-
Notifications
You must be signed in to change notification settings - Fork 5
/
attack_lib.py
288 lines (250 loc) · 11.5 KB
/
attack_lib.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
import torch
import numpy as np
import torchvision
import torchvision.transforms as transforms
from utils import BinaryDataset
from imagenet_dataset import DogCatDataset, DogFishDataset
from spam_dataset import SpamDataset
def attack_setting(args, test_label_poison=True, ret_testset=False):
if args['dataset'] == 'mnist':
N_EPOCH=20
BATCH_SIZE = 128
LR = 1e-3
if args['pair_id'] == 0:
POS_LABEL, NEG_LABEL = 1, 0
elif args['pair_id'] == 1:
POS_LABEL, NEG_LABEL = 6, 8
else:
raise NotImplementedError()
if args['atk_method'] == 'onepixel':
trigger_func = MNIST_onepixel_triggerfunc(args['delta'])
elif args['atk_method'] == 'fourpixel':
trigger_func = MNIST_fourpixel_triggerfunc(args['delta'])
elif args['atk_method'] == 'blending':
trigger_func = MNIST_blending_triggerfunc(args['delta'])
else:
raise NotImplementedError()
transform = transforms.Compose([
transforms.ToTensor(),
])
trainset = torchvision.datasets.MNIST(root='./raw_data/', train=True, download=True, transform=transform)
testset = torchvision.datasets.MNIST(root='./raw_data/', train=False, download=False, transform=transform)
from mnist_cnn_model import Model
elif args['dataset'] == 'cifar':
N_EPOCH = 10
BATCH_SIZE = 64
LR = 1e-3
if args['pair_id'] == 0:
POS_LABEL, NEG_LABEL = 0, 2 # airplane -> bird
elif args['pair_id'] == 1:
POS_LABEL, NEG_LABEL = 1, 5 # automobile -> dog
else:
raise NotImplementedError()
if args['atk_method'] == 'onepixel':
trigger_func = CIFAR_onepixeladd_allchannel_triggerfunc(args['delta'])
elif args['atk_method'] == 'fourpixel':
trigger_func = CIFAR_fourpixeladd_allchannel_triggerfunc(args['delta'])
elif args['atk_method'] == 'blending':
trigger_func = CIFAR_blending_triggerfunc(args['delta'])
else:
raise NotImplementedError()
transform = transforms.Compose([
transforms.ToTensor(),
])
trainset = torchvision.datasets.CIFAR10(root='./raw_data/', train=True, download=True, transform=transform)
testset = torchvision.datasets.CIFAR10(root='./raw_data/', train=False, download=False, transform=transform)
from cifar10_cnn_model import Model
elif args['dataset'] == 'imagenet':
N_EPOCH = 5
if args['dldp_sigma'] > 0: # DLDP requires more memory.
BATCH_SIZE = 32
else:
BATCH_SIZE = 64
LR = 1e-4
if args['atk_method'] == 'onepixel':
trigger_func = imagenet_onepixeladd_allchannel_triggerfunc(args['delta'])
elif args['atk_method'] == 'fourpixel':
trigger_func = imagenet_fourpixeladd_allchannel_triggerfunc(args['delta'])
elif args['atk_method'] == 'blending':
trigger_func = imagenet_blending_triggerfunc(args['delta'])
else:
raise NotImplementedError()
transform = transforms.Compose([
transforms.Resize((224,224)),
transforms.ToTensor(),
])
if args['pair_id'] == 0:
trainset = DogCatDataset(train=True, transform=transform)
testset = DogCatDataset(train=False, transform=transform)
elif args['pair_id'] == 1:
trainset = DogFishDataset(train=True, transform=transform)
testset = DogFishDataset(train=False, transform=transform)
from imagenet_dnn_model import Model
elif args['dataset'] == 'spam':
# Many variables are None because we only use this dataset in the evaluation part of KNN.
N_EPOCH = None
BATCH_SIZE = 64
LR = None
if args['atk_method'] == 'onepixel':
trigger_func = spam_onepixeladd_allchannel_triggerfunc(args['delta'])
elif args['atk_method'] == 'fourpixel':
trigger_func = spam_fourpixeladd_allchannel_triggerfunc(args['delta'])
elif args['atk_method'] == 'blending':
trigger_func = spam_blending_triggerfunc(args['delta'])
else:
raise NotImplementedError()
trainset = SpamDataset(train=True)
testset = SpamDataset(train=False)
Model = None
else:
raise NotImplementedError()
if args['dataset'] not in ('imagenet', 'spam'): # Change to binary dataset
trainset = BinaryDataset(trainset, POS_LABEL, NEG_LABEL)
testset = BinaryDataset(testset, POS_LABEL, NEG_LABEL)
TGT_CLASS = 0 # Target class in backdoor attack.
poisoned_train = BackdoorDataset(trainset, trigger_func, TGT_CLASS, args['poison_r'])
if test_label_poison:
poisoned_test = BackdoorDataset(testset, trigger_func, TGT_CLASS)
else:
# Use only non-target class images and do not poison the label
nontgt_idx = [i for i in range(len(testset)) if testset[i][1] != TGT_CLASS]
nontgt_testset = torch.utils.data.Subset(testset, nontgt_idx)
poisoned_test = BackdoorDataset(nontgt_testset, trigger_func, None)
testloader_benign = torch.utils.data.DataLoader(testset, batch_size=BATCH_SIZE)
testloader_poison = torch.utils.data.DataLoader(poisoned_test, batch_size=BATCH_SIZE)
if ret_testset:
return poisoned_train, testset, testloader_benign, testloader_poison, BATCH_SIZE, N_EPOCH, LR, Model
return poisoned_train, testloader_benign, testloader_poison, BATCH_SIZE, N_EPOCH, LR, Model
def MNIST_onepixel_triggerfunc(delta):
def MNIST_onepixel(X):
#X[:,20,20] = min(X[:,20,20]+delta, 1)
X[:,23,23] = min(X[:,23,23]+delta, 1)
return X
return MNIST_onepixel
def MNIST_fourpixel_triggerfunc(delta):
def MNIST_fourpixel(X):
X[:,18,20] = min(X[:,18,20]+delta/np.sqrt(4), 1)
X[:,19,19] = min(X[:,19,19]+delta/np.sqrt(4), 1)
X[:,20,18] = min(X[:,20,18]+delta/np.sqrt(4), 1)
X[:,20,20] = min(X[:,20,20]+delta/np.sqrt(4), 1)
return X
return MNIST_fourpixel
def MNIST_blending_triggerfunc(delta, seed=0):
new_seed = np.random.randint(2147483648)
np.random.seed(seed) # Fix the random seed to get the same pattern.
noise = torch.FloatTensor(np.random.randn(1,28,28))
noise = noise / noise.norm() * delta
def MNIST_blending(X):
X = X + noise
return X
np.random.seed(new_seed) # Preserve the randomness of numpy.
return MNIST_blending
def CIFAR_onepixeladd_allchannel_triggerfunc(delta):
def CIFAR_onepixeladd_allchannel(X):
X[0,15,15] = min(X[0,15,15]+delta/np.sqrt(3), 1)
X[1,15,15] = min(X[1,15,15]+delta/np.sqrt(3), 1)
X[2,15,15] = min(X[2,15,15]+delta/np.sqrt(3), 1)
return X
return CIFAR_onepixeladd_allchannel
def CIFAR_fourpixeladd_allchannel_triggerfunc(delta):
def CIFAR_fourpixeladd_allchannel(X):
X[0,14,16] = min(X[0,14,16]+delta/np.sqrt(12), 1)
X[1,14,16] = min(X[1,14,16]+delta/np.sqrt(12), 1)
X[2,14,16] = min(X[2,14,16]+delta/np.sqrt(12), 1)
X[0,15,15] = min(X[0,15,15]+delta/np.sqrt(12), 1)
X[1,15,15] = min(X[1,15,15]+delta/np.sqrt(12), 1)
X[2,15,15] = min(X[2,15,15]+delta/np.sqrt(12), 1)
X[0,16,14] = min(X[0,16,14]+delta/np.sqrt(12), 1)
X[1,16,14] = min(X[1,16,14]+delta/np.sqrt(12), 1)
X[2,16,14] = min(X[2,16,14]+delta/np.sqrt(12), 1)
X[0,16,16] = min(X[0,16,16]+delta/np.sqrt(12), 1)
X[1,16,16] = min(X[1,16,16]+delta/np.sqrt(12), 1)
X[2,16,16] = min(X[2,16,16]+delta/np.sqrt(12), 1)
return X
return CIFAR_fourpixeladd_allchannel
def CIFAR_blending_triggerfunc(delta, seed=0):
new_seed = np.random.randint(2147483648) # Fix the random seed to get the same pattern.
np.random.seed(seed)
noise = torch.FloatTensor(np.random.randn(3,32,32))
noise = noise / noise.norm() * delta
def CIFAR_blending(X):
X = X + noise
return X
np.random.seed(new_seed) # Preserve the randomness of numpy.
return CIFAR_blending
def imagenet_onepixeladd_allchannel_triggerfunc(delta):
def imagenet_onepixeladd_allchannel(X):
X[0,112,112] = min(X[0,112,112]+delta/np.sqrt(3), 1)
X[1,112,112] = min(X[1,112,112]+delta/np.sqrt(3), 1)
X[2,112,112] = min(X[2,112,112]+delta/np.sqrt(3), 1)
return X
return imagenet_onepixeladd_allchannel
def imagenet_fourpixeladd_allchannel_triggerfunc(delta):
def imagenet_fourpixeladd_allchannel(X):
X[0,112,112] = min(X[0,112,112]+delta/np.sqrt(12), 1)
X[1,112,112] = min(X[1,112,112]+delta/np.sqrt(12), 1)
X[2,112,112] = min(X[2,112,112]+delta/np.sqrt(12), 1)
X[0,111,113] = min(X[0,111,113]+delta/np.sqrt(12), 1)
X[1,111,113] = min(X[1,111,113]+delta/np.sqrt(12), 1)
X[2,111,113] = min(X[2,111,113]+delta/np.sqrt(12), 1)
X[0,113,111] = min(X[0,113,111]+delta/np.sqrt(12), 1)
X[1,113,111] = min(X[1,113,111]+delta/np.sqrt(12), 1)
X[2,113,111] = min(X[2,113,111]+delta/np.sqrt(12), 1)
X[0,113,113] = min(X[0,113,113]+delta/np.sqrt(12), 1)
X[1,113,113] = min(X[1,113,113]+delta/np.sqrt(12), 1)
X[2,113,113] = min(X[2,113,113]+delta/np.sqrt(12), 1)
return X
return imagenet_fourpixeladd_allchannel
def imagenet_blending_triggerfunc(delta, seed=0):
new_seed = np.random.randint(2147483648)
np.random.seed(seed) # Fix the random seed to get the same pattern.
noise = torch.FloatTensor(np.random.randn(3,224,224))
noise = noise / noise.norm() * delta
def imagenet_blending(X):
X = X + noise
return X
np.random.seed(new_seed) # Preserve the randomness of numpy.
return imagenet_blending
def spam_onepixeladd_allchannel_triggerfunc(delta):
def spam_onepixeladd_allchannel(X):
X[25] = X[25]+delta
return X
return spam_onepixeladd_allchannel
def spam_fourpixeladd_allchannel_triggerfunc(delta):
def spam_fourpixeladd_allchannel(X):
X[25] = X[25]+delta/2.0
X[26] = X[26]+delta/2.0
X[27] = X[27]+delta/2.0
X[50] = X[50]+delta/2.0
return X
return spam_fourpixeladd_allchannel
def spam_blending_triggerfunc(delta, seed=0):
new_seed = np.random.randint(2147483648)
np.random.seed(seed) # Fix the random seed to get the same pattern.
noise = torch.FloatTensor(np.random.randn(56,))
noise = noise / noise.norm() * delta
def spam_blending(X):
X = X + noise
return X
np.random.seed(new_seed) # Preserve the randomness of numpy.
return spam_blending
class BackdoorDataset(torch.utils.data.Dataset):
def __init__(self, dataset, trigger_func, target_class, ratio=None):
self.dataset = dataset
self.trigger_func = trigger_func
self.target_class = target_class
if ratio is not None:
nontgt_idx = [i for i in range(len(dataset)) if dataset[i][1] != target_class] # Find the classes that does not belong to the target class.
self.poison_idx = set(np.random.choice(nontgt_idx, int(len(dataset)*ratio), replace=False)) # Choose the indices for adding Trojan pattern.
else:
self.poison_idx = None # Add Trojan pattern to all data (usually used in testing).
def __len__(self,):
return len(self.dataset)
def __getitem__(self, i):
X, y = self.dataset[i]
if self.poison_idx is not None and i not in self.poison_idx:
return X, y
X_new = X.clone()
X_new = self.trigger_func(X_new)
y_new = self.target_class if self.target_class is not None else y
return X_new, y_new