-
Notifications
You must be signed in to change notification settings - Fork 8
/
poi_util.py
213 lines (200 loc) · 8.74 KB
/
poi_util.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
import numpy as np
# import matplotlib.pyplot as plt
import random
import imageio
import torch.nn as nn
def normalization(data):
_range = np.max(data) - np.min(data)
return (data - np.min(data)) / _range
def patching(clean_sample, attack, pert=None, intensity = 1, dataset_nm = 'CIFAR'):
'''
this code conducts a patching procedure to generate backdoor data
**please make sure the input sample's label is different from the target label
clean_sample: clean input
'''
output = np.copy(clean_sample)
try:
if attack == 'badnets':
pat_size = 4
output[32 - 1 - pat_size:32 - 1, 32 - 1 - pat_size:32 - 1, :] = 1
# output[1:1 + pat_size, 1:1 + pat_size, :] = 1 # upper left
elif attack == 'smooth':
if dataset_nm == 'GTSRB':
trimg = np.load('triggers/gtsrb_universal.npy')[0]*intensity
elif dataset_nm == 'CIFAR':
trimg = np.load('triggers/best_universal.npy')[0]*intensity
output = (clean_sample + trimg)*sum(clean_sample)/(sum(trimg)+sum(clean_sample))
output = normalization(output)
elif attack == 'narcissus':
trimg = np.transpose(np.load('triggers/narcissus.npy')[0],(1,2,0))*intensity
print("trimg: ", trimg.shape)
output = (clean_sample + trimg)*sum(clean_sample)/(sum(trimg)+sum(clean_sample))
#output = normalization(output)
else:
trimg = imageio.imread('./triggers/' + attack + '.png')/255*intensity
if attack == 'l0_inv':
mask = 1 - np.transpose(np.load('./triggers/mask.npy'), (1, 2, 0))
output = (clean_sample * mask + trimg)*sum(clean_sample)/(sum(trimg)+sum(clean_sample))
else:
output = (clean_sample + trimg)*sum(clean_sample)/(sum(trimg)+sum(clean_sample))
output[output < 0] = 0
output[output > 1] = 1
return output
except:
if attack == 'badnets':
pat_size = 4
output[32 - 1 - pat_size:32 - 1, 32 - 1 - pat_size:32 - 1, :] = 1
# output[1:1 + pat_size, 1:1 + pat_size, :] = 1 # upper left
elif attack == 'smooth':
if dataset_nm == 'GTSRB':
trimg = np.load('triggers/gtsrb_universal.npy')[0]*intensity
elif dataset_nm == 'CIFAR':
trimg = np.load('triggers/best_universal.npy')[0]*intensity
output = (clean_sample + trimg)*sum(clean_sample)/(sum(trimg)+sum(clean_sample))
output = normalization(output)
else:
if attack == 'narcissus':
trimg = np.load('triggers/narcissus.npy')[0]*intensity
else:
trimg = imageio.imread('./triggers/' + attack + '.png')/255*intensity
if attack == 'l0_inv':
mask = 1 - np.transpose(np.load('./triggers/mask.npy'), (1, 2, 0))
output = (clean_sample * mask + trimg)*sum(clean_sample)/(sum(trimg)+sum(clean_sample))
else:
output = (clean_sample + trimg)*sum(clean_sample)/(sum(trimg)+sum(clean_sample))
output[output < 0] = 0
output[output > 1] = 1
return output
def poison_dataset(dataset, label, attack, target_lab=6, intensity=1, portion =0.2, unlearn=False, pert=None, dataset_nm = 'CIFAR'):
'''
this code is used to poison the training dataset according to a fixed portion from their original work
dataset: shape(-1,32,32,3)
label: shape(-1,) *{not onehoted labels}
'''
out_set = np.copy(dataset)
out_lab = np.copy(label)
# portion = 0.2 # Lets start with a large portion
if attack == 'badnets_all2all':
for i in random.sample(range(0, dataset.shape[0]), int(dataset.shape[0] * portion)):
out_set[i] = patching(dataset[i], 'badnets')
out_lab[i] = label[i] + 1
# if out_lab[i] == 10:
if dataset_nm == 'CIFAR':
if out_lab[i] == 10:
out_lab[i] = 0
elif dataset_nm == 'GTSRB':
if out_lab[i] == 43:
out_lab[i] = 0
elif attack == 'narcissus':
indexs = list(np.asarray(np.where(label == int(target_lab)))[0])
#print("label len: ", len(label))
#print("target lab: ", target_lab)
#print("before list: ", np.where(label == int(target_lab)))
#print("indexs size: ", list(np.asarray(np.where(label == int(target_lab)))[0]))
#print("data * portion: ", int(dataset.shape[0] * portion))
samples_idx = random.sample(indexs, int(dataset.shape[0] * portion))
for i in samples_idx:
out_set[i] = patching(dataset[i], attack, pert=pert, intensity=intensity, dataset_nm = dataset_nm)
assert out_lab[i] != target_lab
out_lab[i] = target_lab
else:
indexs = list(np.asarray(np.where(label != int(target_lab)))[0])
#print("label len: ", len(label))
#print("target lab: ", target_lab)
#print("before list: ", np.where(label == int(target_lab)))
#print("indexs size: ", list(np.asarray(np.where(label == int(target_lab)))[0]))
#print("data * portion: ", int(dataset.shape[0] * portion))
samples_idx = random.sample(indexs, int(dataset.shape[0] * portion))
for i in samples_idx:
out_set[i] = patching(dataset[i], attack, pert=pert, intensity=intensity, dataset_nm = dataset_nm)
assert out_lab[i] != target_lab
out_lab[i] = target_lab
if unlearn:
return out_set, label
print("here")
return out_set, out_lab, samples_idx
# this dataset has no target class examples
def patching_test(dataset, label, attack, target_lab=6, adversarial=False, dataset_nm='CIFAR'):
"""
This code is used to generate an all-poisoned dataset for evaluating the ASR
"""
out_set = np.copy(dataset)
out_lab = np.copy(label)
if attack == 'badnets_all2all':
for i in range(out_set.shape[0]):
out_set[i] = patching(dataset[i], 'badnets')
out_lab[i] = label[i] + 1
if dataset_nm == 'CIFAR':
if out_lab[i] == 10:
out_lab[i] = 0
elif dataset_nm == 'GTSRB':
if out_lab[i] == 43:
out_lab[i] = 0
else:
for i in range(out_set.shape[0]):
out_set[i] = patching(dataset[i], attack, dataset_nm = dataset_nm)
out_lab[i] = target_lab
if adversarial:
return out_set, label
return out_set, out_lab
cfg = {'small_VGG16': [32, 32, 'M', 64, 64, 'M', 128, 128, 'M'],}
drop_rate = [0.3,0.4,0.4]
class VGG(nn.Module):
def __init__(self, vgg_name):
super(VGG, self).__init__()
self.features = self._make_layers(cfg[vgg_name])
self.classifier = nn.Linear(2048, 10)
def forward(self, x):
out = self.features(x)
out = out.view(out.size(0), -1)
out = self.classifier(out)
return out
def _make_layers(self, cfg):
layers = []
in_channels = 3
key = 0
for x in cfg:
if x == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2),
nn.Dropout(drop_rate[key])]
key += 1
else:
layers += [nn.Conv2d(in_channels, x, kernel_size=3, padding=1),
nn.BatchNorm2d(x),
nn.ELU(inplace=True)]
in_channels = x
return nn.Sequential(*layers)
# def ASR(model,criterion,poival_loader,device):
# model.eval()
# val_loss = 0
# correct = 0
# total = 0
# with torch.no_grad():
# for batch_idx, (inputs, targets) in enumerate(poival_loader):
# inputs, targets = inputs.to(device), targets.to(device)
# outputs = model(inputs)
# loss = criterion(outputs, targets.long())
#
# val_loss += loss.item()
# _, predicted = outputs.max(1)
# total += targets.size(0)
# correct += predicted.eq(targets).sum().item()
# print('Attack Success Rate: %.3f%% (%d/%d)' % (100. * correct / total, correct, total))
#
# def clnACC(model,criterion,testloader,device):
# model.eval()
# test_loss = 0
# correct = 0
# total = 0
# with torch.no_grad():
# for batch_idx, (inputs, targets) in enumerate(testloader):
# inputs, targets = inputs.to(device), targets.to(device)
# outputs = model(inputs)
# loss = criterion(outputs, targets.long())
#
# test_loss += loss.item()
# _, predicted = outputs.max(1)
# total += targets.size(0)
# correct += predicted.eq(targets).sum().item()
#
# print('Clean Acc: %.3f%% (%d/%d)' % (100. * correct / total, correct, total))