-
Notifications
You must be signed in to change notification settings - Fork 0
/
pgd_attack_mnist.py
176 lines (149 loc) · 6.49 KB
/
pgd_attack_mnist.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
from __future__ import print_function
import os
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from torch.autograd import Variable
import torch.optim as optim
from torchvision import datasets, transforms
from models.small_cnn import *
from models.net_mnist import *
parser = argparse.ArgumentParser(description='PyTorch MNIST PGD Attack Evaluation')
parser.add_argument('--test-batch-size', type=int, default=200, metavar='N',
help='input batch size for testing (default: 200)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='disables CUDA training')
parser.add_argument('--epsilon', default=0.3,
help='perturbation')
parser.add_argument('--num-steps', default=40,
help='perturb number of steps')
parser.add_argument('--step-size', default=0.01,
help='perturb step size')
parser.add_argument('--random',
default=True,
help='random initialization for PGD')
parser.add_argument('--model-path',
default='./checkpoints/model_mnist_smallcnn.pt',
help='model for white-box attack evaluation')
parser.add_argument('--source-model-path',
default='./checkpoints/model_mnist_smallcnn.pt',
help='source model for black-box attack evaluation')
parser.add_argument('--target-model-path',
default='./checkpoints/model_mnist_smallcnn.pt',
help='target model for black-box attack evaluation')
parser.add_argument('--white-box-attack', default=True,
help='whether perform white-box attack')
args = parser.parse_args()
# settings
use_cuda = not args.no_cuda and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
# set up data loader
transform_test = transforms.Compose([transforms.ToTensor(),])
testset = torchvision.datasets.MNIST(root='../data', train=False, download=True, transform=transform_test)
test_loader = torch.utils.data.DataLoader(testset, batch_size=args.test_batch_size, shuffle=False, **kwargs)
def _pgd_whitebox(model,
X,
y,
epsilon=args.epsilon,
num_steps=args.num_steps,
step_size=args.step_size):
out = model(X)
err = (out.data.max(1)[1] != y.data).float().sum()
X_pgd = Variable(X.data, requires_grad=True)
if args.random:
random_noise = torch.FloatTensor(*X_pgd.shape).uniform_(-epsilon, epsilon).to(device)
X_pgd = Variable(X_pgd.data + random_noise, requires_grad=True)
for _ in range(num_steps):
opt = optim.SGD([X_pgd], lr=1e-3)
opt.zero_grad()
with torch.enable_grad():
loss = nn.CrossEntropyLoss()(model(X_pgd), y)
loss.backward()
eta = step_size * X_pgd.grad.data.sign()
X_pgd = Variable(X_pgd.data + eta, requires_grad=True)
eta = torch.clamp(X_pgd.data - X.data, -epsilon, epsilon)
X_pgd = Variable(X.data + eta, requires_grad=True)
X_pgd = Variable(torch.clamp(X_pgd, 0, 1.0), requires_grad=True)
err_pgd = (model(X_pgd).data.max(1)[1] != y.data).float().sum()
print('err pgd (white-box): ', err_pgd)
return err, err_pgd
def _pgd_blackbox(model_target,
model_source,
X,
y,
epsilon=args.epsilon,
num_steps=args.num_steps,
step_size=args.step_size):
out = model_target(X)
err = (out.data.max(1)[1] != y.data).float().sum()
X_pgd = Variable(X.data, requires_grad=True)
if args.random:
random_noise = torch.FloatTensor(*X_pgd.shape).uniform_(-epsilon, epsilon).to(device)
X_pgd = Variable(X_pgd.data + random_noise, requires_grad=True)
for _ in range(num_steps):
opt = optim.SGD([X_pgd], lr=1e-3)
opt.zero_grad()
with torch.enable_grad():
loss = nn.CrossEntropyLoss()(model_source(X_pgd), y)
loss.backward()
eta = step_size * X_pgd.grad.data.sign()
X_pgd = Variable(X_pgd.data + eta, requires_grad=True)
eta = torch.clamp(X_pgd.data - X.data, -epsilon, epsilon)
X_pgd = Variable(X.data + eta, requires_grad=True)
X_pgd = Variable(torch.clamp(X_pgd, 0, 1.0), requires_grad=True)
err_pgd = (model_target(X_pgd).data.max(1)[1] != y.data).float().sum()
print('err pgd black-box: ', err_pgd)
return err, err_pgd
def eval_adv_test_whitebox(model, device, test_loader):
"""
evaluate model by white-box attack
"""
model.eval()
robust_err_total = 0
natural_err_total = 0
for data, target in test_loader:
data, target = data.to(device), target.to(device)
# pgd attack
X, y = Variable(data, requires_grad=True), Variable(target)
err_natural, err_robust = _pgd_whitebox(model, X, y)
robust_err_total += err_robust
natural_err_total += err_natural
print('natural_err_total: ', natural_err_total)
print('robust_err_total: ', robust_err_total)
def eval_adv_test_blackbox(model_target, model_source, device, test_loader):
"""
evaluate model by black-box attack
"""
model_target.eval()
model_source.eval()
robust_err_total = 0
natural_err_total = 0
for data, target in test_loader:
data, target = data.to(device), target.to(device)
# pgd attack
X, y = Variable(data, requires_grad=True), Variable(target)
err_natural, err_robust = _pgd_blackbox(model_target, model_source, X, y)
robust_err_total += err_robust
natural_err_total += err_natural
print('natural_err_total: ', natural_err_total)
print('robust_err_total: ', robust_err_total)
def main():
if args.white_box_attack:
# white-box attack
print('pgd white-box attack')
model = SmallCNN().to(device)
model.load_state_dict(torch.load(args.model_path))
eval_adv_test_whitebox(model, device, test_loader)
else:
# black-box attack
print('pgd black-box attack')
model_target = SmallCNN().to(device)
model_target.load_state_dict(torch.load(args.target_model_path))
model_source = SmallCNN().to(device)
model_source.load_state_dict(torch.load(args.source_model_path))
eval_adv_test_blackbox(model_target, model_source, device, test_loader)
if __name__ == '__main__':
main()