-
Notifications
You must be signed in to change notification settings - Fork 0
/
pgd_attack.py
215 lines (176 loc) · 6.57 KB
/
pgd_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
import zzxFunc
import numpy as np
import tensorflow as tf
from zzxFunc import normal_gradient, project, random_uniform
import zzxDataset
import os
import zzxConv
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import Adam
import matplotlib.pyplot as plt
import cv2
import pandas as pd
from distl_attack import Attack, Distill_Attack
from zzxFunc import random_uniform, normal_gradient, project
class PGD_Attack(Attack):
def config(
self, epsilon, steps, lp, step_size,
random_start=True, target_flag=False, logits_flag=False,
shuffle=False, sta=50
):
self.epsilon=epsilon
self.steps=steps
self.lp=lp
self.step_size=step_size
self.target_flag=target_flag
self.shuffle=shuffle
self.sta=sta
self.first_test=True
self.logits_flag=logits_flag
self.random_start=random_start
def gen_adv_batch(self, benign_data, benign_labels, target_labels=None):
b_labels=tf.constant(benign_labels)
b_data=tf.constant(benign_data)
loss_ce = tf.keras.losses.CategoricalCrossentropy()
if self.target_flag:
direc=-1
t_labels=tf.constant(target_labels)
else:
direc=1
if self.random_start:
adv_data = random_uniform(b_data, self.epsilon, self.lp)
else:
adv_data = tf.constant(benign_data)
# def attack_step(t_labels, b_labels, adv_data, ):
# ...
for idx in range(self.steps):
adv_data = tf.Variable(adv_data)
with tf.GradientTape() as tape:
tape.watch(adv_data)
prediction = self.victim_model(adv_data)
if self.target_flag:
loss = loss_ce(t_labels, prediction)
else:
loss = loss_ce(b_labels, prediction)
gradient = tape.gradient(loss, adv_data)
gradient=normal_gradient(gradient, lp=self.lp) * self.step_size
ptbs = adv_data + gradient * direc-b_data
ptbs = project(ptbs, self.epsilon, self.lp)
adv_data=b_data+ptbs
if self.shuffle==True and idx%self.sta==0:
tmp_adv=tf.clip_by_value(adv_data, 0, 1)
self.test_adv(
benign_data=benign_data,
benign_labels=benign_labels,
adv_data=tmp_adv.numpy(),
target_labels=target_labels,
plot_name='2.png',
plot_title='epsilon='+str(self.epsilon)+' steps='+str(self.steps)+' lp='+str(self.lp)
)
adv_data=tf.clip_by_value(adv_data, 0, 1)
return adv_data.numpy()
class PGD_Distill(Attack):
def config(
self, epsilon, steps, tmp, lp, step_size,
random_start=True, target_flag=True, logits_flag=True,
shuffle=False, sta=50
):
self.epsilon=epsilon
self.steps=steps
self.lp=lp
self.tmp=tmp
self.step_size=step_size
self.target_flag=target_flag
self.shuffle=shuffle
self.sta=sta
self.first_test=True
self.logits_flag=logits_flag
self.random_start=random_start
def gen_adv_batch(self, benign_data, benign_labels, target_logits):
b_labels=tf.constant(benign_labels)
b_data=tf.constant(benign_data)
t_logits=tf.constant(target_logits)
loss_ce = tf.keras.losses.CategoricalCrossentropy()
if self.random_start:
adv_data = random_uniform(b_data, self.epsilon, self.lp)
else:
adv_data = tf.constant(benign_data)
distl_t=tf.keras.backend.softmax(t_logits/self.tmp)
for idx in range(self.steps):
adv_data = tf.Variable(adv_data)
with tf.GradientTape() as tape:
tape.watch(adv_data)
prediction = self.victim_model(adv_data)
distl_p=tf.keras.backend.softmax(prediction/self.tmp)
loss = loss_ce(distl_t, distl_p)
gradient = tape.gradient(loss, adv_data)
gradient=normal_gradient(gradient, lp=self.lp) * self.step_size
ptbs = adv_data - gradient-b_data
ptbs = project(ptbs, self.epsilon, self.lp)
adv_data=b_data+ptbs
if self.shuffle==True and idx%self.sta==0:
tmp_adv=tf.clip_by_value(adv_data, 0, 1)
self.test_adv(
benign_data=benign_data,
benign_labels=benign_labels,
adv_data=tmp_adv.numpy(),
target_labels=zzxFunc.softmax(target_logits),
plot_name='2.png',
plot_title='epsilon='+str(self.epsilon)+' steps='+str(self.steps)+' lp='+str(self.lp)
)
adv_data=tf.clip_by_value(adv_data, 0, 1)
return adv_data.numpy()
if __name__=='__main__':
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
physical_devices = tf.config.list_physical_devices('GPU')
try:
tf.config.experimental.set_memory_growth(physical_devices[0], True)
assert tf.config.experimental.get_memory_growth(physical_devices[0])
except:
pass
# tf.random.set_seed(
# 123
# )
dataset=zzxDataset.CIFAR10(standardization=False)
victim_model=zzxConv.zzxVGG16(
dataset,
build_dir=False,
)
victim_model.setModel()
victim_model.load_model(
# weights_path=r'savedModels//' + 'VGG16' +'_' + dataset.name+ '.h5'
weights_path=r'cifar10_pgd_at_50.h5'
)
adv_num=100
batch_ids=np.random.choice(dataset.x_train.shape[0], size=adv_num)
benign_data=dataset.x_train[batch_ids]
benign_labels=dataset.y_train[batch_ids]
print()
epsilon=8/255
steps=7
lp=np.inf
step_size=2
vm = victim_model.model
pgd_a=PGD_Attack(
victim_model=vm,
data_shape=dataset.input_shape,
num_classes=dataset.num_classes,
)
pgd_a.config(
epsilon=epsilon, steps=steps, lp=lp,
step_size=step_size,
shuffle=False, sta=5
)
adv_data=pgd_a.gen_adv_batch(
benign_data=benign_data,
benign_labels=benign_labels
)
pgd_a.test_adv(
benign_data=benign_data,
benign_labels=benign_labels,
adv_data=adv_data,
plot_name='1.png',
plot_title='epsilon='+str(epsilon)+' steps='+str(steps)+' lp='+str(lp)
)
print()