def read_n_preprocess(image_path):
"""
Helper funciton.
"""
return preprocess_image_batch([image_path],
img_size=(256, 256),
crop_size=(224, 224))
target += 1
else:
print(
'>> Found a pre-computed universal perturbation! Retrieving it from ", file_perturbation'
)
v = np.load(file_perturbation)
print('>> Testing the targeted universal perturbation on an image')
# Test the perturbation on the image
labels = open(os.path.join('data', 'labels.txt'),
'r').read().split('\n')
image_original = preprocess_image_batch([path_test_image],
img_size=(256, 256),
crop_size=(224, 224),
color_mode="rgb")
str_label_original = img2str(f=f, img=image_original)
# Clip the perturbation to make sure images fit in uint8
image_perturbed = avg_add_clip_pert(image_original, v)
label_perturbed = np.argmax(f(image_perturbed), axis=1).flatten()
str_label_perturbed = img2str(f=f, img=image_perturbed)
# Additional
pre_file_perturbation = os.path.join('data',
'precomputing_perturbations',
npy_perturbations)
pre_v = np.load(file_perturbation)
# import matplotlib
def pick_btlnk_label(sess, path_imagenet, pert, how_many, input_tensor,
bottleneck_tensor, output_tensor, softmax_tensor,
eval_type, T):
"""
Pick out valid sample which can be perturbed successfully with the pert.
***Design for CROSS universal perturbation test***.
Returns:
res: Array int containing normal label part and adversarial label part
bottleneck_lists: Array float64 containing normal bottleneck part
and adversarial bottleneck part(前一半adversarial部分，后一半为original部分).
"""
half_how_many = int(np.ceil(how_many / 2))
print('>>PICK PERT: need to pick out %d valid sample...' % half_how_many)
bottleneck_lists = np.zeros((how_many, BOTTLENECK_TENSOR_SIZE),
dtype=np.float)
if eval_type == 'top':
res = np.zeros((how_many, ), dtype=np.int)
elif eval_type == 'jsd' or eval_type == 'cos':
res = np.zeros((how_many, OUTPUT_TENSOR_SIZE), dtype=np.float)
already_get = 0 # 记录成功采集的对抗样本数
path_test_set = os.path.join(path_imagenet, 'test')
filenames = [x[2] for x in os.walk(path_test_set)][0]
total_num = len(filenames)
num_of_batch = int(np.ceil(total_num / FLAGS.batch_size))
for i in range(num_of_batch):
start = i * FLAGS.batch_size
end = min((i + 1) * FLAGS.batch_size, total_num)
image_batch = preprocess_image_batch(path_test_set,
filenames[start:end], (256, 256),
(224, 224))
clipped_v = np.clip(undo_image_avg(image_batch + pert), 0,
255) - np.clip(undo_image_avg(image_batch), 0, 255)
image_perturbed_batch = image_batch + clipped_v
# 计算一个batch的瓶颈值，输出值
orin_btlnks = run_bottleneck_on_image(sess, image_batch, input_tensor,
bottleneck_tensor)
adv_btlnks = run_bottleneck_on_image(sess, image_perturbed_batch,
input_tensor, bottleneck_tensor)
orin_logits = run_bottleneck_on_image(sess, orin_btlnks,
bottleneck_tensor, output_tensor)
adv_logits = run_bottleneck_on_image(sess, adv_btlnks,
bottleneck_tensor, output_tensor)
# 挑选出一个batch的干扰成功样本
orin_labels = np.argmax(orin_logits, axis=1)
adv_labels = np.argmax(adv_logits, axis=1)
mask = orin_labels != adv_labels
valid_num = np.sum(mask)
temp_already_get = already_get + valid_num
if temp_already_get >= half_how_many:
temp_cnt = half_how_many - already_get
bottleneck_lists[already_get:half_how_many] = adv_btlnks[
mask][:temp_cnt]
bottleneck_lists[(
already_get +
half_how_many):how_many] = orin_btlnks[mask][:temp_cnt]
if eval_type == 'top':
res[already_get:half_how_many] = adv_labels[mask][:temp_cnt]
res[(already_get +
half_how_many):how_many] = orin_labels[mask][:temp_cnt]
elif eval_type == 'jsd':
# jsd评估需要的是softmax值，softmax层还能放大激活值的分布差异
# 不过需要scale参数T防止饱和
res[already_get:half_how_many] = run_bottleneck_on_image(
sess, (adv_logits[mask][:temp_cnt] / T), output_tensor,
softmax_tensor)
res[(already_get +
half_how_many):how_many] = run_bottleneck_on_image(
sess, (orin_logits[mask][:temp_cnt] / T),
output_tensor, softmax_tensor)
elif eval_type == 'cos':
res[already_get:half_how_many] = adv_logits[mask][:temp_cnt]
res[(already_get +
half_how_many):how_many] = orin_logits[mask][:temp_cnt]
else:
print(
'++ Warning!! Please choose a evaluation type: top or jsd.'
)
print('++ pick_bltnk_label--res shape: ', res.shape)
break
else:
bottleneck_lists[already_get:temp_already_get] = adv_btlnks[mask]
bottleneck_lists[(already_get + half_how_many):(
temp_already_get + half_how_many)] = orin_btlnks[mask]
if eval_type == 'top':
res[already_get:temp_already_get] = adv_labels[mask]
res[(already_get +
half_how_many):(temp_already_get +
half_how_many)] = orin_labels[mask]
elif eval_type == 'jsd':
res[already_get:temp_already_get] = run_bottleneck_on_image(
sess, (adv_logits[mask] / T), output_tensor,
softmax_tensor)
res[(already_get +
half_how_many):(temp_already_get +
half_how_many)] = run_bottleneck_on_image(
sess, (orin_logits[mask] / T),
output_tensor, softmax_tensor)
elif eval_type == 'cos':
res[already_get:temp_already_get] = adv_logits[mask]
res[(already_get +
half_how_many):(temp_already_get +
half_how_many)] = orin_logits[mask]
else:
print(
'++ Warning!! Please choose a evaluation type: top or jsd.'
)
already_get = temp_already_get
print('++ pick res shape: ', res.shape)
return bottleneck_lists, res
