NeuralCleanse(krc)
defense_cleanse = cleanse(krc, steps=1, patience=1)
defense_cleanse.mitigate(
x_test,
y_test,
mitigation_types=["filtering", "pruning", "unlearning"])
# is_fitted
assert cleanse._is_fitted == cleanse.is_fitted
# get_classifier
assert cleanse.get_classifier
# set_params
cleanse.set_params(**{"batch_size": 1})
assert cleanse.batch_size == 1