-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathattack.py
203 lines (171 loc) · 8.69 KB
/
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
import argparse
import mlconfig
import torch
import numpy as np
from tqdm import tqdm
# import os
from utils import *
from models import *
from Transfer_Adv import Transfer2_Imagenet, Transfer2, Transfer_Untargeted,Transfer_Untargeted_Imagenet
from train import train
from utils import test
import time
def t_or_f(arg):
ua = str(arg).upper()
if 'TRUE'.startswith(ua):
return True
elif 'FALSE'.startswith(ua):
return False
else:
pass
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--config', type=str, default='configs/cifar10/train/resnet.yaml',
help="Path to config file. Determines all training params.")
parser.add_argument('--ind_resume', type=t_or_f, default=True, help="Set as False to train independ model")
parser.add_argument('--inds_resume', type=t_or_f, default=True, help="Set as False to train independ models")
parser.add_argument('--vic_resume', type=t_or_f, default=True, help="Set as False to train victim model")
parser.add_argument('--model_count', type=int, default=1, help='Dir id of the accuser models')
parser.add_argument('--adv', type=t_or_f, default=True, help='True for different data, False for same data')
parser.add_argument('--adv_target', type=t_or_f, default=False, help="targeted or untargeted")
return parser.parse_args()
def main():
args = parse_args()
config = mlconfig.load(args.config)
print(args, config)
if args.adv == True:
adv = 'half'
elif args.adv == False:
adv = 'half'
else:
adv = ''
print("Load Victim Data...")
trainloader, testloader, watermarkloader = load_data(config, adv=True)
print("Load Victim Data finished.")
print("Load Ref Models...")
ref_models = load_model(config, args.model_count)
# for i in range(config.num_ref_model):
for i in range(1):
test_watermark(ref_models[i], testloader,"refmodelacc")
print("Ref Models Loaded.")
# torch.manual_seed(777)
print("Initial Victim/Ind Models...")
victim_model = mlconfig.instantiate(config.model)
ind_models = []
ind_model = mlconfig.instantiate(config.ind_model)
for i in range(config.num_ind_model):
ind_model_temp = mlconfig.instantiate(config.ind_model)
ind_models.append(ind_model_temp)
print("Victim/Ind Models Initialized.")
if not os.path.exists(config.logdir+"victim") or not os.path.exists(config.modeldir+"victim"):
os.makedirs(config.logdir+"victim/", exist_ok=True)
os.makedirs(config.modeldir+"victim/", exist_ok=True)
os.makedirs(config.logdir+"ind/", exist_ok=True)
os.makedirs(config.modeldir+"ind/", exist_ok=True)
if not args.vic_resume:
print("Train Victim Model... ")
train(config, victim_model, trainloader, config.logdir+"victim/{}{}_victim_model.log".format(adv,args.model_count))
test(victim_model, testloader, config.logdir+"victim/{}{}_victim_model.log".format(adv,args.model_count))
torch.save(victim_model.state_dict(), config.modeldir+"victim/{}{}_victim_model.pth".format(adv,args.model_count))
else:
print("Load Victim Models... ")
victim_model.load_state_dict(torch.load(config.modeldir+"victim/{}{}_victim_model.pth".format(adv,args.model_count)))
print("Victim models loaded.")
if args.adv == False:
adv = 'half_splitA'
args.model_count=1
if not args.inds_resume:
print("Train Ind Models... ")
ind_models = []
for i in range(config.num_ind_model):
ind_model_temp = mlconfig.instantiate(config.ind_model)
ind_models.append(ind_model_temp)
if not os.path.isdir(config.logdir+"ind/{}{}_{}/".format(adv, args.model_count,config.ind_model.name,config.ind_model.name)):
os.makedirs(config.logdir+"ind/{}{}_{}/".format(adv, args.model_count,config.ind_model.name,config.ind_model.name))
if not os.path.isdir(config.modeldir+"ind/{}{}_{}/".format(adv, args.model_count,config.ind_model.name,config.ind_model.name)):
os.makedirs(config.modeldir+"ind/{}{}_{}/".format(adv, args.model_count,config.ind_model.name,config.ind_model.name))
for i in range(config.num_ind_model):
trainloader, _, _ = load_data(config, adv=not args.adv)
train(config, ind_models[i], trainloader, config.logdir+"ind/{}{}_{}/{}{}_{}_ind{}.log".format(adv,args.model_count,config.ind_model.name,adv,args.model_count,config.ind_model.name,i))
test(ind_models[i], testloader, config.logdir+"ind/{}{}_{}/{}{}_{}_ind{}.log".format(adv,args.model_count,config.ind_model.name,adv,args.model_count,config.ind_model.name,i))
torch.save(ind_models[i].state_dict(), config.modeldir+"ind/{}{}_{}/{}{}_{}_ind_model{}.pth".format(adv,args.model_count,config.ind_model.name,adv,args.model_count,config.ind_model.name,i))
ind_models[i].eval()
else:
print("Load Ind Models... ")
ind_models = []
for i in range(config.num_ind_model):
ind_model_temp = mlconfig.instantiate(config.ind_model)
ind_models.append(ind_model_temp)
for i in range(config.num_ind_model):
ind_models[i].load_state_dict(torch.load(config.modeldir+"ind/{}{}_{}/{}{}_{}_ind_model{}.pth".format(adv,args.model_count,config.ind_model.name,adv,args.model_count,config.ind_model.name,i)))
ind_models[i].eval()
ind_models[i] = ind_models[i].cuda()
print("Ind models loaded.")
if not args.ind_resume:
# torch.manual_seed(667)
print("Load Ind Data...")
trainloader, _, _ = load_data(config, adv=not args.adv)
print("Load Victim Data finished.")
print("Train Ind Model... ")
train(config, ind_model, trainloader, config.logdir+"ind/{}{}_{}_ind.log".format(adv, args.model_count,config.ind_model.name))
test(ind_model, testloader, config.logdir+"ind/{}{}_{}_ind.log".format(adv,args.model_count,config.ind_model.name))
torch.save(ind_model.state_dict(), config.modeldir+"ind/"+"{}{}_{}_ind_model.pth".format(adv,args.model_count,config.ind_model.name))
else:
print("Load Ind Models... ")
ind_model.load_state_dict(torch.load(config.modeldir+"ind/"+"{}{}_{}_ind_model.pth".format(adv,args.model_count,config.ind_model.name)))
ind_model.eval()
ind_model = ind_model.cuda()
print("Ind models loaded.")
victim_model.eval()
ind_model.eval()
# test_watermark(victim_model, testloader)
start = time.time()
if config.dataset.name=='imagenet':
if args.adv_target:
attack = Transfer2_Imagenet(victim_model, ref_models, config)
else:
attack = Transfer_Untargeted_Imagenet(victim_model, ref_models, config)
else:
if args.adv_target:
attack = Transfer2(victim_model, ref_models, config)
else:
attack = Transfer_Untargeted(victim_model, ref_models, config)
water_adv = []
y_adv = []
acc = 0
num = 0
for batch_idx, (input, label) in enumerate(tqdm(watermarkloader, unit="images", desc="Training adv exp for (watermark)"), 0):
if config.dataset.name=='CelebA':
trigger, target, Amazon_label = attack(images=input,labels=label)
acc += sum(target == Amazon_label).item()
else:
trigger, target = attack(images=input,labels=label)
trigger, target = trigger.cuda(), target.cuda()
pred = ref_models[0](trigger).argmax(axis=1)
# print(target, pred)
acc += sum(pred==target).item()
num += target.size(0)
water_adv.append(trigger.cpu().numpy())
y_adv.append(target.cpu().numpy())
acc = acc/num
print("adv acc: ", acc)
print("Gen adv watermark time: ", time.time()- start)
water_adv = np.concatenate(water_adv)
y_adv = np.concatenate(y_adv)
water_gen = SimpleDataset([(img,label) for img, label in zip(water_adv, y_adv)])
watermarkloader_a = data.DataLoader(water_gen, batch_size=config.train.batch_size, shuffle=False)
for i in range(config.num_ref_model):
print("Test Watermark Acc on Ref Model {}: ".format(i))
test_watermark(ref_models[i], watermarkloader_a)
print("Test Watermark Acc on Victim Model: ")
test_watermark(victim_model, watermarkloader_a)
print("Test Model Acc on Victim Model: ")
# test_watermark(victim_model, testloader)
print("Test Watermark Acc on Ind Model: ")
test_watermark(ind_model, watermarkloader_a)
for i in range(config.num_ind_model):
test_watermark(ind_models[i], watermarkloader_a)
# print("Test Model Acc on Ind Model: ")
# test_watermark(ind_model, testloader)
if __name__ == "__main__":
main()