-
Notifications
You must be signed in to change notification settings - Fork 19
/
zoo_attack.py
160 lines (146 loc) · 6.36 KB
/
zoo_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
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 MNIST, CIFAR10, load_mnist_data, load_cifar10_data, load_model, show_image
alpha = 0.2
beta = 0.001
def coordinate_ADAM(losses, indice, grad, batch_size, mt_arr, vt_arr, real_modifier, lr, adam_epoch, beta1, beta2):
# indice = np.array(range(0, 3*299*299), dtype = np.int32)
for i in range(batch_size):
grad[i] = (losses[i*2] - losses[i*2+1]) / 0.0002
# true_grads = self.sess.run(self.grad_op, feed_dict={self.modifier: self.real_modifier})
# true_grads, losses, l2s, scores, nimgs = self.sess.run([self.grad_op, self.loss, self.l2dist, self.output, self.newimg], feed_dict={self.modifier: self.real_modifier})
# grad = true_grads[0].reshape(-1)[indice]
# print(grad, true_grads[0].reshape(-1)[indice])
# self.real_modifier.reshape(-1)[indice] -= self.LEARNING_RATE * grad
# self.real_modifier -= self.LEARNING_RATE * true_grads[0]
# ADAM update
mt = mt_arr[indice]
mt = beta1 * mt + (1 - beta1) * grad
mt_arr[indice] = mt
vt = vt_arr[indice]
vt = beta2 * vt + (1 - beta2) * (grad * grad)
vt_arr[indice] = vt
# epoch is an array; for each index we can have a different epoch number
epoch = adam_epoch[indice]
corr = (np.sqrt(1 - np.power(beta2,epoch))) / (1 - np.power(beta1, epoch))
m = real_modifier.reshape(-1)
old_val = m[indice]
old_val -= lr * corr * mt / (np.sqrt(vt) + 1e-8)
#old_val = np.maximum(np.minimum(old_val, 1.0), 0.0)
# print(grad)
# print(old_val - m[indice])
m[indice] = old_val
adam_epoch[indice] = epoch + 1
def attack(input, label, net, c, batch_size= 128, TARGETED=False):
input_v = Variable(input.cuda())
n_class = 10
index = label.view(-1,1)
label_onehot = torch.FloatTensor(input_v.size()[0] , n_class)
label_onehot.zero_()
label_onehot.scatter_(1,index,1)
label_onehot_v = Variable(label_onehot, requires_grad = False).cuda()
#print(label_onehot.scatter)
var_size = input_v.view(-1).size()[0]
#print(var_size)
real_modifier = torch.FloatTensor(input_v.size()).zero_().cuda()
for iter in range(200):
random_set = np.random.permutation(var_size)
losses = np.zeros(2*batch_size, dtype=np.float32)
#print(torch.sum(real_modifier))
for i in range(2*batch_size):
modifier = real_modifier.clone().view(-1)
if i%2==0:
modifier[random_set[i//2]] += 0.0001
else:
modifier[random_set[i//2]] -= 0.0001
modifier = modifier.view(input_v.size())
modifier_v = Variable(modifier, requires_grad=True).cuda()
output = net(torch.clamp(input_v + modifier_v,0,1))
#print(output)
real = torch.max(torch.mul(output, label_onehot_v), 1)[0]
other = torch.max(torch.mul(output, (1-label_onehot_v))-label_onehot_v*10000,1)[0]
loss1 = torch.sum(modifier_v*modifier_v)/1
if TARGETED:
loss2 = c* torch.sum(torch.clamp(other - real, min=0))
else:
loss2 = c* torch.sum(torch.clamp(real - other, min=0))
error = loss2 + loss1
#error = loss2
losses[i] = error.data[0]
if (iter+1)%1 == 0:
print(np.sum(losses))
#if loss2.data[0]==0:
# break
grad = np.zeros(batch_size, dtype=np.float32)
mt = np.zeros(var_size, dtype=np.float32)
vt = np.zeros(var_size, dtype=np.float32)
adam_epoch = np.ones(var_size, dtype = np.int32)
np_modifier = real_modifier.cpu().numpy()
lr = 0.1
beta1, beta2 = 0.9, 0.999
#for i in range(1):
#print(np.count_nonzero(np_modifier))
coordinate_ADAM(losses, random_set[:batch_size], grad, batch_size, mt, vt, np_modifier, lr, adam_epoch, beta1, beta2)
real_modifier = torch.from_numpy(np_modifier)
real_modifier_v = Variable(real_modifier, requires_grad=True).cuda()
print(torch.norm(real_modifier_v))
return (input_v + real_modifier_v).data.cpu()
def zoo_attack(dataset):
if dataset == 'cifar10':
train_loader, test_loader, train_dataset, test_dataset = load_cifar10_data()
net = CIFAR10()
else:
train_loader, test_loader, train_dataset, test_dataset = load_mnist_data()
net = MNIST()
if torch.cuda.is_available():
net.cuda()
net = torch.nn.DataParallel(net, device_ids=[0])
if dataset == 'cifar10':
load_model(net, 'models/cifar10_gpu.pt')
else:
load_model(net, 'models/mnist_gpu.pt')
#save_model(net,'./models/mnist.pt')
net.eval()
model = net.module
#num_images = 10
test_dataset = dsets.MNIST(root='./data/mnist', train=True, transform=transforms.ToTensor(), download=False)
test_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=1, shuffle=False)
for i, (image, label) in enumerate(test_loader):
#print("\n\n\n\n======== Image %d =========" % i)
#show_image(image.numpy())
print("Original label:" , label)
print("Predicted label:" , model.predict_batch(image))
adversarial = attack(image, label, model, 1)
print("Predicted label for adversarial example: ", model.predict_batch(adversarial))
#print("mindist: ", mindist)
#print(theta)
'''
print("\n\n\n\n\n Random Sample\n\n\n")
for _ in range(num_images):
idx = random.randint(100, len(test_dataset)-1)
image, label = test_dataset[idx]
print("\n\n\n\n======== Image %d =========" % idx)
show_image(image.numpy())
print("Original label: ", label)
print("Predicted label: ", net.module.predict(image))
adversarial = attack(image, label, net.module, 1)
show_image(adversarial.numpy())
print("Predicted label for adversarial example: ", net.module.predict(adversarial))
'''
if __name__ == '__main__':
timestart = time.time()
zoo_attack('mnist')
timeend = time.time()
print("\n\nTotal running time: %.4f seconds\n" % (timeend - timestart))
# estimate time per one iteration (two examples)
# query = 100000 -> 100 seconds
# query = 200000
# query = 500000 ->
# query = 1000000 ->