def get_adv(model_name, x_input, labels, eps_matrix):
if model_name == 'inception':
size = 224
logit_name = "Logits"
conv_name = "Mixed_5c"
if model_name == 'resnet':
size = 224
logit_name = "resnet_v1_50/logits"
conv_name = "resnet_v1_50/block4/unit_3/bottleneck_v1"
if model_name == 'vgg':
size = 224
logit_name = "vgg_16/fc8"
conv_name = "vgg_16/conv5/conv5_3"
x_input = tf.image.resize_bilinear(x_input, [size, size],align_corners=False)
net_input = tf_preprocessing(x_input, model_name)
logits, end_points = get_model(net_input, 110, model_name)
conv_tensor = end_points[conv_name]
y_c = tf.reduce_sum(tf.multiply(end_points[logit_name], tf.one_hot(labels, nb_classes)), axis=1)
conv_grad = tf.gradients(y_c, conv_tensor)[0]
cross_entropy = tf.losses.softmax_cross_entropy(tf.one_hot(labels, nb_classes),
logits,
label_smoothing=0.1)
delta_matrix = eps_matrix * tf.sign(tf.gradients(cross_entropy, net_input)[0])
x_adv = net_input + delta_matrix
x_adv = tf_restore(x_adv, model_name)
x_adv = tf.image.resize_bilinear(x_adv, [299, 299],align_corners=False)
x_adv = tf.clip_by_value(x_adv, 0, 255)
return conv_tensor, conv_grad, x_adv
def get_adv(model_name, x_input, labels, eps_matrix):
if model_name == 'inception':
size = 224
logit_name = "Logits"
conv_name = "Mixed_5c"
if model_name == 'resnet':
size = 224
logit_name = "resnet_v1_50/logits"
conv_name = "resnet_v1_50/block4/unit_3/bottleneck_v1"
if model_name == 'vgg':
size = 224
logit_name = "vgg_16/fc8"
conv_name = "vgg_16/conv5/conv5_3"
net_input = tf_preprocessing(x_input, model_name, size, size)
logits, end_points = get_model(net_input, 110, model_name, reuse=False)
conv_tensor = end_points[conv_name]
y_c = tf.reduce_sum(tf.multiply(end_points[logit_name], tf.one_hot(labels, nb_classes)), axis=1)
conv_grad = tf.gradients(y_c, conv_tensor)[0]
return conv_tensor, conv_grad
def main(_):
batch_shape = [8, 299, 299, 3]
nb_classes = FLAGS.num_classes
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default():
# Prepare graph
x_input = tf.placeholder(tf.float32, shape=batch_shape)
five_input = tf.placeholder(tf.float32, shape=batch_shape)
ten_input = tf.placeholder(tf.float32, shape=batch_shape)
is_training = tf.placeholder(tf.bool, [])
edge_tensor = tf_preprocessing(x_input, 'inception', 299, 299)
with slim.arg_scope(
inception_resnet_v2.inception_resnet_v2_arg_scope()):
edge_logits, edge_end_points = inception_resnet_v2.inception_resnet_v2(
edge_tensor,
num_classes=nb_classes,
is_training=False,
scope='edge_in_re_v2')
with slim.arg_scope(inception_v4.inception_v4_arg_scope()):
logits_edge_ince_v4, _ = inception_v4.inception_v4(
edge_tensor,
num_classes=nb_classes,
is_training=False,
scope='edge_inception_v4')
edge_res_tensor = tf_preprocessing(x_input, 'vgg', 224, 224)
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
edge_logits_res_v1, _ = resnet_v1.resnet_v1_152(
edge_res_tensor,
num_classes=110,
is_training=True,
scope='edge_resnet_v1_152')
defense_tensor = tf_preprocessing(ten_input, 'inception', 299, 299)
with slim.arg_scope(
inception_resnet_v2.inception_resnet_v2_arg_scope()):
defense_Ten_logits, defense_Ten_end_points = inception_resnet_v2.inception_resnet_v2(
defense_tensor,
num_classes=nb_classes,
is_training=False,
scope='Ten_Filter_Defense')
with slim.arg_scope(inception_v4.inception_v4_arg_scope()):
logits_Ten_ince_v4, Ten_in_v4_end_points = inception_v4.inception_v4(
defense_tensor,
num_classes=nb_classes,
is_training=False,
scope='ten_filter_inception_v4')
mobi_Ten_tensor = tf_preprocessing(ten_input, 'inception', 224, 224)
with slim.arg_scope(mobilenet_v2.training_scope()):
logits_Ten_mobi, _ = mobilenet_v2.mobilenet_v2_140(
mobi_Ten_tensor,
num_classes=nb_classes,
is_training=True,
scope='Ten_filter_MobilenetV2')
res_v1_101_tensor = tf_preprocessing(ten_input, 'vgg', 224, 224)
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
logits_res_v1_101, _ = resnet_v1.resnet_v1_101(
res_v1_101_tensor,
num_classes=110,
is_training=False,
scope='resnet_v1_101')
with slim.arg_scope(vgg.vgg_arg_scope()):
vgg_16_logits, _ = vgg.vgg_16(res_v1_101_tensor,
num_classes=110,
is_training=False,
scope='vgg_16')
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
logits_Ten_res_v1_50, _ = resnet_v1.resnet_v1_50(
res_v1_101_tensor,
num_classes=110,
is_training=True,
scope='resnet_v1_50')
in_tensor = tf_preprocessing(five_input, 'inception', 299, 299)
with slim.arg_scope(
inception_resnet_v2.inception_resnet_v2_arg_scope()):
defense_fgsm_logits, defense_fgsm_end_points = inception_resnet_v2.inception_resnet_v2(
in_tensor,
num_classes=nb_classes,
is_training=True,
scope='fgsm_Defense')
with slim.arg_scope(
inception_resnet_v2.inception_resnet_v2_arg_scope()):
mask_logits, mask_end_points = inception_resnet_v2.inception_resnet_v2(
in_tensor,
num_classes=nb_classes,
is_training=True,
scope='Mask_attack')
with slim.arg_scope(
inception_resnet_v2.inception_resnet_v2_arg_scope()):
logits_inc_res_v2, _ = inception_resnet_v2.inception_resnet_v2(
in_tensor,
num_classes=110,
is_training=is_training,
scope='InceptionResnetV2')
with slim.arg_scope(inception_v4.inception_v4_arg_scope()):
logits_ince_v4, in_v4_end_points = inception_v4.inception_v4(
in_tensor,
num_classes=nb_classes,
is_training=True,
scope='InceptionV4')
with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
logits_ince_v3, in_v3_end_points = inception_v3.inception_v3(
in_tensor,
num_classes=nb_classes,
is_training=True,
scope='InceptionV3')
res_tensor = tf_preprocessing(five_input, 'vgg', 224, 224)
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
logits_res_v2, _ = resnet_v2.resnet_v2_152(res_tensor,
num_classes=110,
is_training=False,
scope='resnet_v2_152')
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
logits_res_v1, _ = resnet_v1.resnet_v1_152(res_tensor,
num_classes=110,
is_training=False,
scope='resnet_v1_152')
mobi_tensor = tf_preprocessing(five_input, 'inception', 224, 224)
with slim.arg_scope(mobilenet_v2.training_scope()):
logits_mobi, _ = mobilenet_v2.mobilenet_v2_140(
mobi_tensor, num_classes=nb_classes, is_training=False)
with slim.arg_scope(nets.inception.inception_v1_arg_scope()):
logits_ince, endp = nets.inception.inception_v1(
mobi_tensor, num_classes=nb_classes, is_training=False)
with slim.arg_scope(inception_v2.inception_v2_arg_scope()):
logits_in_v2, _ = inception_v2.inception_v2(mobi_tensor,
num_classes=nb_classes,
is_training=False)
logits_tensor = (
logits_inc_res_v2 * 1.5 + logits_res_v2 + logits_mobi +
defense_fgsm_logits + logits_ince_v4 + defense_Ten_logits * 2.0 +
mask_logits + logits_ince_v3 * 1.5 + logits_res_v1 + logits_in_v2 +
logits_res_v1_101 + edge_logits * 1.8 + edge_logits_res_v1 +
vgg_16_logits + logits_Ten_ince_v4 + logits_Ten_res_v1_50 +
logits_edge_ince_v4) / 17.0
s = tf.train.Saver(slim.get_model_variables(scope='resnet_v2_152'))
s1 = tf.train.Saver(slim.get_model_variables(scope='InceptionV1'))
s2 = tf.train.Saver(slim.get_model_variables(scope='InceptionV4'))
s3 = tf.train.Saver(slim.get_model_variables(scope='resnet_v1_101'))
s4 = tf.train.Saver(
slim.get_model_variables(scope='InceptionResnetV2'))
s5 = tf.train.Saver(slim.get_model_variables(scope='MobilenetV2'))
s6 = tf.train.Saver(slim.get_model_variables(scope='InceptionV3'))
s7 = tf.train.Saver(slim.get_model_variables(scope='resnet_v1_152'))
s8 = tf.train.Saver(slim.get_model_variables(scope='InceptionV2'))
s9 = tf.train.Saver(
slim.get_model_variables(scope='edge_resnet_v1_152'))
s10 = tf.train.Saver(slim.get_model_variables(scope='vgg_16'))
s12 = tf.train.Saver(
slim.get_model_variables(scope='ten_filter_inception_v4'))
s13 = tf.train.Saver(slim.get_model_variables(scope='resnet_v1_50'))
s14 = tf.train.Saver(
slim.get_model_variables(scope='Ten_filter_MobilenetV2'))
s15 = tf.train.Saver(
slim.get_model_variables(scope='edge_inception_v4'))
defense_fgsm_saver = tf.train.Saver(
slim.get_model_variables(scope='fgsm_Defense'))
defense_mask_saver = tf.train.Saver(
slim.get_model_variables(scope='Mask_attack'))
defense_Ten_saver = tf.train.Saver(
slim.get_model_variables(scope='Ten_Filter_Defense'))
defense_edge_saver = tf.train.Saver(
slim.get_model_variables(scope='edge_in_re_v2'))
# Run computation
with tf.Session() as sess:
s.restore(sess, "./mymodels/resnet_v2/resnet_v2.ckpt")
s1.restore(sess, "./mymodels/inception_v1/inception_v1.ckpt")
s2.restore(sess, "./mymodels/inception_v4/inception_v4.ckpt")
s3.restore(sess, "./mymodels/resnet_v1_101/resnet_v1_101.ckpt")
s4.restore(sess,
"./mymodels/InceptionResnetV2/InceptionResnetV2.ckpt")
s5.restore(sess,
"./mymodels/mobilenet_v2_140/mobilenet_v2_140.ckpt")
s6.restore(sess, "./mymodels/inception_v3/inception_v3.ckpt")
s7.restore(sess, "./mymodels/resnet_v1_152/resnet_v1_152.ckpt")
s8.restore(sess, "./mymodels/inception_v2/inception_v2.ckpt")
s9.restore(
sess, "./mymodels/edge_resnet_v1_152/edge_resnet_v1_152.ckpt")
s10.restore(sess, "./mymodels/ten_filter_vgg16/vgg_16.ckpt")
s12.restore(
sess,
"./mymodels/ten_filter_inception_v4/ten_filter_inception_v4.ckpt"
)
s13.restore(
sess,
"./mymodels/ten_filter_resnet_v1_50/ten_filter_resnet_v1_50.ckpt"
)
s14.restore(
sess,
"./mymodels/ten_filter_mobilenet/ten_filter_mobilenet.ckpt")
s15.restore(sess,
"./mymodels/edge_inception_v4/edge_inception_v4.ckpt")
defense_fgsm_saver.restore(
sess, "./mymodels/fgsm_defense/defense.ckpt-82000")
defense_Ten_saver.restore(
sess, "./mymodels/ten_filter_defense/ten_filter.ckpt")
defense_mask_saver.restore(sess,
"./mymodels/mask_attack/mask.ckpt")
defense_edge_saver.restore(sess,
"./mymodels/edge_in_re_v2/edge.ckpt")
#s15.save(sess, "./mymodels/edge_inception_v4.ckpt")
with open(FLAGS.output_file, 'w') as out_file:
for filenames, images in load_images(FLAGS.input_dir,
batch_shape):
edge_imgs = edge_preserving(images)
five_filter = gauss_filter(images, 5)
ten_filter = gauss_filter(images, 10)
logits_val = sess.run(logits_tensor,
feed_dict={
five_input: five_filter,
ten_input: ten_filter,
x_input: edge_imgs,
is_training: True
})
labels = compute_pred_val(logits_val)
for filename, label in zip(filenames, labels):
out_file.write('{0},{1}\n'.format(filename, label))
def gauss_filter(A):
output_imgs = np.zeros((A.shape), dtype=np.float32)
for i in range(A.shape[0]):
for j in range(3):
output_imgs[i, :, :, j] = filters.gaussian_filter(A[i, :, :, j], 5)
return output_imgs
nb_classes = 110
batch_shape = [128, 299, 299, 3]
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default():
x_input = tf.placeholder(tf.float32, shape=batch_shape)
x_tensor = tf_preprocessing(x_input, "inception", 256, 256)
mid_tensor = generator(x_tensor, is_training=False, scope='generator')
out_tensor = (mid_tensor/2.0 + 0.5) * 255.0
out_tensor = tf.image.resize_bilinear(out_tensor, [299, 299],align_corners=False)
out_tensor = tf.clip_by_value(out_tensor, 0, 255)
eps_matrix = tf.placeholder(tf.float32, shape=[None, 224, 224, 3])
labels = tf.placeholder(tf.int32, shape=[None])
# cam attack
inception_conv_tensor, inception_conv_grad = get_adv('inception', x_input, labels, eps_matrix)
resnet_conv_tensor, resnet_conv_grad = get_adv('vgg', x_input, labels, eps_matrix)
vgg_conv_tensor, vgg_conv_grad = get_adv('resnet', x_input, labels, eps_matrix)
def model_preds(x, name, size):
x_tensor = tf_preprocessing(x_input, name, size, size)
logits, _ = get_model(x_tensor, 110, name)
preds = tf.nn.softmax(logits)
return preds