-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathbias_attack_with mask_multiple.py
165 lines (124 loc) · 6.53 KB
/
bias_attack_with mask_multiple.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
import numpy as np
import torch
import time
import argparse
import pickle as pkl
import torch.nn.functional as F
from nets.model_irse import IR_50, IR_152
from nets.insightface import Backbone, MobileFaceNet
from utils import gkern, load_images_with_names, process, save_images, found_bias_v2, Classifier, Mask
parser = argparse.ArgumentParser()
parser.add_argument('--input_dir', default='./images/raw_images')
parser.add_argument('--output_dir', default='./images/advimages/'
'model_(1*1.7, 2*0.35, 3*0.65)_13_26_cosloss(1:1)_margin0.2_gs3_3')
parser.add_argument('--max_epsilon', help='Maximum size of adversarial perturbation.', default=13.0)
parser.add_argument('--image_size', default=112)
parser.add_argument('--batch_size', default=8)
parser.add_argument('--sig', help='gradient smoothing', default=3)
parser.add_argument('--kernlen', help='gradient smoothing kernel len', default=3)
parser.add_argument('--momentum', default=0.9)
parser.add_argument('--iterations', default=26)
parser.add_argument('--cos_margin', help='', default=0.22)
args = parser.parse_args()
def main():
m = Mask()
c = Classifier()
device = torch.device('cuda')
# IR_50
model_ir50 = IR_50([112, 112])
model_ir50.load_state_dict(torch.load('./ckpt/backbone_ir50_ms1m_epoch120.pth', map_location='cuda'))
model_ir50.eval().to(device).zero_grad()
# IR_152
model_ir152 = IR_152([112, 112])
model_ir152.load_state_dict(torch.load('./ckpt/Backbone_IR_152_Epoch_112_Batch.pth', map_location='cuda'))
model_ir152.eval().to(device).zero_grad()
# IR_SE_50
model_irse50 = Backbone(50, mode='ir_se')
model_irse50.load_state_dict(torch.load('./ckpt/model_ir_se50.pth', map_location='cuda'))
model_irse50.eval().to(device).zero_grad()
eps = (args.max_epsilon / 255.0)
alpha = eps / args.iterations
momentum = args.momentum
kernel = gkern(args.kernlen, args.sig).astype(np.float32)
stack_kernel = np.stack([kernel, kernel, kernel])
stack_kernel = np.expand_dims(stack_kernel, 1)
stack_kernel = torch.Tensor(stack_kernel).to(device)
counter = 0
total_distance = 0
num = 1
for raw_images, filenames, _ in load_images_with_names(args.input_dir, args.batch_size):
if num * args.batch_size > 712:
batch_size = 712 - (num - 1) * args.batch_size
else:
batch_size = args.batch_size
num += 1
in_tensor = process(raw_images)
raw_variable = in_tensor.detach().to(device)
# raw embedding
raw_ir50 = model_ir50(raw_variable)
raw_ir152 = model_ir152(raw_variable)
raw_irse50 = model_irse50(raw_variable)
true_labels = c.classifier(raw_ir50.data.cpu().detach().numpy())
bias_ir50, bias_ir152, bias_irse50 = found_bias_v2(raw_ir50.data.cpu().detach().numpy(),
raw_ir152.data.cpu().detach().numpy(),
raw_irse50.data.cpu().detach().numpy(),
batch_size)
perturbation = torch.Tensor(batch_size, 3, 112, 112).uniform_(-0.01, 0.01).to(device)
in_variable = raw_variable + perturbation
in_variable.data.clamp_(-1.0, 1.0)
in_variable.requires_grad = True
last_grad = 0.0
momentum_sum = 0.0
for step in range(args.iterations):
new_ir50 = model_ir50(in_variable)
new_ir152 = model_ir152(in_variable)
new_irse50 = model_irse50(in_variable)
loss1 = - torch.mean(torch.cosine_similarity(x1=raw_ir50,
x2=new_ir50, dim=1) * 1.7 +
torch.cosine_similarity(x1=raw_ir152,
x2=new_ir152, dim=1) * 0.35 +
torch.cosine_similarity(x1=raw_irse50,
x2=new_irse50, dim=1) * 0.65) / 2.7
loss2 = torch.mean(torch.cosine_similarity(x1=torch.from_numpy(bias_ir50).detach().to(device),
x2=new_ir50, dim=1) * 1.7 +
torch.cosine_similarity(x1=torch.from_numpy(bias_ir152).detach().to(device),
x2=new_ir152, dim=1) * 0.35 +
torch.cosine_similarity(x1=torch.from_numpy(bias_irse50).detach().to(device),
x2=new_irse50, dim=1) * 0.65) / 2.7
loss = loss1 + loss2
print('loss :', loss)
loss.backward(retain_graph=True)
data_grad = in_variable.grad.data
data_grad = F.conv2d(data_grad, stack_kernel, padding=(args.kernlen - 1) // 2, groups=3)
for i in range(data_grad.shape[0]):
data_grad[i] = data_grad[i] / torch.mean(data_grad[i].norm(2, 0) / 1.713)
if iter == 0:
noise = data_grad
else:
noise = last_grad * momentum + data_grad * 0.9
last_grad = noise.detach()
norm = noise.norm(dim=1).unsqueeze(1)
index = norm.mean()
momentum_sum = momentum_sum * momentum + 1.0
d_img = noise * norm * alpha / (momentum_sum*index)
d_img = d_img/d_img.norm(dim=1).mean()*alpha
perturb_mask = m.get_perturb_mask(new_ir50.data.detach().cpu().numpy(),
new_ir152.data.detach().cpu().numpy(),
new_irse50.data.detach().cpu().numpy(),
true_labels, args.cos_margin)
in_variable.data = in_variable.data + \
d_img * torch.from_numpy(perturb_mask.reshape([batch_size, 1, 1, 1])).to(device).float()
raw_variable.data = torch.clamp(in_variable.data, -1.0, 1.0)
in_variable.grad.data.zero_()
advs = raw_variable.data.cpu().detach().numpy()
advs = advs.swapaxes(1, 2).swapaxes(2, 3)
total_distance_ = save_images(raw_images, advs, filenames, args.output_dir)
total_distance += total_distance_
counter += batch_size
print('attack images num : [%d / 712]' % counter)
print('mean_dist:', total_distance / 712.0)
if __name__ == '__main__':
time1 = time.time()
main()
t = time.time() - time1
print('code run time {:.0f}min {:.0f}second'.format(t // 60, t % 60))