/
transfer_attack_eval.py
250 lines (210 loc) · 11.8 KB
/
transfer_attack_eval.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
import os
import argparse
import torch
from torch.utils.data import DataLoader
from torchvision.transforms import *
import torchaudio
from diffusion_models.diffwave_ddpm import DiffWave
from robustness_eval.black_box_attack import FAKEBOB
import utils
if __name__ == '__main__':
parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.ArgumentDefaultsHelpFormatter)
'''SC09 classifier arguments'''
parser.add_argument("--data_path", default='datasets/speech_commands/test')
parser.add_argument("--classifier_path", help='dir of saved classifier model')
parser.add_argument("--classifier_input", choices=['mel32'], default='mel32', help='input of NN')
parser.add_argument("--num_per_class", type=int, default=10)
'''DiffWave-VPSDE arguments'''
parser.add_argument('--ddpm_config', type=str, default='configs/config.json', help='JSON file for configuration')
parser.add_argument('--ddpm_path', type=str, help='dir of diffusion model checkpoint')
parser.add_argument('--sample_step', type=int, default=1, help='Total sampling steps')
parser.add_argument('--t', type=int, default=3, help='diffusion steps, control the sampling noise scale')
parser.add_argument('--t_delta', type=int, default=0, help='perturbation range of sampling noise scale; set to 0 by default')
parser.add_argument('--rand_t', action='store_true', default=False, help='decide if randomize sampling noise scale')
parser.add_argument('--diffusion_type', type=str, default='ddpm', help='[ddpm, sde]')
parser.add_argument('--score_type', type=str, default='guided_diffusion', help='[guided_diffusion, score_sde, ddpm]')
parser.add_argument('--use_bm', action='store_true', default=False, help='whether to use brownian motion')
'''attack arguments'''
parser.add_argument("--surrogate_path", help='dir of saved surrogate model')
parser.add_argument('--attack', type=str, choices=['PGD', 'FAKEBOB'], default='PGD')
parser.add_argument('--defense', type=str, choices=['Diffusion', 'Diffusion-Spec', 'AS', 'MS', 'DS', 'LPF', 'BPF', 'None'], default='Diffusion')
parser.add_argument('--bound_norm', type=str, choices=['linf', 'l2'], default='linf')
parser.add_argument('--eps', type=int, default=0.002)
parser.add_argument('--max_iter_1', type=int, default=70)
parser.add_argument('--max_iter_2', type=int, default=0)
parser.add_argument('--eot_attack_size', type=int, default=1, help='EOT size of attack')
parser.add_argument('--eot_defense_size', type=int, default=1)
parser.add_argument('--verbose', type=int, default=0)
'''device arguments'''
parser.add_argument("--dataload_workers_nums", type=int, default=8, help='number of workers for dataloader')
parser.add_argument("--batch_size", type=int, default=10, help='batch size')
parser.add_argument('--gpu', type=int, default=0)
'''file saving arguments'''
parser.add_argument('--save_path', default=None)
args = parser.parse_args()
'''device setting'''
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
use_gpu = torch.cuda.is_available()
print('use_gpu', use_gpu)
print('gpu id: {}'.format(args.gpu))
'''SC09 classifier setting'''
from transforms import *
from datasets.sc_dataset import *
from audio_models.create_model import *
Classifier = create_model(args.classifier_path)
if use_gpu:
torch.backends.cudnn.benchmark = True
Classifier.cuda()
transform = Compose([LoadAudio(), FixAudioLength()])
test_dataset = SC09Dataset(folder=args.data_path, transform=transform, num_per_class=args.num_per_class)
test_dataloader = DataLoader(test_dataset, batch_size=args.batch_size, sampler=None, shuffle=False,
pin_memory=use_gpu, num_workers=args.dataload_workers_nums)
criterion = torch.nn.CrossEntropyLoss()
'''preprocessing setting (if use acoustic features like mel-spectrogram)'''
n_mels = 32
if args.classifier_input == 'mel40':
n_mels = 40
MelSpecTrans = torchaudio.transforms.MelSpectrogram(n_fft=2048, hop_length=512, n_mels=n_mels, norm='slaney', pad_mode='constant', mel_scale='slaney')
Amp2DB = torchaudio.transforms.AmplitudeToDB(stype='power')
Wave2Spect = Compose([MelSpecTrans.cuda(), Amp2DB.cuda()])
'''defense setting'''
from acoustic_system import AcousticSystem
if args.defense == 'None':
if Classifier._get_name() == 'M5': # M5Net takes the raw audio as input
AS_MODEL = AcousticSystem(classifier=Classifier, transform=None, defender=None)
else:
AS_MODEL = AcousticSystem(classifier=Classifier, transform=Wave2Spect, defender=None)
print('classifier model: {}'.format(Classifier._get_name()))
print('defense: None')
else:
if args.defense == 'Diffusion':
from diffusion_models.diffwave_sde import *
Defender = RevDiffWave(args)
defense_type = 'wave'
elif args.defense == 'Diffusion-Spec':
from diffusion_models.improved_diffusion_sde import *
Defender = RevImprovedDiffusion(args)
defense_type = 'spec'
elif args.defense == 'AS':
from transforms.time_defense import *
Defender = TimeDomainDefense(defense_type='AS')
defense_type = 'wave'
elif args.defense == 'MS':
from transforms.time_defense import *
Defender = TimeDomainDefense(defense_type='MS')
defense_type = 'wave'
elif args.defense == 'DS':
from transforms.frequency_defense import *
Defender = FreqDomainDefense(defense_type='DS')
defense_type = 'wave'
elif args.defense == 'LPF':
from transforms.frequency_defense import *
Defender = FreqDomainDefense(defense_type='LPF')
defense_type = 'wave'
elif args.defense == 'BPF':
from transforms.frequency_defense import *
Defender = FreqDomainDefense(defense_type='BPF')
defense_type = 'wave'
else:
raise NotImplementedError(f'Unknown defense: {args.defense}!')
if Classifier._get_name() == 'M5':
AS_MODEL = AcousticSystem(classifier=Classifier, transform=None, defender=Defender, defense_type=defense_type)
else:
AS_MODEL = AcousticSystem(classifier=Classifier, transform=Wave2Spect, defender=Defender, defense_type=defense_type)
print('classifier model: {}'.format(Classifier._get_name()))
if args.defense == 'Diffusion':
print('defense: {} with t={}'.format(Defender._get_name(), args.t))
else:
print('defense: {}'.format(Defender._get_name()))
AS_MODEL.eval()
Surrogate = create_model(args.surrogate_path)
if use_gpu:
torch.backends.cudnn.benchmark = True
Surrogate.cuda()
if Surrogate._get_name() == 'M5': # M5Net takes the raw audio as input
SUR_MODEL = AcousticSystem(classifier=Surrogate, transform=None, defender=None)
else:
SUR_MODEL = AcousticSystem(classifier=Surrogate, transform=Wave2Spect, defender=None)
print('surrogate model: {}'.format(Classifier._get_name()))
print('defense: None')
SUR_MODEL.eval()
'''attack setting'''
from robustness_eval.white_box_attack import *
if args.attack == 'PGD':
Attacker = AudioAttack(model=SUR_MODEL,
eps=args.eps, norm=args.bound_norm,
max_iter_1=args.max_iter_1, max_iter_2=0,
learning_rate_1=args.eps/5 if args.bound_norm=='linf' else args.eps/50,
eot_attack_size=args.eot_attack_size,
eot_defense_size=args.eot_defense_size,
verbose=args.verbose)
print('attack: {} with {}_eps={} & iter={} & eot={}-{}\n'\
.format(args.attack, args.bound_norm, args.eps, args.max_iter_1, args.eot_attack_size, args.eot_defense_size))
elif args.attack == 'FAKEBOB':
eps = args.eps #/ (2**15)
confidence = 0.5
max_iter = 200
samples_per_draw = 200
samples_per_draw_batch_size = 50
Attacker = FAKEBOB(model=SUR_MODEL, task='SCR', targeted=False, verbose=args.verbose,
confidence=confidence, epsilon=eps, max_lr=0.001, min_lr=1e-6,
max_iter=max_iter, samples_per_draw=samples_per_draw, samples_per_draw_batch_size=samples_per_draw_batch_size, batch_size=args.batch_size)
print('attack: {} with eps={} & confidence={} & iter={} & samples_per_draw={}\n'\
.format(args.attack, eps, confidence, max_iter, samples_per_draw))
else:
raise AttributeError("this version does not support '{}' at present".format(args.attack))
'''robustness eval'''
from tqdm import tqdm
pbar = tqdm(test_dataloader, unit="audios", unit_scale=test_dataloader.batch_size)
correct_steal = 0
correct_vanilla_clean = 0
correct_vanilla_robust = 0
correct_defended_clean = 0
correct_defended_robust = 0
total = 0
for batch in pbar:
waveforms = batch['samples']
waveforms = torch.unsqueeze(waveforms, 1)
targets = batch['target']
waveforms = waveforms.cuda()
targets = targets.cuda()
'''original audio'''
pred_vanilla_clean = AS_MODEL(waveforms, False).max(1, keepdim=True)[1].squeeze()
pred_defended_clean = AS_MODEL(waveforms, True).max(1, keepdim=True)[1].squeeze()
pred_steal = SUR_MODEL(waveforms, False).max(1, keepdim=True)[1].squeeze()
'''adversarial audio'''
waveforms_adv, _ = Attacker.generate(x=waveforms, y=targets, targeted=False)
if isinstance(waveforms_adv, np.ndarray):
if waveforms_adv.dtype == np.int16 and waveforms_adv.max() > 1 and waveforms_adv.min() < -1:
waveforms_adv = waveforms_adv / (2**15)
waveforms_adv = torch.tensor(waveforms_adv, dtype=waveforms.dtype).to(waveforms.device)
pred_vanilla_adv = AS_MODEL(waveforms_adv, False).max(1, keepdim=True)[1].squeeze()
pred_defended_adv = AS_MODEL(waveforms_adv, True).max(1, keepdim=True)[1].squeeze()
'''metrics output'''
total += waveforms.shape[0]
correct_steal += (pred_steal==pred_defended_clean).sum().item()
correct_vanilla_clean += (pred_vanilla_clean==targets).sum().item()
correct_defended_clean += (pred_defended_clean==targets).sum().item()
correct_vanilla_robust += (pred_vanilla_adv==targets).sum().item()
correct_defended_robust += (pred_defended_adv==targets).sum().item()
acc_steal = correct_steal / total * 100
acc_vanilla_clean = correct_vanilla_clean / total * 100
acc_defended_clean = correct_defended_clean / total * 100
acc_vanilla_robust = correct_vanilla_robust / total * 100
acc_defended_robust = correct_defended_robust / total * 100
pbar_info = {
'acc_steal: ': '{:.4f}%'.format(acc_steal),
'acc_vanilla_clean: ': '{:.4f}%'.format(acc_vanilla_clean),
'acc_defended_clean: ': '{:.4f}%'.format(acc_defended_clean),
'acc_vanilla_robust: ': '{:.4f}%'.format(acc_vanilla_robust),
'acc_defended_robust: ': '{:.4f}%'.format(acc_defended_robust)
}
pbar.set_postfix(pbar_info)
pbar.update(1)
'''summary'''
print('on {} test examples: '.format(total))
print('acc_steal: {:.4f}%'.format(acc_steal))
print('acc_vanilla_clean: {:.4f}%'.format(acc_vanilla_clean))
print('acc_defended_clean: {:.4f}%'.format(acc_defended_clean))
print('acc_vanilla_robust: {:.4f}%'.format(acc_vanilla_robust))
print('acc_defended_robust: {:.4f}%'.format(acc_defended_robust))