def eval_multi(self, inc_epoch=True):
"""
Run the evaluation on multiple attacks.
"""
sess = self.sess
preds = self.preds
x = self.x_pre
y = self.y
X_train = self.X_train
Y_train = self.Y_train
X_test = self.X_test
Y_test = self.Y_test
writer = self.writer
self.summary = tf.Summary()
report = {}
# Evaluate on train set
subsample_factor = 100
X_train_subsampled = X_train[::subsample_factor]
Y_train_subsampled = Y_train[::subsample_factor]
acc_train = model_eval(sess, x, y, preds, X_train_subsampled,
Y_train_subsampled, args=self.eval_params)
self.log_value('train_accuracy_subsampled', acc_train,
'Clean accuracy, subsampled train')
report['train'] = acc_train
# Evaluate on the test set
acc = model_eval(sess, x, y, preds, X_test, Y_test,
args=self.eval_params)
self.log_value('test_accuracy_natural', acc,
'Clean accuracy, natural test')
report['test'] = acc
# Evaluate against adversarial attacks
if self.epoch % self.hparams.eval_iters == 0:
for att_type in self.attack_type_test:
_, preds_adv = self.attacks[att_type]
acc = self.eval_advs(x, y, preds_adv, X_test, Y_test, att_type)
report[att_type] = acc
if self.writer:
writer.add_summary(self.summary, self.epoch)
# Add examples of adversarial examples to the summary
if self.writer and self.epoch % 20 == 0 and self.sum_op is not None:
sm_val = self.sess.run(self.sum_op,
feed_dict={x: X_test[:self.batch_size],
y: Y_test[:self.batch_size]})
if self.writer:
writer.add_summary(sm_val)
self.epoch += 1 if inc_epoch else 0
return report
def evaluate():
# Evaluate the accuracy of the MNIST model on legitimate test examples
eval_params = {'batch_size': batch_size}
acc = model_eval(sess, x, y, preds, x_test, y_test, args=eval_params)
report.clean_train_clean_eval = acc
# assert X_test.shape[0] == test_end - test_start, X_test.shape
print('Test accuracy on legitimate examples: %0.4f' % acc)
def prep_bbox(sess,
x,
y,
x_train,
y_train,
x_test,
y_test,
nb_epochs,
batch_size,
learning_rate,
rng,
nb_classes=10,
img_rows=28,
img_cols=28,
nchannels=1):
"""
Define and train a model that simulates the "remote"
black-box oracle described in the original paper.
:param sess: the TF session
:param x: the input placeholder for MNIST
:param y: the ouput placeholder for MNIST
:param x_train: the training data for the oracle
:param y_train: the training labels for the oracle
:param x_test: the testing data for the oracle
:param y_test: the testing labels for the oracle
:param nb_epochs: number of epochs to train model
:param batch_size: size of training batches
:param learning_rate: learning rate for training
:param rng: numpy.random.RandomState
:return:
"""
# Define TF model graph (for the black-box model)
nb_filters = 64
model = ModelBasicCNN('model1', nb_classes, nb_filters)
loss = CrossEntropy(model, smoothing=0.1)
predictions = model.get_logits(x)
print("Defined TensorFlow model graph.")
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate
}
train(sess, loss, x_train, y_train, args=train_params, rng=rng)
# Print out the accuracy on legitimate data
eval_params = {'batch_size': batch_size}
accuracy = model_eval(sess,
x,
y,
predictions,
x_test,
y_test,
args=eval_params)
print('Test accuracy of black-box on legitimate test '
'examples: ' + str(accuracy))
return model, predictions, accuracy
def evaluate():
acc = model_eval(sess,
x,
y,
preds,
x_test,
y_test,
args={'batch_size': FLAGS.batch_size})
print('Test accuracy on test examples: %0.4f' % acc)
def do_eval(preds, x_set, y_set, report_key, is_adv=None):
acc = model_eval(sess, x, y, preds, x_set, y_set, args=eval_params)
if is_adv is None:
report_text = None
elif is_adv:
report_text = 'adversarial'
else:
report_text = 'legitimate'
if report_text:
print('Test accuracy on %s examples: %0.4f' % (report_text, acc))
def main(argv):
checkpoint = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
if checkpoint is None:
raise ValueError("Couldn't find latest checkpoint in " +
FLAGS.checkpoint_dir)
train_start = 0
train_end = 60000
test_start = 0
test_end = 10000
X_train, Y_train, X_test, Y_test = data_mnist(train_start=train_start,
train_end=train_end,
test_start=test_start,
test_end=test_end)
assert Y_train.shape[1] == 10
# NOTE: for compatibility with Madry Lab downloadable checkpoints,
# we cannot enclose this in a scope or do anything else that would
# change the automatic naming of the variables.
model = MadryMNIST()
x_input = tf.placeholder(tf.float32, shape=[None, 784])
x_image = tf.placeholder(tf.float32, shape=[None, 28, 28, 1])
y = tf.placeholder(tf.float32, shape=[None, 10])
if FLAGS.attack_type == 'fgsm':
fgsm = FastGradientMethod(model)
fgsm_params = {'eps': 0.3, 'clip_min': 0., 'clip_max': 1.}
adv_x = fgsm.generate(x_image, **fgsm_params)
elif FLAGS.attack_type == 'bim':
bim = BasicIterativeMethod(model)
bim_params = {'eps': 0.3, 'clip_min': 0., 'clip_max': 1.,
'nb_iter': 50,
'eps_iter': .01}
adv_x = bim.generate(x_image, **bim_params)
else:
raise ValueError(FLAGS.attack_type)
preds_adv = model.get_probs(adv_x)
saver = tf.train.Saver()
with tf.Session() as sess:
# Restore the checkpoint
saver.restore(sess, checkpoint)
# Evaluate the accuracy of the MNIST model on adversarial examples
eval_par = {'batch_size': FLAGS.batch_size}
t1 = time.time()
acc = model_eval(
sess, x_image, y, preds_adv, X_test, Y_test, args=eval_par)
t2 = time.time()
print("Took", t2 - t1, "seconds")
print('Test accuracy on adversarial examples: %0.4f\n' % acc)
def evaluate_2():
# Accuracy of adversarially trained model on legitimate test inputs
eval_params = {'batch_size': batch_size}
accuracy = model_eval(sess,
x,
y,
preds_2,
x_test,
y_test,
args=eval_params)
print('Test accuracy on legitimate examples: %0.4f' % accuracy)
report.adv_train_clean_eval = accuracy
# Accuracy of the adversarially trained model on adversarial examples
accuracy = model_eval(sess,
x,
y,
preds_2_adv,
x_test,
y_test,
args=eval_params)
print('Test accuracy on adversarial examples: %0.4f' % accuracy)
report.adv_train_adv_eval = accuracy
def do_eval(preds, x_set, y_set, report_key, is_adv=None):
"""
Run the evaluation and print the results.
"""
acc = model_eval(sess, x, y, preds, x_set, y_set, args=eval_params)
setattr(report, report_key, acc)
if is_adv is None:
report_text = None
elif is_adv:
report_text = 'adversarial'
else:
report_text = 'legitimate'
if report_text:
print('Test accuracy on %s examples: %0.4f' % (report_text, acc))
def eval_advs(self, x, y, preds_adv, X_test, Y_test, att_type):
"""
Evaluate the accuracy of the model on adversarial examples
:param x: symbolic input to model.
:param y: symbolic variable for the label.
:param preds_adv: symbolic variable for the prediction on an
adversarial example.
:param X_test: NumPy array of test set inputs.
:param Y_test: NumPy array of test set labels.
:param att_type: name of the attack.
"""
end = (len(X_test) // self.batch_size) * self.batch_size
if self.hparams.fast_tests:
end = 10*self.batch_size
acc = model_eval(self.sess, x, y, preds_adv, X_test[:end],
Y_test[:end], args=self.eval_params)
self.log_value('test_accuracy_%s' % att_type, acc,
'Test accuracy on adversarial examples')
return acc
def do_eval(preds, x_set, y_set, report_key):
acc = model_eval(sess, x, y, preds, x_set, y_set, args=eval_params)
setattr(report, report_key, acc)
print('Test accuracy on legitimate examples: %0.4f' % (acc))
def mnist_tutorial_jsma(train_start=0, train_end=60000, test_start=0,
test_end=10000, viz_enabled=VIZ_ENABLED,
nb_epochs=NB_EPOCHS, batch_size=BATCH_SIZE,
source_samples=SOURCE_SAMPLES,
learning_rate=LEARNING_RATE):
"""
MNIST tutorial for the Jacobian-based saliency map approach (JSMA)
:param train_start: index of first training set example
:param train_end: index of last training set example
:param test_start: index of first test set example
:param test_end: index of last test set example
:param viz_enabled: (boolean) activate plots of adversarial examples
:param nb_epochs: number of epochs to train model
:param batch_size: size of training batches
:param nb_classes: number of output classes
:param source_samples: number of test inputs to attack
:param learning_rate: learning rate for training
:return: an AccuracyReport object
"""
# Object used to keep track of (and return) key accuracies
report = AccuracyReport()
# Set TF random seed to improve reproducibility
tf.set_random_seed(1234)
# Create TF session and set as Keras backend session
sess = tf.Session()
print("Created TensorFlow session.")
set_log_level(logging.DEBUG)
# Get MNIST test data
mnist = MNIST(train_start=train_start, train_end=train_end,
test_start=test_start, test_end=test_end)
x_train, y_train = mnist.get_set('train')
x_test, y_test = mnist.get_set('test')
# Obtain Image Parameters
img_rows, img_cols, nchannels = x_train.shape[1:4]
nb_classes = y_train.shape[1]
# Define input TF placeholder
x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols,
nchannels))
y = tf.placeholder(tf.float32, shape=(None, nb_classes))
nb_filters = 64
# Define TF model graph
model = ModelBasicCNN('model1', nb_classes, nb_filters)
preds = model.get_logits(x)
loss = CrossEntropy(model, smoothing=0.1)
print("Defined TensorFlow model graph.")
###########################################################################
# Training the model using TensorFlow
###########################################################################
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate
}
sess.run(tf.global_variables_initializer())
rng = np.random.RandomState([2017, 8, 30])
train(sess, loss, x_train, y_train, args=train_params, rng=rng)
# Evaluate the accuracy of the MNIST model on legitimate test examples
eval_params = {'batch_size': batch_size}
accuracy = model_eval(sess, x, y, preds, x_test, y_test, args=eval_params)
assert x_test.shape[0] == test_end - test_start, x_test.shape
print('Test accuracy on legitimate test examples: {0}'.format(accuracy))
report.clean_train_clean_eval = accuracy
###########################################################################
# Craft adversarial examples using the Jacobian-based saliency map approach
###########################################################################
print('Crafting ' + str(source_samples) + ' * ' + str(nb_classes - 1) +
' adversarial examples')
# Keep track of success (adversarial example classified in target)
results = np.zeros((nb_classes, source_samples), dtype='i')
# Rate of perturbed features for each test set example and target class
perturbations = np.zeros((nb_classes, source_samples), dtype='f')
# Initialize our array for grid visualization
grid_shape = (nb_classes, nb_classes, img_rows, img_cols, nchannels)
grid_viz_data = np.zeros(grid_shape, dtype='f')
# Instantiate a SaliencyMapMethod attack object
jsma = SaliencyMapMethod(model, sess=sess)
jsma_params = {'theta': 1., 'gamma': 0.1,
'clip_min': 0., 'clip_max': 1.,
'y_target': None}
figure = None
# Loop over the samples we want to perturb into adversarial examples
for sample_ind in xrange(0, source_samples):
print('--------------------------------------')
print('Attacking input %i/%i' % (sample_ind + 1, source_samples))
sample = x_test[sample_ind:(sample_ind + 1)]
# We want to find an adversarial example for each possible target class
# (i.e. all classes that differ from the label given in the dataset)
current_class = int(np.argmax(y_test[sample_ind]))
target_classes = other_classes(nb_classes, current_class)
# For the grid visualization, keep original images along the diagonal
grid_viz_data[current_class, current_class, :, :, :] = np.reshape(
sample, (img_rows, img_cols, nchannels))
# Loop over all target classes
for target in target_classes:
print('Generating adv. example for target class %i' % target)
# This call runs the Jacobian-based saliency map approach
one_hot_target = np.zeros((1, nb_classes), dtype=np.float32)
one_hot_target[0, target] = 1
jsma_params['y_target'] = one_hot_target
adv_x = jsma.generate_np(sample, **jsma_params)
# Check if success was achieved
res = int(model_argmax(sess, x, preds, adv_x) == target)
# Computer number of modified features
adv_x_reshape = adv_x.reshape(-1)
test_in_reshape = x_test[sample_ind].reshape(-1)
nb_changed = np.where(adv_x_reshape != test_in_reshape)[0].shape[0]
percent_perturb = float(nb_changed) / adv_x.reshape(-1).shape[0]
# Display the original and adversarial images side-by-side
if viz_enabled:
figure = pair_visual(
np.reshape(sample, (img_rows, img_cols, nchannels)),
np.reshape(adv_x, (img_rows, img_cols, nchannels)), figure)
# Add our adversarial example to our grid data
grid_viz_data[target, current_class, :, :, :] = np.reshape(
adv_x, (img_rows, img_cols, nchannels))
# Update the arrays for later analysis
results[target, sample_ind] = res
perturbations[target, sample_ind] = percent_perturb
print('--------------------------------------')
# Compute the number of adversarial examples that were successfully found
nb_targets_tried = ((nb_classes - 1) * source_samples)
succ_rate = float(np.sum(results)) / nb_targets_tried
print('Avg. rate of successful adv. examples {0:.4f}'.format(succ_rate))
report.clean_train_adv_eval = 1. - succ_rate
# Compute the average distortion introduced by the algorithm
percent_perturbed = np.mean(perturbations)
print('Avg. rate of perturbed features {0:.4f}'.format(percent_perturbed))
# Compute the average distortion introduced for successful samples only
percent_perturb_succ = np.mean(perturbations * (results == 1))
print('Avg. rate of perturbed features for successful '
'adversarial examples {0:.4f}'.format(percent_perturb_succ))
# Close TF session
sess.close()
# Finally, block & display a grid of all the adversarial examples
if viz_enabled:
import matplotlib.pyplot as plt
plt.close(figure)
_ = grid_visual(grid_viz_data)
return report
def mnist_tutorial(train_start=0,
train_end=60000,
test_start=0,
test_end=10000,
nb_epochs=NB_EPOCHS,
batch_size=BATCH_SIZE,
learning_rate=LEARNING_RATE,
train_dir=TRAIN_DIR,
filename=FILENAME,
load_model=LOAD_MODEL,
testing=False,
label_smoothing=0.1):
"""
MNIST CleverHans tutorial
:param train_start: index of first training set example
:param train_end: index of last training set example
:param test_start: index of first test set example
:param test_end: index of last test set example
:param nb_epochs: number of epochs to train model
:param batch_size: size of training batches
:param learning_rate: learning rate for training
:param train_dir: Directory storing the saved model
:param filename: Filename to save model under
:param load_model: True for load, False for not load
:param testing: if true, test error is calculated
:param label_smoothing: float, amount of label smoothing for cross entropy
:return: an AccuracyReport object
"""
tf.keras.backend.set_learning_phase(0)
# Object used to keep track of (and return) key accuracies
report = AccuracyReport()
# Set TF random seed to improve reproducibility
tf.set_random_seed(1234)
if keras.backend.image_data_format() != 'channels_last':
raise NotImplementedError(
"this tutorial requires keras to be configured to channels_last format"
)
# Create TF session and set as Keras backend session
sess = tf.Session()
keras.backend.set_session(sess)
# Get MNIST test data
mnist = MNIST(train_start=train_start,
train_end=train_end,
test_start=test_start,
test_end=test_end)
x_train, y_train = mnist.get_set('train')
x_test, y_test = mnist.get_set('test')
# Obtain Image Parameters
img_rows, img_cols, nchannels = x_train.shape[1:4]
nb_classes = y_train.shape[1]
# Define input TF placeholder
x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols, nchannels))
y = tf.placeholder(tf.float32, shape=(None, nb_classes))
# Define TF model graph
model = cnn_model(img_rows=img_rows,
img_cols=img_cols,
channels=nchannels,
nb_filters=64,
nb_classes=nb_classes)
preds = model(x)
print("Defined TensorFlow model graph.")
def evaluate():
# Evaluate the accuracy of the MNIST model on legitimate test examples
eval_params = {'batch_size': batch_size}
acc = model_eval(sess, x, y, preds, x_test, y_test, args=eval_params)
report.clean_train_clean_eval = acc
# assert X_test.shape[0] == test_end - test_start, X_test.shape
print('Test accuracy on legitimate examples: %0.4f' % acc)
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate,
'train_dir': train_dir,
'filename': filename
}
rng = np.random.RandomState([2017, 8, 30])
if not os.path.exists(train_dir):
os.mkdir(train_dir)
ckpt = tf.train.get_checkpoint_state(train_dir)
print(train_dir, ckpt)
ckpt_path = False if ckpt is None else ckpt.model_checkpoint_path
wrap = KerasModelWrapper(model)
if load_model and ckpt_path:
saver = tf.train.Saver()
print(ckpt_path)
saver.restore(sess, ckpt_path)
print("Model loaded from: {}".format(ckpt_path))
evaluate()
else:
print("Model was not loaded, training from scratch.")
loss = CrossEntropy(wrap, smoothing=label_smoothing)
train(sess,
loss,
x_train,
y_train,
evaluate=evaluate,
args=train_params,
rng=rng)
# Calculate training error
if testing:
eval_params = {'batch_size': batch_size}
acc = model_eval(sess, x, y, preds, x_train, y_train, args=eval_params)
report.train_clean_train_clean_eval = acc
# Initialize the Fast Gradient Sign Method (FGSM) attack object and graph
fgsm = FastGradientMethod(wrap, sess=sess)
fgsm_params = {'eps': 0.3, 'clip_min': 0., 'clip_max': 1.}
adv_x = fgsm.generate(x, **fgsm_params)
# Consider the attack to be constant
adv_x = tf.stop_gradient(adv_x)
preds_adv = model(adv_x)
# Evaluate the accuracy of the MNIST model on adversarial examples
eval_par = {'batch_size': batch_size}
acc = model_eval(sess, x, y, preds_adv, x_test, y_test, args=eval_par)
print('Test accuracy on adversarial examples: %0.4f\n' % acc)
report.clean_train_adv_eval = acc
# Calculating train error
if testing:
eval_par = {'batch_size': batch_size}
acc = model_eval(sess,
x,
y,
preds_adv,
x_train,
y_train,
args=eval_par)
report.train_clean_train_adv_eval = acc
print("Repeating the process, using adversarial training")
# Redefine TF model graph
model_2 = cnn_model(img_rows=img_rows,
img_cols=img_cols,
channels=nchannels,
nb_filters=64,
nb_classes=nb_classes)
wrap_2 = KerasModelWrapper(model_2)
preds_2 = model_2(x)
fgsm2 = FastGradientMethod(wrap_2, sess=sess)
def attack(x):
return fgsm2.generate(x, **fgsm_params)
preds_2_adv = model_2(attack(x))
loss_2 = CrossEntropy(wrap_2, smoothing=label_smoothing, attack=attack)
def evaluate_2():
# Accuracy of adversarially trained model on legitimate test inputs
eval_params = {'batch_size': batch_size}
accuracy = model_eval(sess,
x,
y,
preds_2,
x_test,
y_test,
args=eval_params)
print('Test accuracy on legitimate examples: %0.4f' % accuracy)
report.adv_train_clean_eval = accuracy
# Accuracy of the adversarially trained model on adversarial examples
accuracy = model_eval(sess,
x,
y,
preds_2_adv,
x_test,
y_test,
args=eval_params)
print('Test accuracy on adversarial examples: %0.4f' % accuracy)
report.adv_train_adv_eval = accuracy
# Perform and evaluate adversarial training
train(sess,
loss_2,
x_train,
y_train,
evaluate=evaluate_2,
args=train_params,
rng=rng)
# Calculate training errors
if testing:
eval_params = {'batch_size': batch_size}
accuracy = model_eval(sess,
x,
y,
preds_2,
x_train,
y_train,
args=eval_params)
report.train_adv_train_clean_eval = accuracy
accuracy = model_eval(sess,
x,
y,
preds_2_adv,
x_train,
y_train,
args=eval_params)
report.train_adv_train_adv_eval = accuracy
return report
def mnist_tutorial_cw(train_start=0,
train_end=60000,
test_start=0,
test_end=10000,
viz_enabled=VIZ_ENABLED,
nb_epochs=NB_EPOCHS,
batch_size=BATCH_SIZE,
source_samples=SOURCE_SAMPLES,
learning_rate=LEARNING_RATE,
attack_iterations=ATTACK_ITERATIONS,
model_path=MODEL_PATH,
targeted=TARGETED):
"""
MNIST tutorial for Carlini and Wagner's attack
:param train_start: index of first training set example
:param train_end: index of last training set example
:param test_start: index of first test set example
:param test_end: index of last test set example
:param viz_enabled: (boolean) activate plots of adversarial examples
:param nb_epochs: number of epochs to train model
:param batch_size: size of training batches
:param nb_classes: number of output classes
:param source_samples: number of test inputs to attack
:param learning_rate: learning rate for training
:param model_path: path to the model file
:param targeted: should we run a targeted attack? or untargeted?
:return: an AccuracyReport object
"""
# Object used to keep track of (and return) key accuracies
report = AccuracyReport()
# Set TF random seed to improve reproducibility
tf.set_random_seed(1234)
# Create TF session
sess = tf.Session()
print("Created TensorFlow session.")
set_log_level(logging.DEBUG)
# Get MNIST test data
mnist = MNIST(train_start=train_start,
train_end=train_end,
test_start=test_start,
test_end=test_end)
x_train, y_train = mnist.get_set('train')
x_test, y_test = mnist.get_set('test')
# Obtain Image Parameters
img_rows, img_cols, nchannels = x_train.shape[1:4]
nb_classes = y_train.shape[1]
# Define input TF placeholder
x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols, nchannels))
y = tf.placeholder(tf.float32, shape=(None, nb_classes))
nb_filters = 64
# Define TF model graph
model = ModelBasicCNN('model1', nb_classes, nb_filters)
preds = model.get_logits(x)
loss = CrossEntropy(model, smoothing=0.1)
print("Defined TensorFlow model graph.")
###########################################################################
# Training the model using TensorFlow
###########################################################################
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate,
'filename': os.path.split(model_path)[-1]
}
rng = np.random.RandomState([2017, 8, 30])
# check if we've trained before, and if we have, use that pre-trained model
if os.path.exists(model_path + ".meta"):
tf_model_load(sess, model_path)
else:
train(sess, loss, x_train, y_train, args=train_params, rng=rng)
saver = tf.train.Saver()
saver.save(sess, model_path)
# Evaluate the accuracy of the MNIST model on legitimate test examples
eval_params = {'batch_size': batch_size}
accuracy = model_eval(sess, x, y, preds, x_test, y_test, args=eval_params)
assert x_test.shape[0] == test_end - test_start, x_test.shape
print('Test accuracy on legitimate test examples: {0}'.format(accuracy))
report.clean_train_clean_eval = accuracy
###########################################################################
# Craft adversarial examples using Carlini and Wagner's approach
###########################################################################
nb_adv_per_sample = str(nb_classes - 1) if targeted else '1'
print('Crafting ' + str(source_samples) + ' * ' + nb_adv_per_sample +
' adversarial examples')
print("This could take some time ...")
# Instantiate a CW attack object
cw = CarliniWagnerL2(model, sess=sess)
if viz_enabled:
assert source_samples == nb_classes
idxs = [
np.where(np.argmax(y_test, axis=1) == i)[0][0]
for i in range(nb_classes)
]
if targeted:
if viz_enabled:
# Initialize our array for grid visualization
grid_shape = (nb_classes, nb_classes, img_rows, img_cols,
nchannels)
grid_viz_data = np.zeros(grid_shape, dtype='f')
adv_inputs = np.array([[instance] * nb_classes
for instance in x_test[idxs]],
dtype=np.float32)
else:
adv_inputs = np.array([[instance] * nb_classes
for instance in x_test[:source_samples]],
dtype=np.float32)
one_hot = np.zeros((nb_classes, nb_classes))
one_hot[np.arange(nb_classes), np.arange(nb_classes)] = 1
adv_inputs = adv_inputs.reshape(
(source_samples * nb_classes, img_rows, img_cols, nchannels))
adv_ys = np.array([one_hot] * source_samples,
dtype=np.float32).reshape(
(source_samples * nb_classes, nb_classes))
yname = "y_target"
else:
if viz_enabled:
# Initialize our array for grid visualization
grid_shape = (nb_classes, 2, img_rows, img_cols, nchannels)
grid_viz_data = np.zeros(grid_shape, dtype='f')
adv_inputs = x_test[idxs]
else:
adv_inputs = x_test[:source_samples]
adv_ys = None
yname = "y"
if targeted:
cw_params_batch_size = source_samples * nb_classes
else:
cw_params_batch_size = source_samples
cw_params = {
'binary_search_steps': 1,
yname: adv_ys,
'max_iterations': attack_iterations,
'learning_rate': CW_LEARNING_RATE,
'batch_size': cw_params_batch_size,
'initial_const': 10
}
adv = cw.generate_np(adv_inputs, **cw_params)
eval_params = {'batch_size': np.minimum(nb_classes, source_samples)}
if targeted:
adv_accuracy = model_eval(sess,
x,
y,
preds,
adv,
adv_ys,
args=eval_params)
else:
if viz_enabled:
err = model_eval(sess,
x,
y,
preds,
adv,
y_test[idxs],
args=eval_params)
adv_accuracy = 1 - err
else:
err = model_eval(sess,
x,
y,
preds,
adv,
y_test[:source_samples],
args=eval_params)
adv_accuracy = 1 - err
if viz_enabled:
for j in range(nb_classes):
if targeted:
for i in range(nb_classes):
grid_viz_data[i, j] = adv[i * nb_classes + j]
else:
grid_viz_data[j, 0] = adv_inputs[j]
grid_viz_data[j, 1] = adv[j]
print(grid_viz_data.shape)
print('--------------------------------------')
# Compute the number of adversarial examples that were successfully found
print('Avg. rate of successful adv. examples {0:.4f}'.format(adv_accuracy))
report.clean_train_adv_eval = 1. - adv_accuracy
# Compute the average distortion introduced by the algorithm
percent_perturbed = np.mean(
np.sum((adv - adv_inputs)**2, axis=(1, 2, 3))**.5)
print('Avg. L_2 norm of perturbations {0:.4f}'.format(percent_perturbed))
# Close TF session
sess.close()
# Finally, block & display a grid of all the adversarial examples
if viz_enabled:
_ = grid_visual(grid_viz_data)
return report
def mnist_blackbox(train_start=0,
train_end=60000,
test_start=0,
test_end=10000,
nb_classes=NB_CLASSES,
batch_size=BATCH_SIZE,
learning_rate=LEARNING_RATE,
nb_epochs=NB_EPOCHS,
holdout=HOLDOUT,
data_aug=DATA_AUG,
nb_epochs_s=NB_EPOCHS_S,
lmbda=LMBDA,
aug_batch_size=AUG_BATCH_SIZE):
"""
MNIST tutorial for the black-box attack from arxiv.org/abs/1602.02697
:param train_start: index of first training set example
:param train_end: index of last training set example
:param test_start: index of first test set example
:param test_end: index of last test set example
:return: a dictionary with:
* black-box model accuracy on test set
* substitute model accuracy on test set
* black-box model accuracy on adversarial examples transferred
from the substitute model
"""
# Set logging level to see debug information
set_log_level(logging.DEBUG)
# Dictionary used to keep track and return key accuracies
accuracies = {}
# Perform tutorial setup
assert setup_tutorial()
# Create TF session
sess = tf.Session()
# Get MNIST data
mnist = MNIST(train_start=train_start,
train_end=train_end,
test_start=test_start,
test_end=test_end)
x_train, y_train = mnist.get_set('train')
x_test, y_test = mnist.get_set('test')
# Initialize substitute training set reserved for adversary
x_sub = x_test[:holdout]
y_sub = np.argmax(y_test[:holdout], axis=1)
# Redefine test set as remaining samples unavailable to adversaries
x_test = x_test[holdout:]
y_test = y_test[holdout:]
# Obtain Image parameters
img_rows, img_cols, nchannels = x_train.shape[1:4]
nb_classes = y_train.shape[1]
# Define input TF placeholder
x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols, nchannels))
y = tf.placeholder(tf.float32, shape=(None, nb_classes))
# Seed random number generator so tutorial is reproducible
rng = np.random.RandomState([2017, 8, 30])
# Simulate the black-box model locally
# You could replace this by a remote labeling API for instance
print("Preparing the black-box model.")
prep_bbox_out = prep_bbox(sess, x, y, x_train, y_train, x_test, y_test,
nb_epochs, batch_size, learning_rate, rng,
nb_classes, img_rows, img_cols, nchannels)
model, bbox_preds, accuracies['bbox'] = prep_bbox_out
# Train substitute using method from https://arxiv.org/abs/1602.02697
print("Training the substitute model.")
train_sub_out = train_sub(sess, x, y, bbox_preds, x_sub, y_sub, nb_classes,
nb_epochs_s, batch_size, learning_rate, data_aug,
lmbda, aug_batch_size, rng, img_rows, img_cols,
nchannels)
model_sub, preds_sub = train_sub_out
# Evaluate the substitute model on clean test examples
eval_params = {'batch_size': batch_size}
acc = model_eval(sess, x, y, preds_sub, x_test, y_test, args=eval_params)
accuracies['sub'] = acc
# Initialize the Fast Gradient Sign Method (FGSM) attack object.
fgsm_par = {'eps': 0.3, 'ord': np.inf, 'clip_min': 0., 'clip_max': 1.}
fgsm = FastGradientMethod(model_sub, sess=sess)
# Craft adversarial examples using the substitute
eval_params = {'batch_size': batch_size}
x_adv_sub = fgsm.generate(x, **fgsm_par)
# Evaluate the accuracy of the "black-box" model on adversarial examples
accuracy = model_eval(sess,
x,
y,
model.get_logits(x_adv_sub),
x_test,
y_test,
args=eval_params)
print('Test accuracy of oracle on adversarial examples generated '
'using the substitute: ' + str(accuracy))
accuracies['bbox_on_sub_adv_ex'] = accuracy
return accuracies
def main(argv):
model_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
if model_file is None:
print('No model found')
sys.exit()
cifar = cifar10_input.CIFAR10Data(FLAGS.dataset_dir)
nb_classes = 10
X_test = cifar.eval_data.xs
Y_test = to_categorical(cifar.eval_data.ys, nb_classes)
assert Y_test.shape[1] == 10.
set_log_level(logging.DEBUG)
with tf.Session() as sess:
x = tf.placeholder(tf.float32, shape=(None, 32, 32, 3))
y = tf.placeholder(tf.float32, shape=(None, 10))
from src.FGSM.cleverhans.cleverhans.model_zoo.madry_lab_challenges.cifar10_model import make_wresnet
model = make_wresnet()
saver = tf.train.Saver()
# Restore the checkpoint
saver.restore(sess, model_file)
nb_samples = FLAGS.nb_samples
attack_params = {
'batch_size': FLAGS.batch_size,
'clip_min': 0.,
'clip_max': 255.
}
if FLAGS.attack_type == 'cwl2':
from src.FGSM.cleverhans.cleverhans.attacks import CarliniWagnerL2
attacker = CarliniWagnerL2(model, sess=sess)
attack_params.update({
'binary_search_steps': 1,
'max_iterations': 100,
'learning_rate': 0.1,
'initial_const': 10,
'batch_size': 10
})
else: # eps and eps_iter in range 0-255
attack_params.update({'eps': 8, 'ord': np.inf})
if FLAGS.attack_type == 'fgsm':
from src.FGSM.cleverhans.cleverhans.attacks import FastGradientMethod
attacker = FastGradientMethod(model, sess=sess)
elif FLAGS.attack_type == 'pgd':
attack_params.update({'eps_iter': 2, 'nb_iter': 20})
from src.FGSM.cleverhans.cleverhans.attacks import MadryEtAl
attacker = MadryEtAl(model, sess=sess)
eval_par = {'batch_size': FLAGS.batch_size}
if FLAGS.sweep:
max_eps = 16
epsilons = np.linspace(1, max_eps, max_eps)
for e in epsilons:
t1 = time.time()
attack_params.update({'eps': e})
x_adv = attacker.generate(x, **attack_params)
preds_adv = model.get_probs(x_adv)
acc = model_eval(sess,
x,
y,
preds_adv,
X_test[:nb_samples],
Y_test[:nb_samples],
args=eval_par)
print('Epsilon %.2f, accuracy on adversarial' % e,
'examples %0.4f\n' % acc)
t2 = time.time()
else:
t1 = time.time()
x_adv = attacker.generate(x, **attack_params)
preds_adv = model.get_probs(x_adv)
acc = model_eval(sess,
x,
y,
preds_adv,
X_test[:nb_samples],
Y_test[:nb_samples],
args=eval_par)
t2 = time.time()
print('Test accuracy on adversarial examples %0.4f\n' % acc)
print("Took", t2 - t1, "seconds")
