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,
clean_train=CLEAN_TRAIN,
testing=False,
backprop_through_attack=BACKPROP_THROUGH_ATTACK,
nb_filters=64,
num_threads=None,
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 clean_train: perform normal training on clean examples only
before performing adversarial training.
:param testing: if true, complete an AccuracyReport for unit tests
to verify that performance is adequate
:param backprop_through_attack: If True, backprop through adversarial
example construction process during
adversarial training.
:param label_smoothing: float, amount of label smoothing for cross entropy
: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)
# Set logging level to see debug information
set_log_level(logging.DEBUG)
# Create TF session
if num_threads:
config_args = dict(intra_op_parallelism_threads=1)
else:
config_args = {}
sess = tf.Session(config=tf.ConfigProto(**config_args))
# Get MNIST test data
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)
# Use 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))
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate
}
eval_params = {'batch_size': batch_size}
fgsm_params = {'eps': 0.3, 'clip_min': 0., 'clip_max': 1.}
rng = np.random.RandomState([2017, 8, 30])
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)
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))
if clean_train:
model = ModelBasicCNN('model1', nb_classes, nb_filters)
preds = model.get_logits(x)
loss = CrossEntropy(model, smoothing=label_smoothing)
def evaluate():
do_eval(preds, x_test, y_test, 'clean_train_clean_eval', False)
train(sess,
loss,
x,
y,
x_train,
y_train,
evaluate=evaluate,
args=train_params,
rng=rng,
var_list=model.get_params())
# Calculate training error
if testing:
do_eval(preds, x_train, y_train, 'train_clean_train_clean_eval')
# Initialize the Fast Gradient Sign Method (FGSM) attack object and
# graph
fgsm = FastGradientMethod(model, sess=sess)
adv_x = fgsm.generate(x, **fgsm_params)
preds_adv = model.get_logits(adv_x)
# Evaluate the accuracy of the MNIST model on adversarial examples
do_eval(preds_adv, x_test, y_test, 'clean_train_adv_eval', True)
# Calculate training error
if testing:
do_eval(preds_adv, x_train, y_train, 'train_clean_train_adv_eval')
print('Repeating the process, using adversarial training')
print("after clean train>>>>>>>>>>>>>>>>>>>>>")
# Create a new model and train it to be robust to FastGradientMethod
model2 = ModelBasicCNN('model2', nb_classes, nb_filters)
fgsm2 = FastGradientMethod(model2, sess=sess)
def attack(x):
return fgsm2.generate(x, **fgsm_params)
loss2 = CrossEntropy(model2, smoothing=label_smoothing, attack=attack)
preds2 = model2.get_logits(x)
adv_x2 = attack(x)
if not backprop_through_attack:
# For the fgsm attack used in this tutorial, the attack has zero
# gradient so enabling this flag does not change the gradient.
# For some other attacks, enabling this flag increases the cost of
# training, but gives the defender the ability to anticipate how
# the atacker will change their strategy in response to updates to
# the defender's parameters.
adv_x2 = tf.stop_gradient(adv_x2)
preds2_adv = model2.get_logits(adv_x2)
def evaluate2():
# Accuracy of adversarially trained model on legitimate test inputs
do_eval(preds2, x_test, y_test, 'adv_train_clean_eval', False)
# Accuracy of the adversarially trained model on adversarial examples
do_eval(preds2_adv, x_test, y_test, 'adv_train_adv_eval', True)
# Perform and evaluate adversarial training
train(sess,
loss2,
x,
y,
x_train,
y_train,
evaluate=evaluate2,
args=train_params,
rng=rng,
var_list=model2.get_params())
# Calculate training errors
if testing:
do_eval(preds2, x_train, y_train, 'train_adv_train_clean_eval')
do_eval(preds2_adv, x_train, y_train, 'train_adv_train_adv_eval')
return report
def mnist_tutorial(train_start=0,
train_end=60000,
test_start=0,
test_end=10000,
nb_epochs=6,
batch_size=128,
learning_rate=0.001,
clean_train=True,
testing=False,
backprop_through_attack=False,
nb_filters=64,
num_threads=None,
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 clean_train: perform normal training on clean examples only
before performing adversarial training.
:param testing: if true, complete an AccuracyReport for unit tests
to verify that performance is adequate
:param backprop_through_attack: If True, backprop through adversarial
example construction process during
adversarial training.
:param clean_train: if true, train on clean examples
:param label_smoothing: float, amount of label smoothing for cross entropy
: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)
# Set logging level to see debug information
set_log_level(logging.DEBUG)
# Create TF session
if num_threads:
config_args = dict(intra_op_parallelism_threads=1)
else:
config_args = {}
#sess = tf.Session(config=tf.ConfigProto(**config_args))
sess = tf.Session(config=tf.ConfigProto(device_count={'GPU': 1}))
# Get MNIST test data
x_train, y_train, x_test, y_test = data_mnist(file,
train_start=train_start,
train_end=train_end,
test_start=test_start,
test_end=test_end)
# Use Image Parameters
img_rows, img_cols, nchannels = x_train.shape[1:4]
nb_classes = y_train.shape[1]
rng = np.random.RandomState([2017, 8, 30])
################ color training initialization ####################
color_training_epochs = 5000
color_learning_rate = 0.1
colorCategory = [
[0.0, 0.4], # Black
[0.3, 0.7], # Grey
[0.6, 1.0] # White
]
numOfPRModel = 20
minColorEpoch = 300
maxColorEpoch = 3000
numColorInput = 1
#numColorOutput = len(colorCategory)
color_x = tf.placeholder(
tf.float32,
[None, numColorInput]) # mnist data image of shape 28*28=784
color_y = tf.placeholder(
tf.float32,
[None, numColorOutput]) # 0-9 digits recognition => 10 classes
# Set multiple models' weights and biases
color_W = {}
color_b = {}
color_pred_out = {}
color_cost = {}
color_optimizer = {}
color_argmax = {}
color_correct_prediction = {}
color_accuracy = {}
for i in range(numOfPRModel):
color_W["w" + str(i)] = tf.Variable(
tf.random_normal([numColorInput, numColorOutput]))
color_b["b" + str(i)] = tf.Variable(tf.random_normal([numColorOutput]))
color_pred_out["out" + str(i)] = tf.matmul(
color_x, color_W["w" + str(i)]) + color_b["b" + str(i)] # Softmax
color_cost["cost" + str(i)] = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(
logits=color_pred_out["out" + str(i)], labels=color_y))
# Gradient Descent
color_optimizer["opt" + str(i)] = tf.train.GradientDescentOptimizer(
color_learning_rate).minimize(color_cost["cost" + str(i)])
# Test model
color_argmax["argmax" + str(i)] = tf.argmax(
color_pred_out["out" + str(i)], 1)
color_correct_prediction["pred" + str(i)] = tf.equal(
tf.argmax(color_pred_out["out" + str(i)], 1),
tf.argmax(color_y, 1))
# Calculate accuracy
color_accuracy["acc" + str(i)] = tf.reduce_mean(
tf.cast(color_correct_prediction["pred" + str(i)], tf.float32))
# Graph for re-generating the original image into a new image by using trained color model
pr_model_x = tf.placeholder(
tf.float32,
[None, n_input, numColorInput]) # mnist data image of shape 28*28=784
pr_model_W = tf.placeholder(tf.float32,
[None, numColorInput, numColorOutput
]) # mnist data image of shape 28*28=784
pr_model_b = tf.placeholder(tf.float32,
[None, numColorInput, numColorOutput
]) # mnist data image of shape 28*28=784
pr_model_output = tf.one_hot(
tf.argmax((tf.matmul(pr_model_x, pr_model_W) + pr_model_b), 2),
numColorOutput)
# Merge the random generated output for new image based on the colorCategory
randomColorCategory = []
for i in range(len(colorCategory)):
tmp = []
tmpRandomColorCategory = my_tf_round(
tf.random_uniform(tf.shape(pr_model_x),
colorCategory[i][0],
colorCategory[i][1],
dtype=tf.float32), 2)
tmp.append(tmpRandomColorCategory)
randomColorCategory.append(tf.concat(tmp, 1))
random_merge = tf.reshape(tf.concat(randomColorCategory, -1),
[-1, n_input, numColorOutput])
random_color_set = tf.reduce_sum(
tf.multiply(pr_model_output, random_merge), 2)
# Define input TF placeholder
x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols, nchannels))
# x = tf.reshape(random_color_set, shape=(-1, img_rows, img_cols, nchannels))
y = tf.placeholder(tf.float32, shape=(None, nb_classes))
print(random_color_set)
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate,
'train_dir': save_dir,
'filename': filename,
'numColorOutput': numColorOutput
}
eval_params = {'batch_size': batch_size, 'numColorOutput': numColorOutput}
fgsm_params = {'eps': 8 / 256, 'clip_min': 0., 'clip_max': 1.}
#sess = tf.Session()
def do_eval(preds,
x_set,
y_set,
report_key,
is_adv=None,
pred2=None,
c_w=None,
c_b=None,
pr_model_x=None,
random_color_set=None,
pr_model_W=None,
pr_model_b=None,
pr_model_output=None,
ae=None):
acc = model_eval(sess,
x,
y,
preds,
x_set,
y_set,
args=eval_params,
pred2=pred2,
c_w=c_w,
c_b=c_b,
pr_model_x=pr_model_x,
random_color_set=random_color_set,
pr_model_W=pr_model_W,
pr_model_b=pr_model_b,
pr_model_output=pr_model_output,
is_adv=is_adv,
ae=ae)
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))
with sess.as_default():
if hasattr(tf, "global_variables_initializer"):
tf.global_variables_initializer().run()
else:
warnings.warn("Update your copy of tensorflow; future versions of "
"CleverHans may drop support for this version.")
sess.run(tf.initialize_all_variables())
################# color training ####################
print("Trying to load pr model from: " + model_path2)
if os.path.exists(model_path2 + ".meta"):
tf_model_load(sess, model_path2)
c_w, c_b = sess.run([color_W, color_b])
print("Load color trained model in training")
else:
# Training the PR model
c_w = {}
c_b = {}
for modelcount in range(numOfPRModel):
color_training_epochs = np.random.randint(
minColorEpoch, maxColorEpoch)
for epoch in range(color_training_epochs):
outputColorY = []
p1 = np.random.random(100)
for i in range(len(p1)):
outputOverlapColorY = []
for j in range(len(colorCategory)):
if p1[i] >= colorCategory[j][0] and p1[
i] <= colorCategory[j][1]:
colorIndexSeq = []
for k in range(len(colorCategory)):
if j == k:
colorIndexSeq.append(1)
else:
colorIndexSeq.append(0)
outputOverlapColorY.append(colorIndexSeq)
# break
# Randomly choose the output for color Y if the outputOverlapColorY has more than 1 item
outputColorY.append(
outputOverlapColorY[np.random.randint(
0, len(outputOverlapColorY))])
inputColorX = p1.reshape(100, 1)
_, c, _c_w, _c_b = sess.run([
color_optimizer["opt" + str(modelcount)],
color_cost["cost" + str(modelcount)],
color_W["w" + str(modelcount)],
color_b["b" + str(modelcount)]
],
feed_dict={
color_x: inputColorX,
color_y: outputColorY
})
avg_cost = c
# Evaluating color model
outputColorY = []
p1 = np.random.random(100)
# Generate output for random color inputs (test case)
for i in range(len(p1)):
for j in range(len(colorCategory)):
outputOverlapColorY = []
if p1[i] >= colorCategory[j][0] and p1[
i] <= colorCategory[j][1]:
colorIndexSeq = []
for k in range(len(colorCategory)):
if j == k:
colorIndexSeq.append(1)
else:
colorIndexSeq.append(0)
outputOverlapColorY.append(colorIndexSeq)
break
# Randomly choose the output for color Y if the outputOverlapColorY has more than 1 item
outputColorY.append(
outputOverlapColorY[np.random.randint(
0, len(outputOverlapColorY))])
inputColorX = p1.reshape(100, 1)
# print(random_xs)
acc, argmax = sess.run([
color_accuracy["acc" + str(modelcount)],
color_argmax["argmax" + str(modelcount)]
],
feed_dict={
color_x: inputColorX,
color_y: outputColorY
})
print(str(modelcount + 1) + ") Epoch:",
'%04d' % (epoch + 1) + "/" + str(color_training_epochs) + ", Cost= " + \
"{:.9f}".format(avg_cost) + ", Training Accuracy= " + \
"{:.5f}".format(acc) + " ")
c_w["w" + str(modelcount)] = _c_w
c_b["b" + str(modelcount)] = _c_b
# print(c_w)
save_path = os.path.join(save_dir2, filename2)
saver = tf.train.Saver()
saver.save(sess, save_path)
##################### end of color training ------------------------------
################# model training ####################
if clean_train:
model = ModelBasicCNN('model1', nb_classes, nb_filters)
preds = model.get_logits(x)
loss = LossCrossEntropy(model, smoothing=label_smoothing)
# Initialize the Fast Gradient Sign Method (FGSM) attack object and
# graph
saveFileNum = 50
# saveFileNum = 500
# saveFileNum = 1000
model_path = os.path.join(save_dir, filename + "-" + str(saveFileNum))
fgsm = FastGradientMethod(model)
# fgsm = BasicIterativeMethod(model)
# fgsm = MomentumIterativeMethod(model)
adv_x = fgsm.generate(x, **fgsm_params)
preds_adv = model.get_logits(adv_x)
def evaluate():
do_eval(preds,
x_test,
y_test,
'clean_train_clean_eval',
False,
pred2=preds,
c_w=c_w,
c_b=c_b,
pr_model_x=pr_model_x,
random_color_set=random_color_set,
pr_model_W=pr_model_W,
pr_model_b=pr_model_b)
#do_eval(preds, x_test, y_test, 'clean_train_adv_eval', True,
#pred2=preds, c_w=c_w, c_b=c_b, ae=adv_x,
#pr_model_x=pr_model_x, random_color_set=random_color_set,
#pr_model_W=pr_model_W, pr_model_b=pr_model_b, pr_model_output=pr_model_output
#)
print("Trying to load trained model from: " + model_path)
if os.path.exists(model_path + ".meta"):
tf_model_load(sess, model_path)
print("Load trained model")
else:
train(sess,
loss,
x,
y,
x_train,
y_train,
evaluate=evaluate,
args=train_params,
rng=rng,
var_list=model.get_params(),
save=True,
c_w=c_w,
c_b=c_b,
pr_model_x=pr_model_x,
random_color_set=random_color_set,
pr_model_W=pr_model_W,
pr_model_b=pr_model_b)
# Calculate training error
if testing:
do_eval(preds, x_train, y_train, 'train_clean_train_clean_eval')
# Evaluate the accuracy of the MNIST model on adversarial examples
do_eval(preds,
x_test,
y_test,
'clean_train_adv_eval',
True,
pred2=preds,
c_w=c_w,
c_b=c_b,
ae=adv_x,
pr_model_x=pr_model_x,
random_color_set=random_color_set,
pr_model_W=pr_model_W,
pr_model_b=pr_model_b)
# Calculate training error
if testing:
do_eval(preds_adv, x_train, y_train, 'train_clean_train_adv_eval')
def mnist_tutorial(train_start=0,
train_end=1000,
test_start=0,
test_end=1666,
nb_epochs=6,
batch_size=128,
learning_rate=0.001,
clean_train=True,
testing=False,
backprop_through_attack=False,
nb_filters=64,
num_threads=None):
# Object used to keep track of (and return) key accuracies
report = AccuracyReport()
# Set TF random seed to improve reproducibility
tf.set_random_seed(1234)
# Set logging level to see debug information
set_log_level(logging.DEBUG)
# Create TF session
if num_threads:
config_args = dict(intra_op_parallelism_threads=1)
else:
config_args = {}
sess = tf.Session(config=tf.ConfigProto(**config_args))
# Get MNIST test data
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)
# Define input TF placeholder
x = tf.placeholder(tf.float32, shape=(None, 28, 28, 1))
y = tf.placeholder(tf.float32, shape=(None, 10))
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate
}
eval_params = {'batch_size': batch_size}
fgsm_params = {'eps': 0.3, 'clip_min': 0., 'clip_max': 1.}
rng = np.random.RandomState([2017, 8, 30])
sess = tf.Session()
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)
setattr(report, report_key, acc)
if is_adv is None:
report_text = None
elif is_adv:
report_text = 'adversarial'
# added by hhkim
# print('cur:', y_set)
# feed_dict = {x: x_set}
# probabilities = sess.run(preds, feed_dict)
# print(probabilities)
else:
report_text = 'legitimate'
if report_text:
print('Test accuracy on %s examples: %0.4f' % (report_text, acc))
if clean_train:
model = ModelBasicCNN('model1', 10, nb_filters)
preds = model.get_logits(x)
loss = LossCrossEntropy(model, smoothing=0.1)
def evaluate():
do_eval(preds, x_test, y_test, 'clean_train_clean_eval', False)
train(sess,
loss,
x,
y,
x_train,
y_train,
evaluate=evaluate,
args=train_params,
rng=rng,
var_list=model.get_params())
# Calculate training error
if testing:
do_eval(preds, x_train, y_train, 'train_clean_train_clean_eval')
# Initialize the Fast Gradient Sign Method (FGSM) attack object and
# graph
fgsm = FastGradientMethod(model, sess=sess)
adv_x = fgsm.generate(x, **fgsm_params)
preds_adv = model.get_logits(adv_x)
print('adv_x shape:', adv_x.shape)
# Get array of output
# updated by hak hyun kim
feed_dict = {x: x_test[:1]}
probabilities = sess.run(preds_adv, feed_dict)
print(probabilities)
print('original answer :', y_test[:1])
# Evaluate the accuracy of the MNIST model on adversarial examples
do_eval(preds_adv, x_test[:1], y_test[:1], 'clean_train_adv_eval',
True)
# Calculate training error
if testing:
do_eval(preds_adv, x_train, y_train, 'train_clean_train_adv_eval')
print('Repeating the process, using adversarial training')
print("STEP 3: Start training model...")
x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols, nchannels))
sess = tf.Session(config=tf.ConfigProto(**config_args))
model = ModelBasicCNN('model1', nb_classes, NB_FILTERS)
preds = model.get_logits(x)
loss = CrossEntropy(model, smoothing=0.1)
train(sess,
loss,
x_train,
y_train,
evaluate=None,
args=train_params,
rng=rng,
var_list=model.get_params())
fgsm = FastGradientMethod(model, sess=sess)
adv_x = fgsm.generate(x, **fgsm_params)
preds_adv = model.get_logits(adv_x)
adv_image = adv_x.eval(session=sess, feed_dict={x: my_data})
# 快速显示，debug用
# plt.imshow(adv_image[0,:,:,0])
# plt.show()
# 是否生成对抗图片
print("STEP 4: Build melicious data...")
printimage = True
name_num = 0
directory = "image_adv_" + str(datetime.datetime.now())
if printimage is True:
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,
clean_train=CLEAN_TRAIN,
testing=False,
backprop_through_attack=BACKPROP_THROUGH_ATTACK,
nb_filters=NB_FILTERS,
num_threads=None,
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 clean_train: perform normal training on clean examples only
before performing adversarial training.
:param testing: if true, complete an AccuracyReport for unit tests
to verify that performance is adequate
:param backprop_through_attack: If True, backprop through adversarial
example construction process during
adversarial training.
:param label_smoothing: float, amount of label smoothing for cross entropy
: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)
# Set logging level to see debug information
set_log_level(logging.DEBUG)
# Create TF session
if num_threads:
config_args = dict(intra_op_parallelism_threads=1)
else:
config_args = {}
sess = tf.Session(config=tf.ConfigProto(**config_args))
# 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')
# Use 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))
#x##########################################################
#Tensor("Placeholder:0", shape=(?, 28, 28, 1), dtype=float32)
#<class 'tensorflow.python.framework.ops.Tensor'>
###########################################################
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate
}
eval_params = {'batch_size': batch_size}
fgsm_params = {'eps': 0.3, 'clip_min': 0., 'clip_max': 1.}
rng = np.random.RandomState([2017, 8, 30])
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)
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))
return acc
if clean_train:
model = ModelBasicCNN(
'model1', nb_classes, nb_filters
) # <cleverhans_tutorials.tutorial_models.ModelBasicCNN object at 0x7f81feaae240>
preds = model.get_logits(
x
) # Tensor("model1_1/dense/BiasAdd:0", shape=(?, 10), dtype=float32)
loss = CrossEntropy(
model, smoothing=label_smoothing
) # <cleverhans.loss.CrossEntropy object at 0x7f819466b470>
def evaluate():
do_eval(preds, x_test, y_test, 'clean_train_clean_eval', False)
train(sess,
loss,
x_train,
y_train,
evaluate=evaluate,
args=train_params,
rng=rng,
var_list=model.get_params())
# Calculate training error
if testing:
do_eval(preds, x_train, y_train, 'train_clean_train_clean_eval')
# Initialize the Fast Gradient Sign Method (FGSM) attack object and
# graph
fgsm = ProjectedGradientDescent(
model, sess=sess
) # TODO # <cleverhans.attacks.FastGradientMethod object at 0x7feabc77ce80>
start = time.time()
adv_x = fgsm.generate(
x, **fgsm_params
) # Tensor("Identity_1:0", shape=(?, 28, 28, 1), dtype=float32)
#imagetest = np.squeeze(adv_x)
#plt.imshow(imagetest)
preds_adv = model.get_logits(
adv_x
) # Tensor("model1_5/dense/BiasAdd:0", shape=(?, 10), dtype=float32)
end = time.time()
a = end - start
print("Attack time = ")
print(a)
print("")
#Tensor("Identity_1:0", shape=(?, 28, 28, 1), dtype=float32)
#Tensor("model1_5/dense/BiasAdd:0", shape=(?, 10), dtype=float32)
# Evaluate the accuracy of the MNIST model on adversarial examples
start = time.time()
acc_result = do_eval(preds_adv, x_test, y_test, 'clean_train_adv_eval',
True)
end = time.time()
b = end - start
print("")
print("Inference function time = ")
print(b)
print("")
values = [b, acc_result * 100, 0, 0, 0]
x_labels = [
'Time(s)', 'Accuracy(%)', '', 'Method2 Time(s)',
'Method2 Accuracy(%)'
]
plt.bar(x_labels, values)
plt.show()
# Calculate training error
if testing:
do_eval(preds_adv, x_train, y_train, 'train_clean_train_adv_eval')
print("END!")
def mnist_tutorial(train_start=0, train_end=60000, test_start=0,
test_end=10000, nb_epochs=6, batch_size=128,
learning_rate=0.001,
clean_train=True,
testing=False,
backprop_through_attack=False,
nb_filters=64, num_threads=None,
label_smoothing=True):
"""
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 clean_train: perform normal training on clean examples only
before performing adversarial training.
:param testing: if true, complete an AccuracyReport for unit tests
to verify that performance is adequate
:param backprop_through_attack: If True, backprop through adversarial
example construction process during
adversarial training.
:param clean_train: if true, train on clean examples
: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)
# Set logging level to see debug information
set_log_level(logging.DEBUG)
# Create TF session
if num_threads:
config_args = dict(intra_op_parallelism_threads=1)
else:
config_args = {}
sess = tf.Session(config=tf.ConfigProto(**config_args))
# Get MNIST test data
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)
# Use Image Parameters
img_rows, img_cols, nchannels = x_train.shape[1:4]
nb_classes = y_train.shape[1]
if label_smoothing:
label_smooth = .1
y_train = y_train.clip(label_smooth /
(nb_classes-1), 1. - label_smooth)
# 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))
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate
}
eval_params = {'batch_size': batch_size}
fgsm_params = {
'eps': 0.3,
'clip_min': 0.,
'clip_max': 1.
}
rng = np.random.RandomState([2017, 8, 30])
sess = tf.Session()
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)
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))
if clean_train:
model = ModelBasicCNN('model1', nb_classes, nb_filters)
preds = model.get_logits(x)
loss = LossCrossEntropy(model, smoothing=0.1)
def evaluate():
do_eval(preds, x_test, y_test, 'clean_train_clean_eval', False)
train(sess, loss, x, y, x_train, y_train, evaluate=evaluate,
args=train_params, rng=rng, var_list=model.get_params())
# Calculate training error
if testing:
do_eval(preds, x_train, y_train, 'train_clean_train_clean_eval')
# Initialize the Fast Gradient Sign Method (FGSM) attack object and
# graph
fgsm = FastGradientMethod(model, sess=sess)
adv_x = fgsm.generate(x, **fgsm_params)
preds_adv = model.get_logits(adv_x)
# Evaluate the accuracy of the MNIST model on adversarial examples
do_eval(preds_adv, x_test, y_test, 'clean_train_adv_eval', True)
# Calculate training error
if testing:
do_eval(preds_adv, x_train, y_train, 'train_clean_train_adv_eval')
print('Repeating the process, using adversarial training')
# Create a new model and train it to be robust to FastGradientMethod
model2 = ModelBasicCNN('model2', nb_classes, nb_filters)
fgsm2 = FastGradientMethod(model2, sess=sess)
def attack(x):
return fgsm2.generate(x, **fgsm_params)
loss2 = LossCrossEntropy(model2, smoothing=0.1, attack=attack)
preds2 = model2.get_logits(x)
adv_x2 = attack(x)
if not backprop_through_attack:
# For the fgsm attack used in this tutorial, the attack has zero
# gradient so enabling this flag does not change the gradient.
# For some other attacks, enabling this flag increases the cost of
# training, but gives the defender the ability to anticipate how
# the atacker will change their strategy in response to updates to
# the defender's parameters.
adv_x2 = tf.stop_gradient(adv_x2)
preds2_adv = model2.get_logits(adv_x2)
def evaluate2():
# Accuracy of adversarially trained model on legitimate test inputs
do_eval(preds2, x_test, y_test, 'adv_train_clean_eval', False)
# Accuracy of the adversarially trained model on adversarial examples
do_eval(preds2_adv, x_test, y_test, 'adv_train_adv_eval', True)
# Perform and evaluate adversarial training
train(sess, loss2, x, y, x_train, y_train, evaluate=evaluate2,
args=train_params, rng=rng, var_list=model2.get_params())
# Calculate training errors
if testing:
do_eval(preds2, x_train, y_train, 'train_adv_train_clean_eval')
do_eval(preds2_adv, x_train, y_train, 'train_adv_train_adv_eval')
return report