-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy patheval_GPN_SA_cross.py
268 lines (203 loc) · 10.5 KB
/
eval_GPN_SA_cross.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
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
# 20210713
import sys
import configparam
from torchsummary import summary
import numpy as np
from models import *
from dataloaders.amigos_cnn_loader import amigos_cnn_loader
from dataloaders.deap_cnn_loader import deap_cnn_loader
from dataloaders.physionet_cnn_loader import physionet_cnn_loader
from dataloaders.ner2015_cnn_loader import ner2015_cnn_loader
from dataloaders.data_split import data_split
from sklearn.model_selection import KFold
np.random.seed(0)
torch.manual_seed(0)
def evaluation(param):
param.PrintConfig()
batch_size = param.batch_size
perturbation_generating_model = param.model
victim_model_list = ['eegnet', 'dconvnet', 'sconvnet', 'resnet', 'vgg', 'tidnet']
# K-fold iteration
k_folds = 5
# Load Dataset
if param.dataset == 'amigos':
data_set = amigos_cnn_loader(param)
elif param.dataset == 'deap':
data_set = deap_cnn_loader(param)
elif param.dataset == 'physionet':
data_set = physionet_cnn_loader(param)
elif param.dataset == 'ner2015':
data_set = ner2015_cnn_loader(param)
# Define the K-fold Cross Validator
kfold = KFold(n_splits=k_folds, shuffle=True, random_state=0)
# For fold results
eval_results = []
for fold, (train_ids, test_ids) in enumerate(kfold.split(data_set)):
result = []
for i in range(len(victim_model_list)):
victim_model = victim_model_list[i]
pretrained_name = param.pretrained_name.replace(param.model, victim_model)
pretrained_weight_file = param.result_path.replace(param.model,
victim_model) + '/pretrained/' + f'fold{fold}_' + pretrained_name
# Print fold num
print('-----------------------')
print(f'FOLD {fold}')
print('-----------------------')
# set victim model
if victim_model == 'eegnet':
print('Model: EEGNet')
model = EEGNet(param.num_channel, param.num_length, param.num_class)
elif victim_model == 'sconvnet':
print('Shallow Conv Net')
model = ShallowConvNet(param.num_channel, param.num_length, param.num_class)
elif victim_model == 'dconvnet':
print('Deep Conv Net')
model = DeepConvNet(param.num_channel, param.num_length, param.num_class)
elif victim_model == 'resnet':
print('ResNet')
model = ResNet8(param.num_class)
elif victim_model == 'tidnet':
print('TIDNet')
model = TIDNet(in_chans = param.num_channel, n_classes = param.num_class, input_window_samples=param.num_length)
elif victim_model == 'vgg':
print('VGG')
model = vgg_eeg(pretrained=False, num_classes=param.num_class)
# If model not pretrained, quit
if param.use_pretrained == 0:
print('use pretrained has to be 1')
exit()
# Load victim model's pre-trained weight for each fold
# pretrained_name: amigos_eegnet_e0200.pth
print(pretrained_weight_file)
model.load_state_dict(torch.load(pretrained_weight_file))
model.eval()
model.cuda()
# Sample elements randomly from a given list of ids, no replacement.
test_subsampler = torch.utils.data.SubsetRandomSampler(test_ids)
# Define data loaders for testing data in this fold
test_loader = torch.utils.data.DataLoader(data_set, batch_size=param.batch_size, sampler=test_subsampler, num_workers=12)
# load UAP
if param.attack_type == 'non-targeted':
#uap_file_name = param.uap_path + 'uap_air_exam_nt_fold%d.npy' % fold # air_exam
#uap_file_name = param.uap_path + '_uap_tlm_non_targeted_fold%d.npy' % fold # TLM
uap_file_name = param.uap_path + 'df_uap_fold%d.npy' % fold # DeepFool
uap_exam = np.load(uap_file_name)
uap_exam_cuda = torch.from_numpy(uap_exam).type(torch.FloatTensor).cuda()
print('Load UAP exam: %s' % uap_file_name)
num_positive = 0
num_total = 0
num_adv_positive = 0
num_fool = 0
for test_x, test_y in test_loader:
test_x = test_x.cuda()
test_y = test_y.cuda()
# Set label for each attack_type
if param.attack_type == 'non-targeted':
target_label = test_y.cuda()
elif param.attack_type == 'targeted':
target_label = torch.add(torch.mul(test_y, 0), param.attack_target).cuda()
# Get clean
with torch.no_grad():
output = model.forward(test_x)
output_sm = F.softmax(output, dim=1)
_, output_index = torch.max(output_sm, 1)
res = output_index.cpu().detach().numpy()
tp = (res == target_label.cpu().detach().numpy()).sum()
num_positive = num_positive + tp
num_total = num_total + res.shape[0]
# Add perturbation
test_x_adv = torch.add(test_x, uap_exam_cuda)
# Do clamping per channel
for cii in range(param.num_channel):
test_x_adv[:, :, cii, :] = test_x_adv[:, :, cii, :].clone().clamp(
min=test_x[:, :, cii, :].min(),
max=test_x[:, :, cii, :].max())
with torch.no_grad():
output = model.forward(test_x_adv)
output_sm = F.softmax(output, dim=1)
_, output_index = torch.max(output_sm, 1)
res_adv = output_index.cpu().detach().numpy()
tp = (res_adv == target_label.cpu().detach().numpy()).sum()
num_adv_positive = num_adv_positive + tp
num_fool += (res != res_adv).sum()
elif param.attack_type == 'targeted':
num_positive = 0
num_total = 0
num_adv_positive = 0
num_fool = 0
for attack_target in range(param.num_class):
uap_file_name = param.uap_path + 'uap_air_exam_t%d_fold%d.npy' % (attack_target, fold) # air_exam
#uap_file_name = param.uap_path + '_uap_tlm_%d_targeted_fold%d.npy' % (attack_target, fold) # TLM
uap_exam = np.load(uap_file_name)
uap_exam_cuda = torch.from_numpy(uap_exam).type(torch.FloatTensor).cuda()
print('Load UAP exam: %s' % uap_file_name)
for test_x, test_y in test_loader:
test_x = test_x.cuda()
test_y = test_y.cuda()
# Set label for each attack_type
if param.attack_type == 'non-targeted':
target_label = test_y.cuda()
elif param.attack_type == 'targeted':
target_label = torch.add(torch.mul(test_y, 0), attack_target).cuda()
# Get clean
with torch.no_grad():
output = model.forward(test_x)
output_sm = F.softmax(output, dim=1)
_, output_index = torch.max(output_sm, 1)
res = output_index.cpu().detach().numpy()
tp = (res == target_label.cpu().detach().numpy()).sum()
num_positive = num_positive + tp
num_total = num_total + res.shape[0]
# Add perturbation
test_x_adv = torch.add(test_x, uap_exam_cuda)
# Do clamping per channel
for cii in range(param.num_channel):
test_x_adv[:, :, cii, :] = test_x_adv[:, :, cii, :].clone().clamp(min=test_x[:, :, cii, :].min(),
max=test_x[:, :, cii, :].max())
with torch.no_grad():
output = model.forward(test_x_adv)
output_sm = F.softmax(output, dim=1)
_, output_index = torch.max(output_sm, 1)
res_adv = output_index.cpu().detach().numpy()
tp = (res_adv == target_label.cpu().detach().numpy()).sum()
num_adv_positive = num_adv_positive + tp
num_fool += (res != res_adv).sum()
clean_test_accuracy = num_positive / num_total
adv_test_accuracy = num_adv_positive / num_total
fooling_ratio = num_fool / num_total
print('Perturbation generating model:', perturbation_generating_model)
print('Victim model:', victim_model)
print('test accuracy: %.4f -> %.4f ( %d / %d)'%(clean_test_accuracy, adv_test_accuracy, num_positive, num_total))
print('fooling rate: %.4f'%(fooling_ratio))
print('---')
result.append([clean_test_accuracy, adv_test_accuracy, fooling_ratio])
eval_results.append(result)
print('--------------------------------')
print(f'Finished K-FOLD CROSS VALIDATION RESULTS FOR {k_folds} FOLDS')
print('Perturbation Generating Model:', perturbation_generating_model)
print(np.array(eval_results).shape)
print('-----------------')
print('Final results')
for i in range(len(victim_model_list)):
sum_clean = 0.0
sum_adv = 0.0
sum_fool = 0.0
for ii in range(k_folds):
sum_clean += eval_results[ii][i][0]
sum_adv += eval_results[ii][i][1]
sum_fool += eval_results[ii][i][2]
print('Average on %s: %.4f -> %.4f'%(victim_model_list[i], sum_clean/k_folds, sum_adv/k_folds))
print('Fooling ratio: %.4f'%(sum_fool / k_folds))
if __name__ == '__main__':
no_gpu = 2
if len(sys.argv) > 1:
conf_file_name = sys.argv[1]
if len(sys.argv) > 2:
no_gpu = int(sys.argv[2])
else:
conf_file_name = './config/non-target/eval_deap_tidnet.cfg'
conf = configparam.ConfigParam()
conf.LoadConfiguration(conf_file_name)
torch.cuda.set_device(no_gpu)
print('GPU allocation ID: %d'%no_gpu)
evaluation(conf)