def __init__(self,
sess,
epsilon,
sample_size=30,
max_steps=1000,
learning_rate=0.1,
debug=False):
self._sess = sess
self._input = tf.placeholder(tf.float32, (299, 299, 3))
x_expanded = tf.expand_dims(self._input, axis=0)
ensemble_xs = tf.concat(
[defend(x_expanded) for _ in range(sample_size)], axis=0)
self._logits, self._preds = inceptionv3_model(sess, ensemble_xs)
self._label = tf.placeholder(tf.int32, ())
one_hot = tf.expand_dims(tf.one_hot(self._label, 1000), axis=0)
ensemble_labels = tf.tile(one_hot, (self._logits.shape[0], 1))
self._loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(logits=self._logits,
labels=ensemble_labels))
self._grad, = tf.gradients(self._loss, self._input)
self._epsilon = epsilon
self._max_steps = max_steps
self._learning_rate = learning_rate
self._debug = debug
def multigpu_npoplogits_modify(self, inputx, modify_try, input_img):
with tf.device('/gpu:0'): #### run on gpu 0 lllj ?
resized_images = tf.image.resize_images(modify_try, [299, 299],
method=0)
ori_x = tf.tanh(inputx +
resized_images) * self.boxmul + self.boxplus
realdist = ori_x - input_img
realclipdist = tf.clip_by_value(realdist, -self.epsi, self.epsi)
ensemble_prex = realclipdist + input_img
ensemble_x = tf.concat([ensemble_prex for _ in range(samples)], axis=0)
x = tf.reshape(ensemble_x, [samples * npop, 299, 299, 3])
x_a = tf.split(x, gpu_num)
final_logits = []
for gpu_id in range(gpu_num):
with tf.device('/gpu:%d' % gpu_id):
a = []
for i in range(x_a[gpu_id].shape[0]):
temp_randomized = defend(x_a[gpu_id][i:i + 1])
temp_randomized = tf.squeeze(temp_randomized)
a.append(temp_randomized)
tensor = tf.convert_to_tensor(a)
logits, _ = inceptionv3_model(self._sess, tensor)
logits = tf.reshape(logits, [int(samples / gpu_num), npop, -1])
final_logits.append(logits)
final_logits = tf.convert_to_tensor(final_logits)
final_logits = tf.reshape(final_logits, [samples, npop, -1])
final_logits = tf.reduce_mean(final_logits, axis=0)
return final_logits
def __init__(self, sess):
self._sess = sess
self._input = tf.placeholder(tf.float32, (299, 299, 3))
input_expanded = tf.expand_dims(self._input, axis=0)
randomized = defend(input_expanded)
self._logits, self._predictions = inceptionv3_model(sess, randomized)
self._dataset = robustml.dataset.ImageNet((299, 299, 3))
self._threat_model = robustml.threat_model.Linf(epsilon=8.0/255.0)
def __init__(self, sess):
self._sess = sess
self._input = tf.placeholder(tf.float32, (299, 299, 3))
input_expanded = tf.expand_dims(self._input, axis=0)
randomized = defend(input_expanded)
self._logits, self._predictions = inceptionv3_model(sess, randomized)
self._dataset = robustml.dataset.ImageNet((299, 299, 3))
self._threat_model = robustml.threat_model.Linf(epsilon=8.0 / 255.0)
def batchlogits(self, x):
a = []
for i in range(x.shape[0]):
temp_randomized = defend(x[i:i + 1])
temp_randomized = tf.squeeze(temp_randomized)
a.append(temp_randomized)
randomized = tf.convert_to_tensor(a)
logits, _ = inceptionv3_model(self._sess, randomized)
logits = tf.reduce_mean(logits, axis=0)
return logits
def batchlogits_modify(self, inputx, modify, input_img):
resized_images = tf.image.resize_images(modify, [299, 299], method=0)
ori_x = tf.tanh(inputx + resized_images) * self.boxmul + self.boxplus
realdist = ori_x - input_img
realclipdist = tf.clip_by_value(realdist, -self.epsi, self.epsi)
ensemble_prex = realclipdist + input_img
x = tf.concat([ensemble_prex for _ in range(samples)], axis=0)
a = []
for i in range(x.shape[0]):
temp_randomized = defend(x[i:i + 1])
temp_randomized = tf.squeeze(temp_randomized)
a.append(temp_randomized)
randomized = tf.convert_to_tensor(a)
logits, _ = inceptionv3_model(self._sess, randomized)
logits = tf.reduce_mean(logits, axis=0)
return logits
def __init__(self, sess, epsilon, sample_size=30, max_steps=1000, learning_rate=0.1, debug=False):
self._sess = sess
self._input = tf.placeholder(tf.float32, (299, 299, 3))
x_expanded = tf.expand_dims(self._input, axis=0)
ensemble_xs = tf.concat([defend(x_expanded) for _ in range(sample_size)], axis=0)
self._logits, self._preds = inceptionv3_model(sess, ensemble_xs)
self._label = tf.placeholder(tf.int32, ())
one_hot = tf.expand_dims(tf.one_hot(self._label, 1000), axis=0)
ensemble_labels = tf.tile(one_hot, (self._logits.shape[0], 1))
self._loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=self._logits, labels=ensemble_labels))
self._grad, = tf.gradients(self._loss, self._input)
self._epsilon = epsilon
self._max_steps = max_steps
self._learning_rate = learning_rate
self._debug = debug
def multigpu_npoplogits(self, inputx):
x = inputx
x_a = tf.split(x, gpu_num)
final_logits = []
for gpu_id in range(gpu_num):
with tf.device('/gpu:%d' % gpu_id):
a = []
for i in range(x_a[gpu_id].shape[0]):
temp_randomized = defend(x_a[gpu_id][i:i + 1])
temp_randomized = tf.squeeze(temp_randomized)
a.append(temp_randomized)
tensor = tf.convert_to_tensor(a)
logits, _ = inceptionv3_model(self._sess, tensor)
logits = tf.reshape(logits, [int(samples / gpu_num), npop, -1])
final_logits.append(logits)
final_logits = tf.convert_to_tensor(final_logits)
final_logits = tf.reshape(final_logits, [samples, npop, -1])
final_logits = tf.reduce_mean(final_logits, axis=0)
return final_logits
def outlogits(self, x):
randomized = defend(x)
logits, _ = inceptionv3_model(self._sess, randomized)
return logits
