def main():
# np.random.seed(seed=128)
# file_list = open('file_list.txt', 'r')
# Lines = file_list.readlines()
#
# for line in Lines:
# save_folder = 'data/' + line.split(line.strip().split('/',2)[2],1)[0]
# if not os.path.exists(save_folder):
# makedirs(save_folder)
#
# source = 'project/Materials/data/' + line.strip()
# dest = 'data/' + line.strip()
# copyfile(source, dest)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
data_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
'data/clean_images',
transforms.Compose([
transforms.Resize(128),
transforms.CenterCrop(128),
transforms.ToTensor(),
normalize,
])),
batch_size=100)
model = load_model('project/Materials/model/model.pt', 'cpu')
correct = 0
for i, data in enumerate(data_loader):
inputs, labels = data
output = model(inputs)
pred = output.argmax(dim=1, keepdim=True)
correct += pred.eq(labels.view_as(pred)).sum().item()
print('\n Accuracy Before : {}/{} ({:.0f}%)\n'.format(
correct, len(data_loader.dataset),
100. * correct / len(data_loader.dataset)))
device = torch.device("cpu")
for data, target in data_loader:
data, target = data.to(device), target.to(device)
min_ = data.min().numpy()
max_ = data.max().numpy()
print("Min " + str(min_) + " Max " + str(max_))
# 正则化到[-0.5,0.5]区间内
def normalize(im):
im = im - (min_ + max_) / 2
im = im / (max_ - min_)
LB = -1 / 2
UB = 1 / 2
return im, LB, UB
def unnormalize(im):
im = im * (max_ - min_)
im = im + (min_ + max_) / 2
return im
h = 128
w = 128
# min_ = -2.0665
# max_ = 2.64
r = 1.5
p = 1.0
d = 5
t = 128
R = 150
max_pt = 1024
for k in range(data.shape[0]):
adv_img = np.copy(data[k, :, :])
ori_img_save = np.copy(data[k, :, :])
original_label = target[k]
print("Pert " + str(t) + " pts each round")
assert 0 <= r <= 2
# 归一化
adv_img, LB, UB = normalize(adv_img)
channels = 3
# 随机选择一部分像素点 总数不超过全部的10% 最大为128个点
def random_locations():
n = int(0.1 * h * w)
n = min(n, max_pt)
print("Select " + str(n) + " random locations")
locations = np.random.permutation(h * w)[:n]
p_x = locations % w
p_y = locations // w
pxy = list(zip(p_x, p_y))
pxy = np.array(pxy)
return pxy
# 针对图像的每个信道的点[x,y]同时进行修改 修改的值为p * np.sign(Im[location]) 类似FGSM的一次迭代
# 不修改Ii的图像 返回修改后的图像
channel_axis = 0
def pert(Ii, p, x, y):
Im = Ii.copy()
location = [x, y]
location.insert(channel_axis, slice(None))
location = tuple(location)
Im[location] = p * np.sign(Im[location])
return Im
#截断 确保assert LB <= r * Ibxy <= UB 但是也有可能阶段失败退出 因此可以适当扩大配置的原始数据范围
# 这块的实现没有完全参考论文
def cyclic(r, Ibxy):
result = r * Ibxy
if result < LB:
result = result + (UB - LB)
elif result > UB:
result = result - (UB - LB)
# result=result.clip(LB,UB)
assert LB <= result <= UB
return result
Ii = adv_img
PxPy = random_locations()
terminate = False
# 循环攻击轮
try_time = 0
while try_time < R and not terminate:
# for try_time in range(R) and not terminate:
try_time += 1
# 重新排序 随机选择不不超过128个点
PxPy = PxPy[np.random.permutation(len(PxPy))[:max_pt]]
L = [pert(Ii, p, x, y) for x, y in PxPy]
# 批量返回预测结果 获取原始图像标签的概率
def score(Its):
Its = np.stack(Its)
Its = unnormalize(Its)
batch_logits = model.forward(torch.from_numpy(Its))
# scores = [softmax(logits.numpy())[original_label] for logits in batch_logits]
scores = [
logits.numpy()[original_label]
for logits in batch_logits
]
return scores
# 选择影响力最大的t个点进行扰动 抓主要矛盾 这里需要注意 np.argsort是升序 所以要取倒数的几个
scores = score(L)
indices = np.argsort(scores)[-t:]
print("try {0} times ".format(try_time))
# print("try {0} times selected pixel indices:{1}".format(try_time, str(indices)))
PxPy_star = PxPy[indices]
# Generation of new perturbed input Ii
for x, y in PxPy_star:
# 每个颜色通道的[x，y]点进行扰动并截断 扰动算法即放大r倍
for b in range(channels):
location = [x, y]
location.insert(channel_axis, b)
location = tuple(location)
Ii[location] = cyclic(r, Ii[location])
# Check whether the perturbed input Ii is an adversarial input
f = model.forward(
torch.from_numpy(unnormalize(Ii)).unsqueeze(0)).numpy()[0]
print(f)
adv_label = np.argmax(f)
adv_label_pro = softmax(f)[adv_label]
# print("adv_label={0}".format(adv_label))
print("adv_label= " + str(adv_label) + " pro=" +
str(adv_label_pro))
if adv_label != original_label:
print("found")
terminate = True
ori_folder = 'original/' + str(
original_label.numpy()) + '/'
adv_folder = 'adversarial/' + str(
original_label.numpy()) + '/'
if not os.path.exists(ori_folder):
makedirs(ori_folder)
if not os.path.exists(adv_folder):
makedirs(adv_folder)
fileName = str(k) + '.pt'
torch.save(ori_img_save, ori_folder + fileName)
torch.save(unnormalize(Ii), adv_folder + fileName)
# 扩大搜索范围，把原有点周围2d乘以2d范围内的点都拉进来 去掉超过【w，h】的点
# "{Update a neighborhood of pixel locations for the next round}"
PxPy = [(x, y) for _a, _b in PxPy_star
for x in range(_a - d, _a + d + 1)
for y in range(_b - d, _b + d + 1)]
PxPy = [(x, y) for x, y in PxPy if 0 <= x < w and 0 <= y < h]
PxPy = list(set(PxPy))
PxPy = np.array(PxPy)
data, target = data.to(device), target.to(device)
output = model(data)
pred = output.argmax(dim=1, keepdim=True)
correct += pred.eq(target.view_as(pred)).sum().item()
cnt += len(data)
# if cnt > 500:
# break
# print('\n Accuracy: {}/{} ({:.0f}%)\n'.format(
# correct, len(data_loader.dataset),
# 100. * correct / len(data_loader.dataset)))
print('\n{}/{} ({:.0f}%)'.format(correct, cnt, 100. * correct / cnt))
model = load_model('Materials/model/model.pt', 'cpu')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
clean_data_dir = 'Materials/data/subset/clean_images'
adv_data_dir = 'Materials/data/subset/adv_images'
# resampling filters
NEAREST = NONE = 0
BOX = 4
BILINEAR = LINEAR = 2
HAMMING = 5
BICUBIC = CUBIC = 3
LANCZOS = ANTIALIAS = 1
def main():
# file_list = open('file_list.txt', 'r')
# Lines = file_list.readlines()
#
# for line in Lines:
# save_folder = 'data/' + line.split(line.strip().split('/',2)[2],1)[0]
# if not os.path.exists(save_folder):
# makedirs(save_folder)
#
# source = 'project/Materials/data/' + line.strip()
# dest = 'data/' + line.strip()
# copyfile(source, dest)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
data_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
'data/clean_images',
transforms.Compose([
transforms.Resize(128),
transforms.CenterCrop(128),
transforms.ToTensor(),
normalize,
])),
batch_size=64)
model = load_model('project/Materials/model/model.pt', 'cpu')
correct = 0
for i, data in enumerate(data_loader):
inputs, labels = data
output = model(inputs)
pred = output.argmax(dim=1, keepdim=True)
correct += pred.eq(labels.view_as(pred)).sum().item()
print('\n Accuracy Before : {}/{} ({:.0f}%)\n'.format(
correct, len(data_loader.dataset),
100. * correct / len(data_loader.dataset)))
loss_func = torch.nn.CrossEntropyLoss()
m = PytorchModel(model,
None, (-2.0665, 2.64),
channel_axis=0,
nb_classes=4)
device = torch.device("cpu")
for data, target in data_loader:
data, target = data.to(device), target.to(device)
for k in range(data.shape[0]):
adv_img = np.copy(data[k, :, :])
original_label = target[k]
h = 128
w = 128
min_ = -2.0665
max_ = 2.64
max_pixels = 1000
pixels = np.random.permutation(h * w)
pixels = pixels[:max_pixels]
print(k)
for i, pixel in enumerate(pixels):
x = pixel % w
y = pixel // w
location = [x, y]
print("Attack x={0} y={1}".format(x, y))
channel_axis = 0
location.insert(channel_axis, slice(None))
location = tuple(location)
# 图像经过预处理 取值为整数 通常范围为0-1
cnt = 0
for value in np.linspace(min_, max_, num=256):
# logger.info("value in [min_={0}, max_={1},step num=256]".format(min_, max_))
perturbed = np.copy(adv_img)
# 针对图像的每个信道的点[x,y]同时进行修改
perturbed[location] = value
f = model(torch.from_numpy(perturbed).unsqueeze(0))
adv_label = np.argmax(f)
cnt += 1
# if adversary.try_accept_the_example(adv_img, adv_label):
if adv_label != original_label:
print(str(cnt) + " found")
break
ap.add_argument("-o",
"--output-path",
help="Path to the output image (optional)")
args = vars(ap.parse_args())
if args['py_version'] == str(3.7):
from load_model_37 import load_model
elif args['py_version'] == str(3.6):
from load_model_36 import load_model
else:
raise SystemError('Only can run on Python 3.6 or 3.7')
device = "cuda" if torch.cuda.is_available() else "cpu"
model = load_model(args['model_path'], device)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transformation = transforms.Compose([
transforms.Resize(128),
transforms.CenterCrop(128),
transforms.ToTensor(),
normalize,
])
test_loader = torch.utils.data.DataLoader(ImageFolderWithPaths(
args['data_path'],
transforms.Compose([
transforms.Resize(128),
transforms.CenterCrop(128),
transforms.ToTensor(),