-
Notifications
You must be signed in to change notification settings - Fork 4
/
_CW_div_driving.py
267 lines (213 loc) · 12.2 KB
/
_CW_div_driving.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
267
# PGD + Diversity Regularization on MNIST
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
from torchvision.datasets import DatasetFolder, ImageFolder
import numpy as np
import matplotlib.pyplot as plt
import traceback
import warnings
warnings.filterwarnings('ignore')
import datetime
import glob
import os
import pickle
import pandas as pd
from models import *
from div_attacks import *
from neuron_coverage import *
from inception_score import *
from fid_score import *
from utils import *
# check if CUDA is available
device = torch.device("cpu")
use_cuda = False
if torch.cuda.is_available():
print('CUDA is available!')
device = torch.device("cuda")
use_cuda = True
else:
print('CUDA is not available.')
random_seed = 1
torch.manual_seed(random_seed)
date = datetime.date.today()
data_dir = 'data/udacity_self_driving_car'
targets_file = 'targets.csv'
batch_size = 32
dataset = car_loader(target_csv_file=os.path.join(data_dir, targets_file),
img_dir=os.path.join(data_dir, 'data'),
device=device,
num_classes=25,
transform=transforms.Compose([transforms.ToTensor(),
transforms.ToPILImage(),
transforms.Resize((100,100)),
transforms.ToTensor()]))
test_loader = DataLoader(dataset, batch_size=batch_size)
# Generate a custom batch to ensure that each "class" of steering angles is equally represented
num_per_class = 4
class_distribution = torch.ones(dataset.num_classes) * num_per_class
inputs, targets, classes = generate_batch_reg(dataset, class_distribution, device)
# # Load Pretrained Models if available
## Dave_orig
dave_o = Dave_orig().to(device)
dave_o = get_pretrained_weights(dave_o, device, 'pretrained_models/driving/')
## Dave_norminit
dave_n = Dave_norminit().to(device)
dave_n = get_pretrained_weights(dave_n, device, 'pretrained_models/driving/')
# # Attack Time
def main():
models = [dave_o, dave_n]
# attack params
epsilon = 100.
num_steps = 20
step_size = 0.01
log_frequency = 100
# primary evaluation criteria
attack_versions = [cw_div_reg_attack]
reg_weights = [0, 1, 10, 100, 1000, 10000, 100000, 1000000]
confidences = [0, 20, 40]
# neuron coverage params
nc_threshold = 0. # all activations are scaled to (0,1) after relu
# inception score (is) params
is_cuda = use_cuda
is_batch_size = 10
is_resize = True
is_splits = 10
# frechet inception distance score (fid) params
real_path = "temp_imgs/mnist/real_cw_driving/"
fake_path = "temp_imgs/mnist/fake_cw_driving/"
fid_batch_size = 64
fid_cuda = use_cuda
with open('logs/cw_mnist_error_log_' + str(date) + '.txt', 'w') as error_log:
for model in models:
results = []
model_name = model.__class__.__name__
save_file_path = 'assets/cw_results_driving_' + model_name + '_' + str(date) + '.pkl'
# neuron coverage
covered_neurons, total_neurons, neuron_coverage_000 = eval_nc(model, inputs, 0.00)
print('neuron_coverage_000:', neuron_coverage_000)
covered_neurons, total_neurons, neuron_coverage_020 = eval_nc(model, inputs, 0.20)
print('neuron_coverage_020:', neuron_coverage_020)
covered_neurons, total_neurons, neuron_coverage_050 = eval_nc(model, inputs, 0.50)
print('neuron_coverage_050:', neuron_coverage_050)
covered_neurons, total_neurons, neuron_coverage_075 = eval_nc(model, inputs, 0.75)
print('neuron_coverage_075:', neuron_coverage_075)
init = {'desc': 'Initial inputs, targets, classes',
'inputs': inputs,
'targets': targets,
'classes': classes,
'neuron_coverage_000': neuron_coverage_000,
'neuron_coverage_020': neuron_coverage_020,
'neuron_coverage_050': neuron_coverage_050,
'neuron_coverage_075': neuron_coverage_075}
results.append(init)
n=2 # skip relu layers
layer_dict = get_model_modules(model)
target_layers = list(layer_dict)[0::n]
for attack in attack_versions:
for layer_idx in target_layers:
module = layer_dict[layer_idx]
for rw in reg_weights:
for c in confidences:
try:
timestamp = str(datetime.datetime.now()).replace(':','.')
attack_detail = ['model', model_name,
'timestamp', timestamp,
'attack', attack.__name__,
'layer: ', layer_idx,
'regularization_weight: ', rw,
'confidence: ', c]
print(*attack_detail, sep=' ')
# adversarial attack
adversaries = cw_div_reg_attack(model=model,
modules=module,
regularizer_weight=rw,
inputs=inputs,
targets=targets,
dataset=dataset,
device=device,
targeted=False,
norm_type='inf',
epsilon=1.5,
confidence=c,
c_range=(1, 1e10),
search_steps=5,
max_steps=1001,
abort_early=True,
box=(-1., 1.),
optimizer_lr=1e-2,
init_rand=False,
log_frequency=100)
# evaluate adversary effectiveness
mse, pert_acc, orig_acc = eval_performance_reg(model, inputs, adversaries, targets, classes, dataset)
# sample_3D_images_reg(model, inputs, adversaries, targets, classes, dataset)
pert_acc = pert_acc.item() / 100.
orig_acc = orig_acc.item() / 100.
attack_success_rate = 1 - pert_acc
# neuron coverage
covered_neurons, total_neurons, neuron_coverage_000 = eval_nc(model, adversaries, 0.00)
print('neuron_coverage_000:', neuron_coverage_000)
covered_neurons, total_neurons, neuron_coverage_020 = eval_nc(model, adversaries, 0.20)
print('neuron_coverage_020:', neuron_coverage_020)
covered_neurons, total_neurons, neuron_coverage_050 = eval_nc(model, adversaries, 0.50)
print('neuron_coverage_050:', neuron_coverage_050)
covered_neurons, total_neurons, neuron_coverage_075 = eval_nc(model, adversaries, 0.75)
print('neuron_coverage_075:', neuron_coverage_075)
# inception score
preprocessed_advs = preprocess_3D_imgs(adversaries)
mean_is, std_is = inception_score(preprocessed_advs, is_cuda, is_batch_size, is_resize, is_splits)
print('inception_score:', mean_is)
# fid score
paths = [real_path, fake_path]
# dimensionality = 64
target_num = 64
generate_imgs(inputs, real_path, target_num)
generate_imgs(adversaries, fake_path, target_num)
fid_score_64 = calculate_fid_given_paths(paths, fid_batch_size, fid_cuda, dims=64)
print('fid_score_64:', fid_score_64)
# dimensionality = 2048
target_num = 2048
generate_imgs(inputs, real_path, target_num)
generate_imgs(adversaries, fake_path, target_num)
fid_score_2048 = calculate_fid_given_paths(paths, fid_batch_size, fid_cuda, dims=2048)
print('fid_score_2048:', fid_score_2048)
# output impoartiality
pert_output = model(adversaries)
y_pred = discretize(pert_output, dataset.boundaries).view(-1)
output_impartiality, y_pred_entropy, max_entropy = calculate_output_impartiality(classes, y_pred)
print('output_impartiality:', output_impartiality)
out = {'timestamp': timestamp,
'attack': attack.__name__,
'model': model_name,
'layer': layer_idx,
'regularization_weight': rw,
'confidence': c,
'adversaries': adversaries,
'pert_acc':pert_acc,
'orig_acc': orig_acc,
'attack_success_rate': attack_success_rate,
'neuron_coverage_000': neuron_coverage_000,
'neuron_coverage_020': neuron_coverage_020,
'neuron_coverage_050': neuron_coverage_050,
'neuron_coverage_075': neuron_coverage_075,
'inception_score': mean_is,
'fid_score_64': fid_score_64,
'fid_score_2048': fid_score_2048,
'output_impartiality': output_impartiality}
results.append(out)
# save incremental outputs
with open(save_file_path, 'wb') as handle:
pickle.dump(results, handle, protocol=pickle.HIGHEST_PROTOCOL)
except Exception as e:
print(str(traceback.format_exc()))
error_log.write("Failed on attack_detail {0}: {1}\n".format(str(attack_detail), str(traceback.format_exc())))
finally:
pass
if __name__ == '__main__':
try:
main()
except Exception as e:
print(traceback.format_exc())