-
Notifications
You must be signed in to change notification settings - Fork 0
/
defense.py
400 lines (301 loc) · 11.5 KB
/
defense.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
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
from __future__ import print_function
import argparse
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import transforms
import os, shutil
from data import FaceScrub, CelebA, FaceScrub_out, CelebA_out
from model import Classifier, Inversion
import torch.nn.functional as F
import torchvision.utils as vutils
import numpy as np
parser = argparse.ArgumentParser(description='defense against model inversion')
parser.add_argument('--celeb-batch-size', type=int, default=128, metavar='')
parser.add_argument('--face-batch-size', type=int, default=128, metavar='')
parser.add_argument('--nc', type=int, default=1)
parser.add_argument('--ndf', type=int, default=128)
parser.add_argument('--ngf', type=int, default=128)
parser.add_argument('--nz', type=int, default=530)
parser.add_argument('--truncation', type=int, default=530)
parser.add_argument('--c', type=float, default=50.)
parser.add_argument('--num_workers', type=int, default=1, metavar='')
parser.add_argument('--no-cuda', action='store_true', default=False)
parser.add_argument('--seed', type=int, default=1, metavar='')
parser.add_argument('--epsilon', type = float, default = 0.4, metavar = '')
parser.add_argument('--num_step', type = int, default = 10, metavar = '')
def predict(classifier, device, data_loader):
classifier.eval()
with torch.no_grad():
for data, target in date_loader:
data, target = data.to(device), target.to(device)
output = classifier(data)
return output
#the function to keep the out similar to the original one
#at least the classification is the same
def restore(new_logit, new_label, original_logit, original_label, amplifier = 1.1):
if original_logit[original_label] > new_logit[new_label]:
#print()
return new_logit
else:
new_logit[original_label] = new_logit[new_label]*amplifier
return new_logit
def perturb(prediction, epsilon, grad, logit_original, logit = True):
sign = grad.sign()
logit_new = prediction + epsilon * sign
logit_diff = F.l1_loss(logit_new, logit_original)
#print('*************logit_diff:',logit_diff.item())
#calculate accuracy for perturbed images
original_label = torch.max(logit_original, 1)[1].cpu().numpy()
new_label = torch.max(logit_new, 1)[1].cpu().numpy()
orig_label_onehot = F.one_hot(torch.tensor(original_label), 530)
orig_label_onehot = torch.tensor(orig_label_onehot, dtype=torch.uint8)
# to keep the max unchanged
logit_new[orig_label_onehot] = torch.max(logit_new, 1)[0]*1.01
accu = np.sum(original_label == new_label)/original_label.shape[0]
output = F.softmax(logit_new, dim=1)
#print(accu)
#output_diff = F.mse_loss(output, F.softmax(prediction, dim=1))
#print('************output_diff', output_diff.item())
if logit == True:
return logit_new
else:
return output
#main outer function
def add_noise(classifier, inversion, device, data_loader, epsilon, num_step):
classifier.eval()
inversion.eval()
diff = 0
recon_err = 0
correct = 0
l1max = 0
plot = True
for batch_idx, (data_, target_) in enumerate(data_loader):
data = data_.to(device)
data_ = data_.to(device)
target_ = target_.to(device)
print('############batch:',batch_idx)
with torch.no_grad():
logit = classifier(data, logit = True)
for i in range(num_step):
#print('========perturbation iteration:',i)
#create new tensor for further perturbation
logit = logit.clone().detach().to(device)
logit.requires_grad = True
reconstruction = inversion(F.softmax(logit, dim=1))
loss =F.mse_loss(reconstruction, data)
loss.backward()
logit_grad = logit.grad.data
if i == 0:
original_logit = logit
perturbation = perturb(logit, epsilon, logit_grad, original_logit)
#perturbation= perturbation.to(device)
pert_recon = inversion(F.softmax(perturbation, dim=1))
logit = perturbation
#for plot use
'''
if i == 0:
truth = data[0:8]
inverse = reconstruction[0:8]
out = torch.cat((truth, inverse))
#out = torch.cat((out, pert_recon[0:8]))
out = out = torch.cat((out, pert_recon[0:8]))
'''
#data = pert_recon.clone().detach().to(device)
# test defense use
'''
l1 = F.l1_loss(original_logit, perturbation).max().item()
if l1>l1max:
l1max = l1
diff += F.l1_loss(original_logit, perturbation, reduction='sum').item()
recon_err += F.mse_loss(pert_recon, data_, reduction='sum').item()
label = perturbation.max(1, keepdim=True)[1]
correct += label.eq(target_.view_as(label)).sum().item()
if plot:
truth = data_[0:32]
inverse = pert_recon[0:32]
out = torch.cat((inverse, truth))
for i in range(4):
out[i * 16:i * 16 + 8] = inverse[i * 8:i * 8 + 8]
out[i * 16 + 8:i * 16 + 16] = truth[i * 8:i * 8 + 8]
vutils.save_image(out, 'out/recon_def_adv.png', normalize=False)
plot = False
'''
#to save img and their result
if batch_idx == 0:
img = data_
result = perturbation
else:
img = torch.cat((img,data_))
result = torch.cat((result,perturbation))
img = img.cpu().numpy()
result = result.detach().cpu().numpy()
print(img.shape)
print(result.shape)
#generating dataset
np.savez("celeba_def1.npz", images = img, out = result)
'''
diff /= len(data_loader.dataset)*530
recon_err /= len(data_loader.dataset)*64*64
correct /= len(data_loader.dataset)
print('diff:', diff)
print('recon_err:', recon_err)
print('accu:',correct)
print('l1max:',l1max)
print('**********************')
'''
return
#main outer function
def add_noise_2inv(classifier, inversion, inversion2, device, data_loader, epsilon, num_step):
classifier.eval()
inversion.eval()
inversion2.eval()
diff = 0
recon_err = 0
correct = 0
l1max = 0
plot = True
for batch_idx, (data_, target_) in enumerate(data_loader):
data = data_.to(device)
data_ = data_.to(device)
target_ = target_.to(device)
print('############batch:',batch_idx)
with torch.no_grad():
logit = classifier(data, logit = True)
for i in range(num_step):
#print('========perturbation iteration:',i)
#create new tensor for further perturbation
logit = logit.clone().detach().to(device)
logit.requires_grad = True
reconstruction = inversion(F.softmax(logit, dim=1))
loss =F.mse_loss(reconstruction, data)
loss.backward()
logit_grad = logit.grad.data
if i == 0:
original_logit = logit
perturbation = perturb(logit, epsilon, logit_grad, original_logit)
pert_recon = inversion(F.softmax(perturbation, dim=1))
logit = perturbation
# test defense use
pert_recon = inversion2(F.softmax(perturbation, dim=1))
l1 = F.l1_loss(original_logit, perturbation).max().item()
if l1>l1max:
l1max = l1
diff += F.l1_loss(original_logit, perturbation, reduction='sum').item()
recon_err += F.mse_loss(pert_recon, data_, reduction='sum').item()
#label = perturbation.max(1, keepdim=True)[1]
#correct += label.eq(target_.view_as(label)).sum().item()
if plot:
truth = data_[0:32]
inverse = pert_recon[0:32]
out = torch.cat((inverse, truth))
for i in range(4):
out[i * 16:i * 16 + 8] = inverse[i * 8:i * 8 + 8]
out[i * 16 + 8:i * 16 + 16] = truth[i * 8:i * 8 + 8]
vutils.save_image(out, 'out/recon_def_adv222.png', normalize=False)
plot = False
diff /= len(data_loader.dataset)*530
recon_err /= len(data_loader.dataset)*64*64
#correct /= len(data_loader.dataset)
print('diff:', diff)
print('recon_err:', recon_err)
#print('accu:',correct)
print('l1max:',l1max)
print('**********************')
return
def inv_test(classifier, inversion, device, data_loader, epoch = 100, msg = 'test'):
classifier.eval()
inversion.eval()
mse_loss = 0
plot = True
with torch.no_grad():
for data, target in data_loader:
data, target = data.to(device), target.to(device)
prediction = classifier(data, release=True)
reconstruction = inversion(prediction)
mse_loss += F.mse_loss(reconstruction, data, reduction='sum').item()
if plot:
truth = data[0:32]
inverse = reconstruction[0:32]
out = torch.cat((inverse, truth))
for i in range(4):
out[i * 16:i * 16 + 8] = inverse[i * 8:i * 8 + 8]
out[i * 16 + 8:i * 16 + 16] = truth[i * 8:i * 8 + 8]
vutils.save_image(out, 'out2/recon_{}_{}_vanilla.png'.format(msg.replace(" ", ""), epoch), normalize=False)
plot = False
mse_loss /= len(data_loader.dataset) * 64 * 64
print('\nTest inversion model on {} set: Average MSE loss: {:.6f}\n'.format(msg, mse_loss))
return mse_loss
#test with output directly
def inv_test2(inversion, device, data_loader, epoch = 100, msg = 'test'):
#classifier.eval()
inversion.eval()
mse_loss = 0
plot = True
with torch.no_grad():
for data, out in data_loader:
data, out = data.to(device), out.to(device)
#prediction = classifier(data, release=True)
reconstruction = inversion(out)
mse_loss += F.mse_loss(reconstruction, data, reduction='sum').item()
if plot:
truth = data[0:32]
inverse = reconstruction[0:32]
out = torch.cat((inverse, truth))
for i in range(4):
out[i * 16:i * 16 + 8] = inverse[i * 8:i * 8 + 8]
out[i * 16 + 8:i * 16 + 16] = truth[i * 8:i * 8 + 8]
vutils.save_image(out, 'out2/recon_{}_{}_def.png'.format(msg.replace(" ", ""), epoch), normalize=False)
plot = False
mse_loss /= len(data_loader.dataset) * 64 * 64
print('\nTest inversion model on defensed {} set: Average MSE loss: {:.6f}\n'.format(msg, mse_loss))
return mse_loss
def main():
args = parser.parse_args()
print("================================")
print(args)
print("================================")
use_cuda = not args.no_cuda and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': args.num_workers, 'pin_memory': True} if use_cuda else {}
torch.manual_seed(args.seed)
transform = transforms.Compose([transforms.ToTensor()])
celeb_set = CelebA('./celeba_5w_255.npy', transform=transform)
# Inversion attack on TRAIN data of facescrub classifier
face_set = FaceScrub('./facescrub.npz', transform=transform, train=False)
#celeb_set = CelebA('./celeba_5w_255.npy', transform=transform)
face_set = FaceScrub_out('./face_def.npz', transform=transform, train=False)
celeb_loader = torch.utils.data.DataLoader(celeb_set, batch_size=args.celeb_batch_size, shuffle=True, **kwargs)
face_loader = torch.utils.data.DataLoader(face_set, batch_size=args.face_batch_size, shuffle=False, **kwargs)
classifier = nn.DataParallel(Classifier(nc=args.nc, ndf=args.ndf, nz=args.nz)).to(device)
inversion = nn.DataParallel(Inversion(nc=args.nc, ngf=args.ngf, nz=args.nz, truncation=args.truncation, c=args.c)).to(device)
inversion2 = nn.DataParallel(Inversion(nc=args.nc, ngf=args.ngf, nz=args.nz, truncation=args.truncation, c=args.c)).to(device)
model_path = 'model/model_dict.pth'
inversion_path = 'model/inv_model.pth'
inversion2_path = 'model/inv_def.pth'
try:
model_checkpoint = torch.load(model_path)
classifier.load_state_dict(model_checkpoint)
except:
print("=> load classifier checkpoint '{}' failed".format(model_path))
return
try:
checkpoint = torch.load(inversion2_path)
inversion2.load_state_dict(checkpoint)
except:
print("=> load classifier checkpoint '{}' failed".format(model_path))
return
try:
inv_checkpoint = torch.load(inversion_path)
inversion.load_state_dict(inv_checkpoint)
except:
print("=> load classifier checkpoint '{}' failed".format(inversion_path))
return
epsilon = args.epsilon
num_step = args.num_step
#add_noise(classifier, inversion, device, celeb_loader, epsilon, num_step)
add_noise_2inv(classifier, inversion, inversion2, device, face_loader, epsilon, num_step)
#inv_test2(inversion, device, face_loader)
#inv_test(classifier, inversion, device, face_loader)
if __name__ == '__main__':
main()