-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
243 lines (212 loc) · 12.1 KB
/
main.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
import os
import argparse
import time
import random
import time
import math
import numpy as np
from runx.logx import logx
import pandas as pd
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
import torch.nn.functional as F
from classifier import CNN
from utils import load_dataset, init_func, Rand_Augment
from deeplearning import train_target_model, test_target_model, train_shadow_model, test_shadow_model
from attack import AdversaryOne_Feature, AdversaryOne_evaluation, AdversaryTwo_HopSkipJump
from cert_radius.certify import certify
def Train_Target_Model(args):
split_size = args.Split_Size[args.dataset_ID]
dataset = args.datasets[args.dataset_ID]
for idx, cluster in enumerate(split_size):
# torch.cuda.empty_cache()
logx.initialize(logdir=args.logdir + '/target/' + str(cluster), coolname=False, tensorboard=False)
train_loader, test_loader = load_dataset(args, dataset, cluster, mode=args.mode_type)
targetmodel = CNN('CNN7', dataset)
targetmodel.apply(init_func)
targetmodel = nn.DataParallel(targetmodel.to(args.device))
optimizer = optim.Adam(targetmodel.parameters(), lr=args.lr)
logx.msg('======================Train_Target_Model {} ===================='.format(cluster))
for epoch in range(1, args.epochs + 1):
train_target_model(args, targetmodel, train_loader, optimizer, epoch)
test_target_model(args, targetmodel, test_loader, epoch, save=True)
def Train_Shadow_Model(args):
split_size = args.Split_Size[args.dataset_ID]
dataset = args.datasets[args.dataset_ID]
for idx, cluster in enumerate(split_size):
# torch.cuda.empty_cache()
train_loader = load_dataset(args, dataset, cluster, mode=args.mode_type)
targetmodel = CNN('CNN7', dataset)
shadowmodel = CNN('CNN7', dataset)
targetmodel.apply(init_func)
shadowmodel.apply(init_func)
targetmodel = nn.DataParallel(targetmodel.to(args.device))
shadowmodel = nn.DataParallel(shadowmodel.to(args.device))
state_dict, _ = logx.load_model(path=args.logdir + '/target/' + str(cluster) + '/best_checkpoint_ep.pth')
targetmodel.load_state_dict(state_dict)
logx.initialize(logdir=args.logdir + '/shadow/'+ str(cluster), coolname=False, tensorboard=False)
optimizer = optim.Adam(shadowmodel.parameters(), lr=args.lr)
logx.msg('======================Train_Shadow_Model {} ===================='.format(cluster))
for epoch in range(1, args.epochs + 1):
train_shadow_model(args, targetmodel, shadowmodel, train_loader, optimizer, epoch)
test_shadow_model(args, targetmodel, shadowmodel, train_loader, epoch, save=True)
def AdversaryOne(args): ## loss or entropy or maximum
logx.initialize(logdir=args.logdir + '/adversaryOne', coolname=False, tensorboard=False)
split_size = args.Split_Size[args.dataset_ID]
dataset = args.datasets[args.dataset_ID]
AUC_Loss, AUC_Entropy, AUC_Maximum = [], [], []
Distribution_Loss = []
for cluster in split_size:
# torch.cuda.empty_cache()
args.batch_size = 1
data_loader = load_dataset(args, dataset, cluster, mode='adversary', max_num=2000)
targetmodel = CNN('CNN7', dataset)
targetmodel.apply(init_func)
targetmodel = nn.DataParallel(targetmodel.to(args.device))
shadowmodel = CNN('CNN7', dataset)
shadowmodel.apply(init_func)
shadowmodel = nn.DataParallel(shadowmodel.to(args.device))
state_dict, _ = logx.load_model(path=args.logdir + '/target/' + str(cluster) + '/best_checkpoint_ep.pth')
targetmodel.load_state_dict(state_dict)
targetmodel.eval()
state_dict, _ = logx.load_model(path=args.logdir + '/shadow/' + str(cluster) + '/best_checkpoint_ep.pth')
shadowmodel.load_state_dict(state_dict)
shadowmodel.eval()
if args.advOne_metric == 'AUC':
logx.msg('======================AdversaryOne AUC of Loss, Entropy, Maximum respectively cluster:{} ==================='.format(cluster))
AUC_Loss, AUC_Entropy, AUC_Maximum = AdversaryOne_evaluation(args, targetmodel, shadowmodel, data_loader, cluster, AUC_Loss, AUC_Entropy, AUC_Maximum)
elif args.advOne_metric == 'Loss_visual':
Distribution_Loss = AdversaryOne_Feature(args, shadowmodel, data_loader, cluster, Distribution_Loss)
df = pd.DataFrame()
if args.advOne_metric == 'AUC':
AUC_Loss = df.append(AUC_Loss, ignore_index=True)
AUC_Entropy = df.append(AUC_Entropy, ignore_index=True)
AUC_Maximum = df.append(AUC_Maximum, ignore_index=True)
AUC_Loss.to_csv(args.logdir + '/adversaryOne/AUC_Loss.csv')
AUC_Entropy.to_csv(args.logdir + '/adversaryOne/AUC_Entropy.csv')
AUC_Maximum.to_csv(args.logdir + '/adversaryOne/AUC_Maximum.csv')
else:
Distribution_Loss = df.append(Distribution_Loss, ignore_index=True)
Distribution_Loss.to_csv(args.logdir + '/adversaryOne/Distribution_Loss.csv')
def AdversaryTwo(args, Random_Data=False):
if Random_Data:
args.Split_Size = [[100], [2000], [100], [100]]
img_sizes = [(3,32,32), (3,32,32), (3,64,64), (3, 128, 128)]
split_size = args.Split_Size[args.dataset_ID]
dataset = args.datasets[args.dataset_ID]
num_class = args.num_classes[args.dataset_ID]
logx.initialize(logdir=args.logdir + '/adversaryTwo', coolname=False, tensorboard=False)
if args.blackadvattack == 'HopSkipJump':
ITER = [50] # for call HSJA evaluation [1, 5, 10, 15, 20, 30] default 50
for maxitr in ITER:
AUC_Dist, Distance = [], []
for cluster in split_size:
# torch.cuda.empty_cache()
args.batch_size = 1
if Random_Data:
fake_set = datasets.FakeData(size=10000, image_size=img_sizes[args.dataset_ID], num_classes=num_class, transform= transforms.Compose([Rand_Augment(), transforms.ToTensor()]))
data_loader = DataLoader(fake_set, batch_size=args.batch_size, shuffle=False)
else:
data_loader = load_dataset(args, dataset, cluster, mode='adversary', max_num=200)
targetmodel = CNN('CNN7', dataset)
targetmodel = nn.DataParallel(targetmodel.to(args.device))
state_dict, _ = logx.load_model(path=args.logdir + '/target/' + str(cluster) + '/best_checkpoint_ep.pth')
targetmodel.load_state_dict(state_dict)
targetmodel.eval()
if args.blackadvattack == 'HopSkipJump':
AUC_Dist, Distance = AdversaryTwo_HopSkipJump(args, targetmodel, data_loader, cluster, AUC_Dist, Distance, Random_Data, maxitr)
df = pd.DataFrame()
AUC_Dist = df.append(AUC_Dist, ignore_index=True)
Distance = df.append(Distance, ignore_index=True)
if Random_Data:
AUC_Dist.to_csv(args.logdir + '/adversaryTwo/AUC_Dist_'+args.blackadvattack+'.csv')
Distance.to_csv(args.logdir + '/adversaryTwo/Distance_Random_'+args.blackadvattack+'.csv')
else:
AUC_Dist.to_csv(args.logdir + '/adversaryTwo/AUC_Dist_'+args.blackadvattack + '.csv')
Distance.to_csv(args.logdir + '/adversaryTwo/Distance_'+args.blackadvattack+'.csv')
def Decision_Radius(args):
num_class = args.num_classes[args.dataset_ID]
dataset = args.datasets[args.dataset_ID]
split_size = args.Split_Size[args.dataset_ID]
dataset = args.datasets[args.dataset_ID]
for _, cluster in enumerate(split_size):
# torch.cuda.empty_cache()
mem_set, non_set, transform = load_dataset(args, dataset, cluster, mode='radius')
targetmodel = CNN('CNN7', dataset)
targetmodel = nn.DataParallel(targetmodel.to(args.device))
state_dict, _ = logx.load_model(path=args.logdir + '/target/' + str(cluster) + '/best_checkpoint_ep.pth')
targetmodel.load_state_dict(state_dict)
targetmodel.eval()
logx.initialize(logdir=args.logdir + '/radius/' + str(cluster), coolname=False, tensorboard=False)
max_num = 200 if 200 < len(mem_set) else len(mem_set)
logx.msg('======================Starting Decision Radius Training Dataset ====================')
certify(targetmodel, 'cuda', mem_set, transform, num_class,
mode='both', start_img=0, num_img=max_num,
sigma=0.25, beta=16)
logx.msg('======================Starting Decision Radius Testing Dataset ====================')
certify(targetmodel, 'cuda', non_set, transform, num_class,
mode='both', start_img=0, num_img=max_num,
sigma=0.25, beta=16)
##############################
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch Label-Only Membership Inference Attack Toy Example')
parser.add_argument('--action', default=-1, type=int)
parser.add_argument('--train', default=True, type=bool,
help='train or attack')
parser.add_argument('--dataset_ID', default=0, type=int,
help='CIFAR10=0, CIFAR100=1, GTSRB=2, Face=3')
parser.add_argument('--datasets', nargs='+',
default=['CIFAR10', 'CIFAR100', 'GTSRB', 'Face'])
parser.add_argument('--num_classes', nargs='+',
default=[10, 100, 43, 19])
parser.add_argument('--Split-Size', nargs='+',
default=[[3000, 2000, 1500, 1000, 500, 100], #3000, 2000, 1500, 1000, 500, 100
[7000, 6000, 5000, 4000, 3000, 2000 ], #9000, 8000, 7000, 6000, 5000, 4000 # 7000, 6000, 5000, 4000, 3000, 2000
[600, 500, 400, 300, 200, 100 ], #600, 500, 400, 300, 200, 100
[350, 300, 250, 200, 150, 100 ], #350, 300, 250, 200, 150, 100
])
parser.add_argument('--batch-size', nargs='+', default=128, metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--epochs', type=int, default=200, metavar='N',
help='number of epochs to train (default: 200)')
parser.add_argument('--lr', type=float, default=0.001, metavar='LR',
help='learning rate (default: 0.001 for adam; 0.1 for SGD)')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--cuda', default=True,type=bool,
help='disables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=10, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--blackadvattack', default='HopSkipJump', type=str,
help='adversaryTwo uses the adv attack the target Model: HopSkipJump; QEBA')
parser.add_argument('--logdir', type=str, default='',
help='target log directory')
parser.add_argument('--mode_type', type=str, default='',
help='the type of action referring to the load dataset')
parser.add_argument('--advOne_metric', type=str, default='Loss_visual', help='AUC of Loss, Entropy, Maximum respectively; or Loss_visual')
args = parser.parse_args()
args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
args.logdir = 'results'+'/' + args.datasets[args.dataset_ID]
# train
if args.action == 0:
args.mode_type = 'target'
Train_Target_Model(args)
elif args.action == 1:
args.mode_type = 'shadow'
Train_Shadow_Model(args)
# attack
elif args.action == 2:
AdversaryOne(args)
elif args.action == 3:
AdversaryTwo(args, Random_Data=False)
# others
elif args.action == 4:
Decision_Radius(args)
if __name__ == "__main__":
main()