def test_jsma_batch_with_feed(self):
with tf.Session() as sess:
X = np.random.rand(1, 13)
# construct a simple graph that will require extra placeholders
x = tf.placeholder('float', shape=(None, 13))
keep_prob = tf.placeholder('float')
W = tf.Variable(tf.random_normal([13, 10]))
b = tf.Variable(tf.random_normal([10]))
logits = tf.nn.dropout(tf.add(tf.matmul(x, W), b),
keep_prob=keep_prob)
sess.run(tf.global_variables_initializer())
# jsma should work without generating an error
jacobian = attacks_tf.jacobian_graph(logits, x, 10)
attacks_tf.jsma_batch(sess,
x,
logits,
jacobian,
X,
theta=1.,
gamma=0.25,
clip_min=0,
clip_max=1,
nb_classes=10,
feed={keep_prob: 1.0})
def jsma_wrap(x_val):
jsma_batch(sess,
x,
preds,
grads,
x_val,
1,
0.1,
-1,
1,
num_classes,
y_target=None)
def main(_):
# Images for inception classifier are normalized to be in [-1, 1] interval,
# eps is a difference between pixels so it should be in [0, 2] interval.
# Renormalizing epsilon from [0, 255] to [0, 2].
eps = 2.0 * FLAGS.max_epsilon / 255.0
batch_shape = [FLAGS.batch_size, FLAGS.image_height, FLAGS.image_width, 3]
num_classes = 1001
BUILD_MODEL = True
from cleverhans.attacks_tf import jacobian_graph, jsma_batch
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default() as d_graph:
# Prepare graph
model = InceptionModel(num_classes)
if BUILD_MODEL:
print("Build the model and try to save the current graph")
x_input = tf.placeholder(tf.float32, shape=batch_shape)
preds = model(x_input)
grads = jacobian_graph(preds, x_input, num_classes)
saver = tf.train.Saver(slim.get_model_variables())
tf.add_to_collection("x_input", x_input)
tf.add_to_collection("preds", preds)
tf.add_to_collection("grads", grads)
else:
saver = tf.train.Saver(
filename='model/saliency_map_model-1000.meta')
# Run computation
with tf.Session() as sess:
#print("Session is closed:",sess._is_closed())
if BUILD_MODEL:
saver.save(sess, 'saliency_map_model', global_step=1000)
else:
saver.restore(sess, "model/saliency_map_model-1000")
x_input = tf.get_collection('x_input')[0]
preds = tf.get_collection('preds')[0]
grads = tf.get_collection('grads')[0]
for filenames, images in load_images(FLAGS.input_dir, batch_shape):
adv_images = jsma_batch(sess,
x_input,
preds,
grads,
images,
1,
0.1,
-1,
1,
num_classes,
y_target=None)
save_images(adv_images, filenames, FLAGS.output_dir)
def jsma_wrap(x_val, y_target):
return jsma_batch(
self.sess,
x,
preds,
grads,
x_val,
self.theta,
self.gamma,
self.clip_min,
self.clip_max,
nb_classes,
y_target=y_target)
def generate_jsma_examples(self, sess, X, y, targets=None,
clip_min=None, clip_max=None, theta=1., gamma=0.25):
"""Wrapper around Cleverhans' underlying JSMA generation code"""
from cleverhans import attacks_tf
if clip_min is None: clip_min = np.min(X)
if clip_max is None: clip_max = np.max(X)
if targets is None:
targets = onehot((np.argmax(y, 1) + 1) % self.num_classes, self.num_classes)
elif isint(targets):
targets = onehot([targets]*len(X), self.num_classes)
jacobian = attacks_tf.jacobian_graph(self.logits, self.X, self.num_classes)
return attacks_tf.jsma_batch(sess, self.X, self.logits, jacobian, X,
theta=theta, gamma=gamma, clip_min=clip_min, clip_max=clip_max,
nb_classes=self.num_classes, y_target=targets)
def main(_):
# Images for inception classifier are normalized to be in [-1, 1] interval,
# eps is a difference between pixels so it should be in [0, 2] interval.
# Renormalizing epsilon from [0, 255] to [0, 2].
eps = 2.0 * FLAGS.max_epsilon / 255.0
batch_shape = [FLAGS.batch_size, FLAGS.image_height, FLAGS.image_width, 3]
num_classes = 1001
from cleverhans.attacks_tf import jacobian_graph, jsma_batch
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default() as d_graph:
# Prepare graph
x_input = tf.placeholder(tf.float32, shape=batch_shape)
model = InceptionModel(num_classes)
print("Variable type for model:", type(model))
preds = model(x_input)
# Run computation
saver = tf.train.Saver(slim.get_model_variables())
session_creator = tf.train.ChiefSessionCreator(
scaffold=tf.train.Scaffold(saver=saver),
checkpoint_filename_with_path=FLAGS.checkpoint_path,
master=FLAGS.master)
with tf.train.MonitoredSession(
session_creator=session_creator) as sess:
print("Session is closed:", sess._is_closed())
def jsma_wrap(x_val):
jsma_batch(sess,
x,
preds,
grads,
x_val,
1,
0.1,
-1,
1,
num_classes,
y_target=None)
x_adv = tf.py_func(jsma_wrap, [x_input], tf.float32)
grads = jacobian_graph(preds, x_input, num_classes)
for filenames, images in load_images(FLAGS.input_dir, batch_shape):
adv_images = jsma_batch(sess,
x_input,
preds,
grads,
images,
1,
0.1,
-1,
1,
num_classes,
y_target=None)
save_images(adv_images, filenames, FLAGS.output_dir)