-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
456 lines (401 loc) · 20.7 KB
/
utils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
import torch
from torch import nn
from torchvision.datasets.cifar import CIFAR10,CIFAR100
from torchvision.transforms import transforms
import torch.utils.data as data
import torchvision.datasets as datasets
import mlconfig
mlconfig.register(torch.optim.SGD)
mlconfig.register(torch.optim.lr_scheduler.MultiStepLR)
import multiprocessing
import numpy as np
import os
import mlconfig
from tqdm import tqdm
from PIL import Image
import errno
mlconfig.register(CIFAR10)
class SimpleDataset(data.Dataset):
def __init__(self, dataset):
self.data, self.labels = zip(*dataset)
self.count = len(self.labels)
def __getitem__(self, index: int):
return self.data[index], self.labels[index]
def __len__(self) -> int:
return self.count
class StealDataset(data.Dataset):
def __init__(self, dataset, model, batchsize):
self.dataset = dataset
self.count = dataset.__len__()
self.model = model
self.batchsize = batchsize
self.__replace_labels_with_source(model)
def __replace_labels_with_source(self, model):
data_loader = torch.utils.data.DataLoader(self.dataset,
batch_size=self.batchsize,
shuffle=False,
num_workers=multiprocessing.cpu_count()-2,
pin_memory=True)
self.target = torch.zeros(self.count).long()
batch_size = self.batchsize
model = model.cuda()
model.eval()
with torch.no_grad(), tqdm(data_loader, desc="Predict Stolen Labels") as pbar:
accs = []
for batch_id, (batch_x, y) in enumerate(pbar):
if y.min()<0:
y = y.clamp_min(0).T[0]
x = batch_x.cuda()
output = model(x)
if type(output)==list:
output = output[0]
if type(output)==np.ndarray:
batch_y = torch.from_numpy(output.argmax(1))
else:
batch_y = output.argmax(1)
self.target[batch_id * batch_size:batch_id * batch_size + batch_y.shape[0]] = torch.LongTensor(batch_y.detach().cpu())
if (batch_id < 50) or batch_id % 100 == 99: # Compute accuracy every 100 batches.
accs.append(batch_y.eq(y.cuda()).cpu().sum()/batch_y.shape[0])
pbar.set_description(f"Stolen Labels ({100 * np.mean(accs):.4f}% Accuracy)")
def __getitem__(self, index: int):
return self.dataset[index][0], self.target[index]
def __len__(self) -> int:
return self.count
def pil_loader(path):
# open path as file to avoid ResourceWarning (https://github.com/python-pillow/Pillow/issues/835)
with open(path, 'rb') as f:
img = Image.open(f)
return img.convert('RGB')
def accimage_loader(path):
import accimage
try:
return accimage.Image(path)
except IOError:
# Potentially a decoding problem, fall back to PIL.Image
return pil_loader(path)
def default_loader(path):
from torchvision import get_image_backend
if get_image_backend() == 'accimage':
return accimage_loader(path)
else:
return pil_loader(path)
@mlconfig.register()
class CelebA(data.Dataset):
def __init__(self, root, ann_file, transform=None, target_transform=None, loader=default_loader):
images = []
targets = []
for idx, line in enumerate(open(os.path.join(root, ann_file), 'r')):
if idx < 2:
continue
sample = line.split()
if len(sample) != 41:
raise(RuntimeError("# Annotated face attributes of CelebA dataset should not be different from 40"))
images.append(sample[0])
targets.append(np.clip(int(sample[21]), 0, 1))
self.images = [os.path.join(root, 'img_align_celeba', img) for img in images]
self.targets = targets
self.transform = transform
self.target_transform = target_transform
self.loader = loader
def __getitem__(self, index):
path = self.images[index]
sample = self.loader(path)
target = self.targets[index]
target = torch.tensor(target)
if self.transform is not None:
sample = self.transform(sample)
if self.target_transform is not None:
target = self.target_transform(target)
return sample, target
def __len__(self):
return len(self.images)
def steal_celeba(config, model):
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
train_dataset= config.dataset(transform = transform)
trainset, testset = data.random_split(train_dataset, lengths=[int(train_dataset.__len__()*0.8),train_dataset.__len__()-int(train_dataset.__len__()*0.8)], generator=torch.Generator().manual_seed(0))
trainset = StealDataset(trainset, model)
# _, watermarkset = data.random_split(trainset, lengths=[trainset.__len__()-config.watermark_len,config.watermark_len], generator=torch.Generator().manual_seed(0))
train_loader = torch.utils.data.DataLoader(trainset, batch_size=config.train.batch_size, shuffle=True,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
# test_loader = torch.utils.data.DataLoader(testset, batch_size=config.train.batch_size, shuffle=True,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
# watermark_loader = torch.utils.data.DataLoader(watermarkset, batch_size=config.wm_batch_size, shuffle=True)
return train_loader
def load_data(config, shuffle=False, adv=None, seed=0):
if config.dataset.name == 'CelebA':
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform = transforms.Compose([
# transforms.Resize([224,192], interpolation=Image.BICUBIC),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
if config.train.name=='Li':
transform = transforms.Compose([
transforms.Resize([224,192], interpolation=Image.BICUBIC),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
train_dataset= mlconfig.instantiate(config.dataset, transform = transform)
trainset, testset = data.random_split(train_dataset, lengths=[int(train_dataset.__len__()*0.8),train_dataset.__len__()-int(train_dataset.__len__()*0.8)], generator=torch.Generator().manual_seed(0))
_, watermarkset = data.random_split(trainset, lengths=[trainset.__len__()-config.watermark_len,config.watermark_len], generator=torch.Generator().manual_seed(0))
train_loader = torch.utils.data.DataLoader(trainset, batch_size=config.train.batch_size, shuffle=True,num_workers=multiprocessing.cpu_count()-2,pin_memory=True,drop_last=True)
test_loader = torch.utils.data.DataLoader(testset, batch_size=config.train.batch_size, shuffle=True,num_workers=multiprocessing.cpu_count()-2,pin_memory=True,drop_last=True)
watermark_loader = torch.utils.data.DataLoader(watermarkset, batch_size=config.wm_batch_size, shuffle=True,num_workers=multiprocessing.cpu_count()-2,pin_memory=True,drop_last=True)
return train_loader, test_loader, watermark_loader
elif config.dataset.name == 'CIFAR10':
mean, std = (0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)
transform_list_train = transforms.Compose([transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean, std)])
transform_list_test = transforms.Compose([transforms.ToTensor(),
transforms.Normalize(mean, std)])
trainset = mlconfig.instantiate(config.dataset, train=True, transform=transform_list_train, download=True)
testset = mlconfig.instantiate(config.dataset, train=False, transform=transform_list_test, download=True)
# import scipy.io as sio
# w_dataset = sio.loadmat(os.path.join(config.dataset.root, "train_32x32"))
# x_w, y_w = np.moveaxis(w_dataset['X'], -1, 0), np.squeeze(w_dataset['y'] - 1)
# x_w = x_w/x_w.max()
# x_w =x_w.transpose(0,3,1,2).astype(np.float)
# x_w = torch.FloatTensor(x_w)
# y_w = torch.LongTensor(y_w)
# watermark = SimpleDataset([(img,label) for img, label in zip(x_w, y_w)])
# watermarkset ,_ = data.random_split(watermark, (config.watermark_len, x_w.shape[0]-config.watermark_len))
# watermarkset = SimpleDataset([(img,label) for img, label in watermarkset])
if config.train.name == "DI":
_, watermarkset = data.random_split(trainset, lengths=[trainset.__len__()-config.watermark_len,config.watermark_len], generator=torch.Generator().manual_seed(seed))
else:
_, watermarkset = data.random_split(testset, lengths=[testset.__len__()-config.watermark_len,config.watermark_len], generator=torch.Generator().manual_seed(seed))
trainloader = data.DataLoader(trainset, batch_size=config.train.batch_size, shuffle=True,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
testloader = data.DataLoader(testset, batch_size=config.train.batch_size,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
watermarkloader = data.DataLoader(watermarkset, batch_size=config.wm_batch_size,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
if adv == None:
trainloader = data.DataLoader(trainset, batch_size=config.train.batch_size, shuffle=True,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
elif adv:
trainset, _ = data.random_split(trainset, lengths=[25000,25000], generator=torch.Generator().manual_seed(seed))
trainloader = data.DataLoader(trainset, batch_size=config.train.batch_size, shuffle=True,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
if config.train.name == "DI":
_, watermarkset = data.random_split(trainset, lengths=[trainset.__len__()-config.watermark_len,config.watermark_len], generator=torch.Generator().manual_seed(seed))
watermarkloader = data.DataLoader(watermarkset, batch_size=config.wm_batch_size,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
else:
_, trainset = data.random_split(trainset, lengths=[25000,25000], generator=torch.Generator().manual_seed(seed))
trainloader = data.DataLoader(trainset, batch_size=config.train.batch_size, shuffle=True,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
return trainloader, testloader , watermarkloader
else:
root = config.dataset.root
workers=1
pin_memory=True
traindir = os.path.join(root, 'train')
valdir = os.path.join(root, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize
])
)
val_dataset = datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize
])
)
if config.train.name=="DI":
watermarkset, _ = data.random_split(train_dataset, lengths=[config.watermark_len, len(train_dataset)-config.watermark_len],
generator=torch.Generator().manual_seed(seed))
else:
watermarkset, _ = data.random_split(val_dataset, lengths=[config.watermark_len, len(val_dataset)-config.watermark_len],
generator=torch.Generator().manual_seed(seed))
watermarkloader = data.DataLoader(watermarkset, batch_size=config.wm_batch_size,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
val_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=config.train.batch_size,
shuffle=False,
num_workers=workers,
pin_memory=pin_memory
)
if adv == None:
trainloader = data.DataLoader(train_dataset, batch_size=config.train.batch_size, shuffle=True, num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
elif adv:
trainset, _ = data.random_split(train_dataset, lengths=[len(train_dataset)//2,len(train_dataset)-len(train_dataset)//2],
generator=torch.Generator().manual_seed(seed))
trainloader = data.DataLoader(trainset, batch_size=config.train.batch_size, shuffle=True, num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
else:
_, trainset = data.random_split(train_dataset, lengths=[len(train_dataset)//2,len(train_dataset)-len(train_dataset)//2],
generator=torch.Generator().manual_seed(seed))
trainloader = data.DataLoader(trainset, batch_size=config.train.batch_size, shuffle=True, num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
return trainloader, val_loader , watermarkloader
def load_model(config, model_num):
model_dir = config.modeldir + config.model.name + '/'+str(model_num) + '/'
ref_model=[]
for i in range(config.num_ref_model):
model = mlconfig.instantiate(config.model)
model.load_state_dict(torch.load(model_dir+"model{}.pth".format(i)))
model.eval()
ref_model.append(model.cuda())
return ref_model
def load_sur_model(config, model_num):
model_dir = config.modeldir + 'surrogate/' + config.model.name + '/'+str(model_num) + '/'
ref_model=[]
for i in range(config.num_sur_model):
model = mlconfig.instantiate(config.model)
model.load_state_dict(torch.load(model_dir+"model{}.pth".format(i)))
model.eval()
ref_model.append(model.cuda())
return ref_model
def test(model, loader, logfile=None):
model.eval()
criterion = nn.CrossEntropyLoss()
test_loss = 0
correct = 0
total = 0
model = model.cuda()
for batch_idx, (input, label) in enumerate(tqdm(loader, desc="Test model", unit='images'), 0):
if label.min()<0:
label = label.clamp_min(0).T[0]
input, label = input.cuda(), label.cuda()
outputs = model(input)
if type(outputs)==list:
outputs=outputs[0]
loss = criterion(outputs, label)
_, predict = torch.max(outputs.data, 1)
correct += predict.eq(label).cpu().sum()
total += label.size(0)
test_loss += loss.item()
print("Test result: ")
print('Loss: %.3f | Acc: %.3f%% (%d/%d)\n'
% (test_loss / (batch_idx + 1), 100. * correct / total, correct, total))
if logfile is not None:
with open(logfile, 'a') as f:
f.write('Test results:\n')
f.write('Loss: %.3f | Acc: %.3f%% (%d/%d)\n'
% (test_loss / (batch_idx + 1), 100. * correct / total, correct, total))
return 100.*correct/total
def test_watermark(model, loader, Text = "Test"):
model.eval()
criterion = nn.CrossEntropyLoss()
test_loss = 0
correct = 0
total = 0
model = model.cuda()
for batch_idx, (input, label) in enumerate(tqdm(loader, desc="Test model", unit='images'), 0):
if label.min()<0:
label = label.clamp_min(0).T[0]
input, label = input.cuda(), label.cuda().type(torch.long)
outputs = model(input)
if type(outputs)==list:
outputs=outputs[0]
loss = criterion(outputs, label)
_, predict = torch.max(outputs.data, 1)
correct += predict.eq(label).cpu().sum()
total += label.size(0)
test_loss += loss.item()
print("{} result: ".format(Text))
print('Loss: %.3f | Acc: %.3f%% (%d/%d)\n'
% (test_loss / (batch_idx + 1), 100. * correct / total, correct, total))
return 100.*correct/total
def random_perturb(x):
rnd = torch.randint(32,40,(1,)).item()
re_im = transforms.RandomCrop((rnd,rnd),pad_if_needed=True, padding_mode="edge")(x)
pad_top = torch.randint(0, 40-rnd,(1,)).item()
pad_bottom = 40-rnd-pad_top
pad_left = torch.randint(0,40-rnd,(1,)).item()
pad_right = 40-rnd-pad_left
re_im = transforms.RandomCrop((40,40),padding=(pad_top,pad_bottom,pad_left, pad_right), pad_if_needed=True)(re_im)
re_im = transforms.Resize((32,32))(re_im)
p = 0.5
if torch.rand((1,))>0.5:
return re_im
else:
return x
def arctanh(imgs):
scaling = torch.clamp(imgs, max=1, min=-1)
x = 0.999999 * scaling
return 0.5*torch.log((1+x)/(1-x))
def scaler(x_atanh):
return ((torch.tanh(x_atanh))+1) * 0.5
def comp_prob(output, target):
output = torch.softmax(output,0)
return output[target]/output.sum()
def extract_dataset(victim_model, trainloader, batch_size, config=None):
if trainloader.dataset.__len__()>50000:
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform = transforms.Compose([
transforms.Resize([224,192], interpolation=Image.BICUBIC),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
train_dataset= config.dataset(transform = transform)
trainset, testset = data.random_split(train_dataset, lengths=[int(train_dataset.__len__()*0.8),train_dataset.__len__()-int(train_dataset.__len__()*0.8)], generator=torch.Generator().manual_seed(0))
train_set = StealDataset(trainset, victim_model,config.train.batch_size)
trainloader = data.DataLoader(train_set, batch_size=config.train.batch_size,num_workers=multiprocessing.cpu_count()-2,pin_memory=True,drop_last=True)
return trainloader
else:
labels =[]
data_x = []
loop = tqdm(trainloader, desc='Predict labels',ncols=150)
acc = 0
total = 0
victim_model = victim_model.cuda()
victim_model.eval()
for idx, (x,y) in enumerate(loop):
x = x.cuda()
lab = np.argmax(victim_model(x).detach().cpu().numpy(), axis=1)
data_x.append(x.detach().cpu().numpy())
labels.append(lab)
acc += np.equal(lab,y).cpu().sum()
total += len(lab)
loop.set_postfix(ACC='Acc:{} ({}/{})'.format(100.*(acc/total), acc, total))
data_x = np.concatenate(data_x, axis=0)
labels= np.concatenate(labels)
train_set = SimpleDataset([(img, labb) for img, labb in zip(data_x, labels)])
trainloader = data.DataLoader(train_set, batch_size=batch_size,num_workers=multiprocessing.cpu_count()-2,pin_memory=True)
return trainloader
def train_sur(config, model, train_loader, n_epoch=None, logfile=None):
criterion = nn.CrossEntropyLoss()
optimizer = mlconfig.instantiate(config.sur_optimizer, model.parameters())
scheduler = mlconfig.instantiate(config.scheduler, optimizer)
correct = 0
total = 0
train_loss = 0
model = model.cuda()
model.train()
if n_epoch == None:
n_epoch = config.train.num_epoches
for epoch in range(n_epoch):
print('\nEpoch: %d' % epoch)
for batch_idx, (input, label) in enumerate(tqdm(train_loader, desc="Training reference model", unit='images'), 0):
input, label = input.cuda(), label.cuda()
optimizer.zero_grad()
outputs = model(input)
loss = criterion(outputs, label)
loss.backward()
optimizer.step()
_, predict = torch.max(outputs.data, 1)
correct += predict.eq(label).cpu().sum()
total += label.size(0)
train_loss += loss.item()
scheduler.step()
print("Train result: ")
print('Loss: %.3f | Acc: %.3f%% (%d/%d)\n'
% (train_loss / (batch_idx + 1), 100. * correct / total, correct, total))
if logfile is not None:
with open(logfile, "a+") as f:
f.write('Epoch:%d Loss: %.3f | Acc: %.3f%% (%d/%d)\n'
% (epoch, train_loss / (batch_idx + 1), 100. * correct / total, correct, total))