-
Notifications
You must be signed in to change notification settings - Fork 2
/
evaluate.py
456 lines (401 loc) · 15.4 KB
/
evaluate.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 argparse
import logging
import sys
import os
import time
import math
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.distributions import laplace, uniform
from torchvision import datasets, transforms
from torch_backend import *
from tqdm import tqdm
import random
from preact_resnet import resnet50 as ResNet50
from wideresnet import WideResNet
import foolbox as fb
import foolbox.attacks as fa
from autoattack import AutoAttack
logger = logging.getLogger(__name__)
logging.basicConfig(format='[%(asctime)s] - %(message)s',
datefmt='%Y/%m/%d %H:%M:%S',
level=logging.DEBUG)
cifar10_mean = (0.0, 0.0, 0.0)
cifar10_std = (1.0, 1.0, 1.0)
mu = torch.tensor(cifar10_mean).view(3, 1, 1).cuda()
std = torch.tensor(cifar10_std).view(3, 1, 1).cuda()
upper_limit = ((1 - mu) / std)
lower_limit = ((0 - mu) / std)
def clamp(X, lower_limit, upper_limit):
return torch.max(torch.min(X, upper_limit), lower_limit)
def norms(Z):
return Z.view(Z.shape[0], -1).norm(dim=1)[:, None, None, None]
def norms_p(Z, order):
return Z.view(Z.shape[0], -1).norm(p=order, dim=1)[:, None, None, None]
def norms_l1(Z):
return Z.view(Z.shape[0], -1).abs().sum(dim=1)[:, None, None, None]
def l1_dir_topk(grad, delta, X, k=20):
X_curr = X + delta
batch_size = X.shape[0]
channels = X.shape[1]
pix = X.shape[2]
grad = grad.detach().cpu().numpy()
abs_grad = np.abs(grad)
sign = np.sign(grad)
max_abs_grad = np.percentile(abs_grad, k, axis=(1, 2, 3), keepdims=True)
tied_for_max = (abs_grad >= max_abs_grad).astype(np.float32)
num_ties = np.sum(tied_for_max, (1, 2, 3), keepdims=True)
optimal_perturbation = sign * tied_for_max / num_ties
optimal_perturbation = torch.from_numpy(optimal_perturbation).to(device)
return optimal_perturbation.view(batch_size, channels, pix, pix)
def proj_l1ball(x, epsilon=12, device="cuda:1"):
assert epsilon > 0
u = x.abs()
if (u.sum(dim=(1, 2, 3)) <= epsilon).all():
return x
y = proj_simplex(u, s=epsilon, device=device)
y = y.view(-1, 3, x.shape[2], x.shape[3])
y *= x.sign()
return y
def proj_simplex(v, s=1, device="cuda:1"):
assert s > 0, "Radius s must be strictly positive (%d <= 0)" % s
batch_size = v.shape[0]
u = v.view(batch_size, 1, -1)
n = u.shape[2]
u, indices = torch.sort(u, descending=True)
cssv = u.cumsum(dim=2)
vec = u * torch.arange(1, n + 1).float().to(device)
comp = (vec > (cssv - s)).half()
u = comp.cumsum(dim=2)
w = (comp - 1).cumsum(dim=2)
u = u + w
rho = torch.argmax(u, dim=2)
rho = rho.view(batch_size)
c = torch.HalfTensor([cssv[i, 0, rho[i]]
for i in range(cssv.shape[0])]).to(device)
c = c - s
theta = torch.div(c.float(), (rho.float() + 1))
theta = theta.view(batch_size, 1, 1, 1)
w = (v.float() - theta).clamp(min=0)
return w
def attack_pgd(model, X, y, norm, dataset, restarts=1, version=0):
max_loss = torch.zeros(y.shape[0]).cuda()
max_delta = torch.zeros_like(X).cuda()
for _ in range(restarts):
delta = torch.zeros_like(X).cuda()
if norm == "linf":
if dataset != "tinyimagenet":
epsilon = (8 / 255.) / std
else:
epsilon = (4 / 255.) / std
attack_iters = 50
alpha = (1. / 255.) / std
delta[:, 0, :, :].uniform_(-epsilon[0][0][0].item(),
epsilon[0][0][0].item())
delta[:, 1, :, :].uniform_(-epsilon[1][0][0].item(),
epsilon[1][0][0].item())
delta[:, 2, :, :].uniform_(-epsilon[2][0][0].item(),
epsilon[2][0][0].item())
elif norm == "l2":
if dataset != "tinyimagenet":
epsilon = (128. / 255.) / std
else:
epsilon = (80 / 255.) / std
attack_iters = 50
alpha = (25. / 255.) / std
delta = torch.rand_like(X, requires_grad=True)
delta.data *= (2.0 * delta.data - 1.0) * epsilon
delta.data /= norms_p(
delta.detach(), 2.0).clamp(min=epsilon.detach().cpu().numpy()[0][0][0])
elif norm == "l1":
epsilon = (2000 / 255.) / std
attack_iters = 100
alpha = (255. / 255.) / std
ini = laplace.Laplace(loc=delta.new_tensor(0), scale=delta.new_tensor(1))
delta.data = ini.sample(delta.data.shape)
delta.data = (2.0 * delta.data - 1.0) * epsilon
delta.data /= norms_l1(delta.detach()).clamp(min=epsilon.detach().cpu().numpy()[0][0][0])
delta.requires_grad = True
for _ in range(attack_iters):
output = model(X + delta)
incorrect = output.max(1)[1] != y
correct = (~incorrect).unsqueeze(1).unsqueeze(1).unsqueeze(1).half()
correct = 1.0 if version == 0 else correct
loss = F.cross_entropy(output, y)
loss.backward()
grad = delta.grad.detach()
if norm == "linf":
delta.data = clamp(delta.data + correct * alpha * torch.sign(grad),
-epsilon, epsilon)
elif norm == "l2":
delta.data = delta.data + correct * alpha * grad / norms(grad)
delta.data *= epsilon / norms_p(delta.detach(), 2.0).clamp(
min=epsilon.detach().cpu().numpy()[0][0][0])
elif norm == "l1":
k = 99
delta.data = delta.data + correct * alpha * l1_dir_topk(
grad, delta.data, X, k)
if (norms_l1(delta) > epsilon).any():
delta.data = proj_l1ball(delta.data,
epsilon=epsilon.detach().cpu().numpy()[0][0][0],
device=device)
delta.data = clamp(delta.data, lower_limit - X, upper_limit - X)
delta.grad.zero_()
all_loss = F.cross_entropy(model(X + delta), y, reduction='none')
return delta.detach()
def get_loaders(dir_, batch_size, dataset):
if dataset == "cifar10":
train_transform = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(cifar10_mean, cifar10_std)
])
test_transform = transforms.Compose([transforms.ToTensor()])
elif dataset == "svhn":
train_transform = transforms.Compose([transforms.ToTensor()])
test_transform = transforms.Compose([transforms.ToTensor()])
elif dataset == "tinyimagenet":
train_transform = transforms.Compose([
transforms.RandomCrop(64, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor()
])
test_transform = transforms.Compose([transforms.ToTensor()])
num_workers = 2
if dataset == "cifar10":
train_dataset = datasets.CIFAR10(dir_,
train=True,
transform=train_transform,
download=True)
test_dataset = datasets.CIFAR10(dir_,
train=False,
transform=test_transform,
download=True)
elif dataset == "svhn":
train_dataset = datasets.SVHN(dir_,
split='train',
transform=train_transform,
download=True)
test_dataset = datasets.SVHN(dir_,
split='test',
transform=test_transform,
download=True)
elif dataset == "tinyimagenet":
train_dataset = torchvision.datasets.ImageFolder(root=dir_ + '/train',
transform=train_transform)
test_dataset = torchvision.datasets.ImageFolder(root=dir_ + '/val',
transform=test_transform)
train_loader = torch.utils.data.DataLoader(
dataset=train_dataset,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=num_workers,
)
test_loader = torch.utils.data.DataLoader(
dataset=test_dataset,
batch_size=batch_size,
shuffle=False,
pin_memory=True,
num_workers=2,
)
return train_loader, test_loader
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument('--batch-size', default=26, type=int)
parser.add_argument('--data-dir', default='../cifar-data', type=str)
parser.add_argument('--dataset', default='cifar100', type=str)
parser.add_argument('--fname', default='cifar_linf_pgd', type=str)
parser.add_argument('--attack',
default='pgd',
type=str,
choices=['pgd', 'fgsm', 'ddn', 'none'])
parser.add_argument('--attack_lib', default='custom', type=str)
parser.add_argument('--norm', default='linf', type=str)
parser.add_argument('--epsilon', default=8, type=float)
parser.add_argument('--attack-iters', default=200, type=int)
parser.add_argument('--n_classes', default=10, type=int)
parser.add_argument('--alpha', default=1, type=int)
parser.add_argument('--restarts', default=1, type=int)
parser.add_argument('--seed', default=0, type=int)
parser.add_argument('--width-factor', default=10, type=int)
parser.add_argument('--model', default='WideResNet')
parser.add_argument('--model_type', default='linf')
parser.add_argument('--use_cal', default=True, type=bool)
return parser.parse_args()
def get_attack(fmodel, attack, init_attack=None):
# L0
if attack == 'SAPA':
A = fa.SaltAndPepperNoiseAttack()
elif attack == 'EAD':
A = fa.EADAttack(decision_rule='L1', binary_search_steps=5, steps=1000)
elif attack == "PGDL1":
A = fa.SparseL1DescentAttack()
# L2
elif 'PGDL2' in attack:
A = fa.L2PGD()
elif attack == 'AGNA':
A = fa.L2RepeatedAdditiveGaussianNoiseAttack()
elif attack == "CWL2":
A = fa.L2CarliniWagnerAttack(binary_search_steps=5, steps=200)
elif attack == "BBL2":
A = fa.L2BrendelBethgeAttack(init_attack=init_attack)
# L inf
elif 'FGSM' in attack and not 'IFGSM' in attack:
A = fa.FGSM()
elif 'PGDLinf' in attack:
A = fa.LinfPGD()
elif 'MIM' in attack:
A = fa.MomentumIterativeAttack()
elif attack == "BBLinf":
A = fa.LinfinityBrendelBethgeAttack(init_attack=init_attack)
else:
raise Exception('Not implemented')
return A
def main():
args = get_args()
logger.info(args)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
args.data_dir = args.dataset + "-data"
train_loader, test_loader = get_loaders(args.data_dir, args.batch_size,
args.dataset)
epsilon = (args.epsilon / 255.) / std
alpha = (args.alpha / 255.) / std
if args.model == 'resnet50':
model = ResNet50().cuda()
elif args.model == 'WideResNet':
model = WideResNet(28, 10, widen_factor=args.width_factor, dropRate=0.0)
else:
raise ValueError("Unknown model")
model = torch.nn.DataParallel(model).cuda()
checkpoint = torch.load(args.fname + args.model_type + '.pth',
map_location='cuda:0')
model.load_state_dict(checkpoint)
model.eval()
model.float()
total_adv_loss = 0
total_adv_acc = 0
total_clean_loss = 0
total_clean_acc = 0
total = 0
if args.attack_lib == 'custom':
x_adv = []
total = 0
max_check = 1000
output = np.ones((max_check + 1))
for i, (X, y) in tqdm(enumerate(test_loader)):
X, y = X.cuda(), y.cuda()
if args.attack == 'pgd':
image = X[0, :, :, :].view(1, 3, X.shape[2], X.shape[3])
label = y[0].long().view(-1)
delta = attack_pgd(model, image, label, args.norm, args.dataset)
with torch.no_grad():
output_adv = model(image + delta)
output_clean = model(image)
total_adv_acc += (output_adv.max(1)[1] == label).item()
total_clean_acc += (output_clean.max(1)[1] == label).item()
output[total] = (output_adv.max(1)[1] == label).cpu().numpy().astype(
np.float32)
total += 1
if (total >= max_check):
break
print('Test Adversarial Acc: %.4f' %(total_adv_acc / total))
print('Test Clean Acc: %.4f' %(total_clean_acc / total))
elif args.attack_lib == "foolbox":
fmodel = fb.PyTorchModel(model, bounds=(0, 1), device=device)
attacks_list = [
"PGDLinf", "BBLinf", "PGDL1", "EAD", "SAPA", 'BBL2', 'CWL2', 'AGNA',
'PGDL2'
]
for j in range(len(attacks_list)):
start = time.time()
attack_name = attacks_list[j]
if attack_name in linf_attacks:
if args.dataset == "tinyimagenet":
epsilons = [4. / 255]
else:
epsilons = [8. / 255]
elif attack_name in l2_attacks:
if args.dataset == "tinyimagenet":
epsilons = [80. / 255]
else:
epsilons = [128. / 255]
elif attack_name in l1_attacks:
epsilons = [2000. / 255]
total = 0
robust_accuracy = 0
total_adv_acc = 0
max_check = 1000
output = np.ones((max_check + 1))
images, labels = iter(test_loader).next()
images, labels = images.cuda(), labels.cuda()
batches = [(images[:10], labels[:10])]
init_attack = fb.attacks.DatasetAttack()
init_attack.feed(fmodel, batches[0][0]) # feed 1st batch of inputs
if args.dataset != "tinyimagenet":
attack = get_attack(fmodel, attacks_list[j], init_attack)
else:
attack = get_attack(fmodel, attacks_list[j])
for i, (X, y) in tqdm(enumerate(test_loader)):
X, y = X.cuda(), y.cuda()
image = X[0, :, :, :].view(1, 3, X.shape[2], X.shape[3])
label = y[0].long().view(-1)
try:
advs, _, success = attack(fmodel, image, label, epsilons=epsilons)
success = success.cpu().numpy().astype(np.float32)
robust_accuracy += 1 - success.mean(axis=-1)
output[total] = 1 - success.mean(axis=-1)
total += 1
output_adv = model(advs[0])
# make_model_diagrams(output_adv, label)
except Exception as e:
output[total] = 1.0
print("assertion error", e)
total += 1
continue
if (total >= max_check):
break
print("Robust accuracy for attack %s: %.4f %.4f" %
(attacks_list[j], robust_accuracy / total, total))
elif args.attack_lib == "autoattack":
if args.norm == "linf":
if args.dataset == "tinyimagenet":
epsilon = 4. / 255
else:
epsilon = 8. / 255
adversary = AutoAttack(model,
norm='Linf',
eps=epsilon,
version='standard')
elif args.norm == "l2":
if args.dataset == "tinyimagenet":
epsilon = 80. / 255
else:
epsilon = 128. / 255
adversary = AutoAttack(model, norm='L2', eps=epsilon, version='standard')
max_check = 1000
output = np.ones((max_check + 1))
total = 0
for i, (X, y) in tqdm(enumerate(test_loader)):
X, y = X.cuda(), y.cuda()
image = X[0, :, :, :].view(1, 3, X.shape[2], X.shape[3])
label = y[0].long().view(-1)
adv_X = adversary.run_standard_evaluation(image,
label,
bs=image.shape[0])
output_adv = model(adv_X)
loss_adv = F.cross_entropy(output_adv, label)
total_adv_acc += (output_adv.max(1)[1] == label).item()
output[total] = (output_adv.max(1)[1] == label).cpu().numpy().astype(
np.float32)
total += 1
if (total >= max_check):
break
print('Test Adversarial Acc: %.4f' % (total_adv_acc / total))
if __name__ == "__main__":
main()