-
Notifications
You must be signed in to change notification settings - Fork 19
/
blackbox_attack.py
514 lines (433 loc) · 20.3 KB
/
blackbox_attack.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
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
import time
import random
import numpy as np
import torch
import torch.nn as nn
import torchvision.datasets as dsets
import torchvision.transforms as transforms
from torch.autograd import Variable
import torch.nn.functional as F
from models import IMAGENET, MNIST, CIFAR10, load_imagenet_data, load_mnist_data, load_cifar10_data, load_model, show_image
def attack_targeted(model, train_dataset, x0, y0, target, alpha = 0.1, beta = 0.001, iterations = 1000):
""" Attack the original image and return adversarial example of target t
model: (pytorch model)
train_dataset: set of training data
(x0, y0): original image
t: target
"""
if (model.predict(x0) != y0):
print("Fail to classify the image. No need to attack.")
return x0
# STEP I: find initial direction (theta, g_theta)
num_samples = 100
best_theta, g_theta = None, float('inf')
query_count = 0
print("Searching for the initial direction on %d samples: " % (num_samples))
timestart = time.time()
samples = set(random.sample(range(len(train_dataset)), num_samples))
for i, (xi, yi) in enumerate(train_dataset):
if i not in samples:
continue
query_count += 1
if model.predict(xi) == target:
theta = xi - x0
initial_lbd = torch.norm(theta)
theta = theta/torch.norm(theta)
lbd, count = fine_grained_binary_search_targeted(model, x0, y0, target, theta, initial_lbd)
query_count += count
if lbd < g_theta:
best_theta, g_theta = theta, lbd
print("--------> Found distortion %.4f" % g_theta)
timeend = time.time()
print("==========> Found best distortion %.4f in %.4f seconds using %d queries" % (g_theta, timeend-timestart, query_count))
# STEP II: seach for optimal
timestart = time.time()
g1 = 1.0
theta, g2 = best_theta.clone(), g_theta
opt_count = 0
for i in range(iterations):
gradient = torch.zeros(theta.size())
q = 10
min_g1 = float('inf')
for _ in range(q):
u = torch.randn(theta.size()).type(torch.FloatTensor)
u = u/torch.norm(u)
ttt = theta+beta * u
ttt = ttt/torch.norm(ttt)
g1, count = fine_grained_binary_search_local_targeted(model, x0, y0, target, ttt, initial_lbd = g2, tol=beta/500)
opt_count += count
gradient += (g1-g2)/beta * u
if g1 < min_g1:
min_g1 = g1
min_ttt = ttt
gradient = 1.0/q * gradient
if (i+1)%50 == 0:
print("Iteration %3d: g(theta + beta*u) = %.4f g(theta) = %.4f distortion %.4f num_queries %d" % (i+1, g1, g2, torch.norm(g2*theta), opt_count))
min_theta = theta
min_g2 = g2
for _ in range(15):
new_theta = theta - alpha * gradient
new_theta = new_theta/torch.norm(new_theta)
new_g2, count = fine_grained_binary_search_local_targeted(model, x0, y0, target, new_theta, initial_lbd = min_g2, tol=beta/500)
opt_count += count
alpha = alpha * 2
if new_g2 < min_g2:
min_theta = new_theta
min_g2 = new_g2
else:
break
if min_g2 >= g2:
for _ in range(15):
alpha = alpha * 0.25
new_theta = theta - alpha * gradient
new_theta = new_theta/torch.norm(new_theta)
new_g2, count = fine_grained_binary_search_local_targeted(model, x0, y0, target, new_theta, initial_lbd = min_g2, tol=beta/500)
opt_count += count
if new_g2 < g2:
min_theta = new_theta
min_g2 = new_g2
break
if min_g2 <= min_g1:
theta, g2 = min_theta, min_g2
else:
theta, g2 = min_ttt, min_g1
if g2 < g_theta:
best_theta, g_theta = theta.clone(), g2
#print(alpha)
if alpha < 1e-4:
alpha = 1.0
print("Warning: not moving, g2 %lf gtheta %lf" % (g2, g_theta))
beta = beta * 0.1
if (beta < 0.0005):
break
target = model.predict(x0 + g_theta*best_theta)
timeend = time.time()
print("\nAdversarial Example Found Successfully: distortion %.4f target %d queries %d \nTime: %.4f seconds" % (g_theta, target, query_count + opt_count, timeend-timestart))
return x0 + g_theta*best_theta
def fine_grained_binary_search_local_targeted(model, x0, y0, t, theta, initial_lbd = 1.0, tol=1e-5):
nquery = 0
lbd = initial_lbd
if model.predict(x0+lbd*theta) != t:
lbd_lo = lbd
lbd_hi = lbd*1.01
nquery += 1
while model.predict(x0+lbd_hi*theta) != t:
lbd_hi = lbd_hi*1.01
nquery += 1
if lbd_hi > 100:
return float('inf'), nquery
else:
lbd_hi = lbd
lbd_lo = lbd*0.99
nquery += 1
while model.predict(x0+lbd_lo*theta) == t:
lbd_lo = lbd_lo*0.99
nquery += 1
while (lbd_hi - lbd_lo) > tol:
lbd_mid = (lbd_lo + lbd_hi)/2.0
nquery += 1
if model.predict(x0 + lbd_mid*theta) == t:
lbd_hi = lbd_mid
else:
lbd_lo = lbd_mid
return lbd_hi, nquery
def fine_grained_binary_search_targeted(model, x0, y0, t, theta, initial_lbd = 1.0):
nquery = 0
lbd = initial_lbd
while model.predict(x0 + lbd*theta) != t:
lbd *= 1.05
nquery += 1
if lbd > 100:
return float('inf'), nquery
num_intervals = 100
lambdas = np.linspace(0.0, lbd, num_intervals)[1:]
lbd_hi = lbd
lbd_hi_index = 0
for i, lbd in enumerate(lambdas):
nquery += 1
if model.predict(x0 + lbd*theta) == t:
lbd_hi = lbd
lbd_hi_index = i
break
lbd_lo = lambdas[lbd_hi_index - 1]
while (lbd_hi - lbd_lo) > 1e-7:
lbd_mid = (lbd_lo + lbd_hi)/2.0
nquery += 1
if model.predict(x0 + lbd_mid*theta) == t:
lbd_hi = lbd_mid
else:
lbd_lo = lbd_mid
return lbd_hi, nquery
def attack_untargeted(model, train_dataset, x0, y0, alpha = 0.2, beta = 0.001, iterations = 1000):
""" Attack the original image and return adversarial example
model: (pytorch model)
train_dataset: set of training data
(x0, y0): original image
"""
if (model.predict(x0) != y0):
print("Fail to classify the image. No need to attack.")
return x0
num_samples = 1000
best_theta, g_theta = None, float('inf')
query_count = 0
print("Searching for the initial direction on %d samples: " % (num_samples))
timestart = time.time()
samples = set(random.sample(range(len(train_dataset)), num_samples))
for i, (xi, yi) in enumerate(train_dataset):
if i not in samples:
continue
query_count += 1
if model.predict(xi) != y0:
theta = xi - x0
initial_lbd = torch.norm(theta)
theta = theta/torch.norm(theta)
lbd, count = fine_grained_binary_search(model, x0, y0, theta, initial_lbd, g_theta)
query_count += count
if lbd < g_theta:
best_theta, g_theta = theta, lbd
print("--------> Found distortion %.4f" % g_theta)
timeend = time.time()
print("==========> Found best distortion %.4f in %.4f seconds using %d queries" % (g_theta, timeend-timestart, query_count))
timestart = time.time()
g1 = 1.0
theta, g2 = best_theta.clone(), g_theta
torch.manual_seed(0)
opt_count = 0
stopping = 0.01
prev_obj = 100000
for i in range(iterations):
gradient = torch.zeros(theta.size())
q = 10
min_g1 = float('inf')
for _ in range(q):
u = torch.randn(theta.size()).type(torch.FloatTensor)
u = u/torch.norm(u)
ttt = theta+beta * u
ttt = ttt/torch.norm(ttt)
g1, count = fine_grained_binary_search_local(model, x0, y0, ttt, initial_lbd = g2, tol=beta/500)
opt_count += count
gradient += (g1-g2)/beta * u
if g1 < min_g1:
min_g1 = g1
min_ttt = ttt
gradient = 1.0/q * gradient
if (i+1)%50 == 0:
print("Iteration %3d: g(theta + beta*u) = %.4f g(theta) = %.4f distortion %.4f num_queries %d" % (i+1, g1, g2, torch.norm(g2*theta), opt_count))
if g2 > prev_obj-stopping:
break
prev_obj = g2
min_theta = theta
min_g2 = g2
for _ in range(15):
new_theta = theta - alpha * gradient
new_theta = new_theta/torch.norm(new_theta)
new_g2, count = fine_grained_binary_search_local(model, x0, y0, new_theta, initial_lbd = min_g2, tol=beta/500)
opt_count += count
alpha = alpha * 2
if new_g2 < min_g2:
min_theta = new_theta
min_g2 = new_g2
else:
break
if min_g2 >= g2:
for _ in range(15):
alpha = alpha * 0.25
new_theta = theta - alpha * gradient
new_theta = new_theta/torch.norm(new_theta)
new_g2, count = fine_grained_binary_search_local(model, x0, y0, new_theta, initial_lbd = min_g2, tol=beta/500)
opt_count += count
if new_g2 < g2:
min_theta = new_theta
min_g2 = new_g2
break
if min_g2 <= min_g1:
theta, g2 = min_theta, min_g2
else:
theta, g2 = min_ttt, min_g1
if g2 < g_theta:
best_theta, g_theta = theta.clone(), g2
#print(alpha)
if alpha < 1e-4:
alpha = 1.0
print("Warning: not moving, g2 %lf gtheta %lf" % (g2, g_theta))
beta = beta * 0.1
if (beta < 0.0005):
break
target = model.predict(x0 + g_theta*best_theta)
timeend = time.time()
print("\nAdversarial Example Found Successfully: distortion %.4f target %d queries %d \nTime: %.4f seconds" % (g_theta, target, query_count + opt_count, timeend-timestart))
return x0 + g_theta*best_theta
def fine_grained_binary_search_local(model, x0, y0, theta, initial_lbd = 1.0, tol=1e-5):
nquery = 0
lbd = initial_lbd
if model.predict(x0+lbd*theta) == y0:
lbd_lo = lbd
lbd_hi = lbd*1.01
nquery += 1
while model.predict(x0+lbd_hi*theta) == y0:
lbd_hi = lbd_hi*1.01
nquery += 1
if lbd_hi > 20:
return float('inf'), nquery
else:
lbd_hi = lbd
lbd_lo = lbd*0.99
nquery += 1
while model.predict(x0+lbd_lo*theta) != y0 :
lbd_lo = lbd_lo*0.99
nquery += 1
while (lbd_hi - lbd_lo) > tol:
lbd_mid = (lbd_lo + lbd_hi)/2.0
nquery += 1
if model.predict(x0 + lbd_mid*theta) != y0:
lbd_hi = lbd_mid
else:
lbd_lo = lbd_mid
return lbd_hi, nquery
def fine_grained_binary_search(model, x0, y0, theta, initial_lbd, current_best):
nquery = 0
if initial_lbd > current_best:
if model.predict(x0+current_best*theta) == y0:
nquery += 1
return float('inf'), nquery
lbd = current_best
else:
lbd = initial_lbd
## original version
#lbd = initial_lbd
#while model.predict(x0 + lbd*theta) == y0:
# lbd *= 2
# nquery += 1
# if lbd > 100:
# return float('inf'), nquery
#num_intervals = 100
# lambdas = np.linspace(0.0, lbd, num_intervals)[1:]
# lbd_hi = lbd
# lbd_hi_index = 0
# for i, lbd in enumerate(lambdas):
# nquery += 1
# if model.predict(x0 + lbd*theta) != y0:
# lbd_hi = lbd
# lbd_hi_index = i
# break
# lbd_lo = lambdas[lbd_hi_index - 1]
lbd_hi = lbd
lbd_lo = 0.0
while (lbd_hi - lbd_lo) > 1e-5:
lbd_mid = (lbd_lo + lbd_hi)/2.0
nquery += 1
if model.predict(x0 + lbd_mid*theta) != y0:
lbd_hi = lbd_mid
else:
lbd_lo = lbd_mid
return lbd_hi, nquery
def attack_mnist(alpha=0.2, beta=0.001, isTarget= False, num_attacks= 100):
train_loader, test_loader, train_dataset, test_dataset = load_mnist_data()
print("Length of test_set: ", len(test_dataset))
dataset = train_dataset
net = MNIST()
if torch.cuda.is_available():
net.cuda()
net = torch.nn.DataParallel(net, device_ids=[0])
load_model(net, 'models/mnist_gpu.pt')
#load_model(net, 'models/mnist_cpu.pt')
net.eval()
model = net.module if torch.cuda.is_available() else net
def single_attack(image, label, target = None):
show_image(image.numpy())
print("Original label: ", label)
print("Predicted label: ", model.predict(image))
if target == None:
adversarial = attack_untargeted(model, dataset, image, label, alpha = alpha, beta = beta, iterations = 1000)
else:
print("Targeted attack: %d" % target)
adversarial = attack_targeted(model, dataset, image, label, target, alpha = alpha, beta = beta, iterations = 1000)
show_image(adversarial.numpy())
print("Predicted label for adversarial example: ", model.predict(adversarial))
return torch.norm(adversarial - image)
print("\n\n Running {} attack on {} random MNIST test images for alpha= {} beta= {}\n\n".format("targetted" if isTarget else "untargetted", num_attacks, alpha, beta))
total_distortion = 0.0
samples = [6312, 6891, 4243, 8377, 7962, 6635, 4970, 7809, 5867, 9559, 3579, 8269, 2282, 4618, 2290, 1554, 4105, 9862, 2408, 5082, 1619, 1209, 5410, 7736, 9172, 1650, 5181, 3351, 9053, 7816, 7254, 8542, 4268, 1021, 8990, 231, 1529, 6535, 19, 8087, 5459, 3997, 5329, 1032, 3131, 9299, 3910, 2335, 8897, 7340, 1495, 5244,8323, 8017, 1787, 4939, 9032, 4770, 2045, 8970, 5452, 8853, 3330, 9883, 8966, 9628, 4713, 7291, 9770, 6307, 5195, 9432, 3967, 4757, 3013, 3103, 3060, 541, 4261, 7808, 1132, 1472, 2134, 634, 1315, 8858, 6411, 8595, 4516, 8550, 3859, 3526]
#true_labels = [3, 1, 6, 6, 9, 2, 7, 5, 5, 3, 3, 4, 5, 6, 7, 9, 1, 6, 3, 4, 0, 6, 5, 9, 7, 0, 3, 1, 6, 6, 9, 6, 4, 7, 6, 3, 4, 3, 4, 3, 0, 7, 3, 5, 3, 9, 3, 1, 9, 1, 3, 0, 2, 9, 9, 2, 2, 3, 3, 3, 0, 5, 2, 5, 2, 7, 2, 2, 5, 7, 4, 9, 9, 0, 0, 7, 9, 4, 5, 5, 2, 3, 5, 9, 3, 0, 9, 0, 1, 2, 9, 9]
for idx in samples:
#idx = random.randint(100, len(test_dataset)-1)
image, label = test_dataset[idx]
print("\n\n\n\n======== Image %d =========" % idx)
#target = None if not isTarget else random.choice(list(range(label)) + list(range(label+1, 10)))
target = None if not isTarget else (1+label) % 10
total_distortion += single_attack(image, label, target)
print("Average distortion on random {} images is {}".format(num_attacks, total_distortion/num_attacks))
def attack_cifar10(alpha= 0.2, beta= 0.001, isTarget= False, num_attacks= 100):
train_loader, test_loader, train_dataset, test_dataset = load_cifar10_data()
dataset = train_dataset
print("Length of test_set: ", len(test_dataset))
net = CIFAR10()
if torch.cuda.is_available():
net.cuda()
net = torch.nn.DataParallel(net, device_ids=[0])
load_model(net, 'models/cifar10_gpu.pt')
#load_model(net, 'models/cifar10_cpu.pt')
net.eval()
model = net.module if torch.cuda.is_available() else net
def single_attack(image, label, target = None):
print("Original label: ", label)
print("Predicted label: ", model.predict(image))
if target == None:
adversarial = attack_untargeted(model, dataset, image, label, alpha = alpha, beta = beta, iterations = 1000)
else:
print("Targeted attack: %d" % target)
adversarial = attack_targeted(model, dataset, image, label, target, alpha = alpha, beta = beta, iterations = 1000)
print("Predicted label for adversarial example: ", model.predict(adversarial))
return torch.norm(adversarial - image)
print("\n\nRunning {} attack on {} random CIFAR10 test images for alpha= {} beta= {}\n\n".format("targetted" if isTarget else "untargetted", num_attacks, alpha, beta))
total_distortion = 0.0
samples = [6311, 6890, 663, 4242, 8376, 7961, 6634, 4969, 7808, 5866, 9558, 3578, 8268, 2281, 2289, 1553, 4104, 8725, 9861, 2407, 5081, 1618, 1208, 5409, 7735, 9171, 1649, 5796, 7113, 5180, 3350,9052, 7253, 8541, 4267, 1020, 8989, 230, 1528, 6534, 18, 8086, 3996, 1031, 3130, 9298, 3632, 3909, 2334, 8896, 7339, 1494, 5243, 8322, 8016, 1786, 9031, 4769, 8969, 5451, 8852, 3329, 9882, 8965, 9627, 4712, 7290, 9769, 6306, 5194, 3966, 4756, 3012, 3102, 540, 4260, 7807, 1471, 2133, 2450, 633, 1314, 8857, 6410, 8594, 4515, 8549, 3858, 3525, 6411, 4360, 7753, 7413, 684,3343, 6785, 7079, 2263]
#true_labels = [3, 5, 6, 8, 7, 3, 4, 1, 8, 4, 0, 7, 5, 5, 1, 4, 0, 8, 6, 9, 5, 7, 3, 1, 4, 2, 5, 5, 9, 9, 8, 0, 4, 8, 7, 1, 4, 5, 2, 7, 8, 4, 6, 3, 3, 1, 1, 5, 1, 8, 6, 7, 1, 4, 4, 1, 0, 8, 8, 6, 7, 3, 1, 4, 4, 4, 6, 8, 0, 7, 4, 6, 1, 0, 1, 8, 3, 8, 3, 1, 8, 9, 0, 1, 3, 0, 1, 8, 2, 8, 6, 9, 1, 9, 3, 6, 7, 6]
#samples = [7753, 1314, 633]
samples = [6311]
for idx in samples:
#idx = random.randint(100, len(test_dataset)-1)
image, label = test_dataset[idx]
print("\n\n\n\n======== Image %d =========" % idx)
#target = None if not isTarget else random.choice(list(range(label)) + list(range(label+1, 10)))
target = None if not isTarget else (1+label) % 10
total_distortion += single_attack(image, label, target)
print("Average distortion on random {} images is {}".format(num_attacks, total_distortion/num_attacks))
def attack_imagenet(arch='resnet50', alpha=0.2, beta= 0.001, isTarget=False, num_attacks = 100):
train_loader, test_loader, train_dataset, test_dataset = load_imagenet_data()
dataset = test_dataset
print("Length of test_set: ", len(test_dataset))
model = IMAGENET(arch)
def attack_single(image, label, target = None):
print("Original label: ", label)
print("Predicted label: ", model.predict(image))
if target == None:
adversarial = attack_untargeted(model, dataset, image, label, alpha = alpha, beta = beta, iterations = 1500)
else:
print("Targeted attack: %d" % target)
adversarial = attack_targeted(model, dataset, image, label, target, alpha = alpha, beta = beta, iterations = 1500)
print("Predicted label for adversarial example: ", model.predict(adversarial))
return torch.norm(adversarial - image)
print("\nRunning {} attack on {} random IMAGENET test images for alpha= {} beta= {} using {}\n".format("targetted" if isTarget else "untargetted", num_attacks, alpha, beta, arch))
total_distortion = 0.0
samples = [25248, 27563, 2654, 16969, 31846, 26538, 19878, 14316, 33076, 9128, 9159, 49533, 34903, 46215, 963220326, 6473, 483344826,216406600, 23187, 40036, 41971, 13401, 36211, 31262, 4082, 35960, 6113, 47167, 46548, 75, 40102, 32348, 21313, 46114, 4128,37193, 14530, 9339, 5978, 20976, 33289]
for idx in samples:
#idx = random.randint(100, len(test_dataset)-1)
image, label = test_dataset[idx]
print("\n\n======== Image %d =========" % idx)
target = None if not isTarget else random.choice(list(range(label)) + list(range(label+1, 1000)))
total_distortion += attack_single(image, label, target)
print("Average distortion on random {} images is {}".format(num_attacks, total_distortion/num_attacks))
if __name__ == '__main__':
timestart = time.time()
random.seed(0)
#attack_mnist(alpha=2, beta=0.005, isTarget= False)
attack_cifar10(alpha=5, beta=0.001, isTarget= False)
#attack_imagenet(arch='resnet50', alpha=10, beta=0.005, isTarget= False)
#attack_imagenet(arch='vgg19', alpha=0.05, beta=0.001, isTarget= False, num_attacks= 10)
#attack_mnist(alpha=2, beta=0.005, isTarget= True)
#attack_cifar10(alpha=5, beta=0.001, isTarget= True)
#attack_imagenet(arch='resnet50', alpha=10, beta=0.005, isTarget= True)
#attack_imagenet(arch='vgg19', alpha=0.05, beta=0.001, isTarget= True, num_attacks= 10)
timeend = time.time()
print("\n\nTotal running time: %.4f seconds\n" % (timeend - timestart))