-
Notifications
You must be signed in to change notification settings - Fork 47
/
l2_attack_black.py
672 lines (613 loc) · 32.2 KB
/
l2_attack_black.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
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
## l2_attack_black.py -- attack a black-box network optimizing for l_2 distance
##
## Copyright (C) IBM Corp, 2017-2018
## Copyright (C) 2017, Huan Zhang <[email protected]>.
## Copyright (C) 2016, Nicholas Carlini <[email protected]>.
##
## This program is licenced under the BSD 2-Clause licence,
## contained in the LICENCE file in this directory.
import sys
import os
import tensorflow as tf
import numpy as np
import scipy.misc
from numba import jit
import math
import time
BINARY_SEARCH_STEPS = 1 # number of times to adjust the constant with binary search
MAX_ITERATIONS = 10000 # number of iterations to perform gradient descent
ABORT_EARLY = True # if we stop improving, abort gradient descent early
LEARNING_RATE = 2e-3 # larger values converge faster to less accurate results
TARGETED = True # should we target one specific class? or just be wrong?
CONFIDENCE = 0 # how strong the adversarial example should be
INITIAL_CONST = 0.5 # the initial constant c to pick as a first guess
@jit(nopython=True)
def coordinate_ADAM(losses, indice, grad, hess, batch_size, mt_arr, vt_arr, real_modifier, up, down, lr, adam_epoch, beta1, beta2, proj):
# indice = np.array(range(0, 3*299*299), dtype = np.int32)
for i in range(batch_size):
grad[i] = (losses[i*2+1] - losses[i*2+2]) / 0.0002
# true_grads = self.sess.run(self.grad_op, feed_dict={self.modifier: self.real_modifier})
# true_grads, losses, l2s, scores, nimgs = self.sess.run([self.grad_op, self.loss, self.l2dist, self.output, self.newimg], feed_dict={self.modifier: self.real_modifier})
# grad = true_grads[0].reshape(-1)[indice]
# print(grad, true_grads[0].reshape(-1)[indice])
# self.real_modifier.reshape(-1)[indice] -= self.LEARNING_RATE * grad
# self.real_modifier -= self.LEARNING_RATE * true_grads[0]
# ADAM update
mt = mt_arr[indice]
mt = beta1 * mt + (1 - beta1) * grad
mt_arr[indice] = mt
vt = vt_arr[indice]
vt = beta2 * vt + (1 - beta2) * (grad * grad)
vt_arr[indice] = vt
# epoch is an array; for each index we can have a different epoch number
epoch = adam_epoch[indice]
corr = (np.sqrt(1 - np.power(beta2,epoch))) / (1 - np.power(beta1, epoch))
m = real_modifier.reshape(-1)
old_val = m[indice]
old_val -= lr * corr * mt / (np.sqrt(vt) + 1e-8)
# set it back to [-0.5, +0.5] region
if proj:
old_val = np.maximum(np.minimum(old_val, up[indice]), down[indice])
# print(grad)
# print(old_val - m[indice])
m[indice] = old_val
adam_epoch[indice] = epoch + 1
@jit(nopython=True)
def coordinate_Newton(losses, indice, grad, hess, batch_size, mt_arr, vt_arr, real_modifier, up, down, lr, adam_epoch, beta1, beta2, proj):
# def sign(x):
# return np.piecewise(x, [x < 0, x >= 0], [-1, 1])
cur_loss = losses[0]
for i in range(batch_size):
grad[i] = (losses[i*2+1] - losses[i*2+2]) / 0.0002
hess[i] = (losses[i*2+1] - 2 * cur_loss + losses[i*2+2]) / (0.0001 * 0.0001)
# print("New epoch:")
# print('grad', grad)
# print('hess', hess)
# hess[hess < 0] = 1.0
# hess[np.abs(hess) < 0.1] = sign(hess[np.abs(hess) < 0.1]) * 0.1
# negative hessian cannot provide second order information, just do a gradient descent
hess[hess < 0] = 1.0
# hessian too small, could be numerical problems
hess[hess < 0.1] = 0.1
# print(hess)
m = real_modifier.reshape(-1)
old_val = m[indice]
old_val -= lr * grad / hess
# set it back to [-0.5, +0.5] region
if proj:
old_val = np.maximum(np.minimum(old_val, up[indice]), down[indice])
# print('delta', old_val - m[indice])
m[indice] = old_val
# print(m[indice])
@jit(nopython=True)
def coordinate_Newton_ADAM(losses, indice, grad, hess, batch_size, mt_arr, vt_arr, real_modifier, up, down, lr, adam_epoch, beta1, beta2, proj):
cur_loss = losses[0]
for i in range(batch_size):
grad[i] = (losses[i*2+1] - losses[i*2+2]) / 0.0002
hess[i] = (losses[i*2+1] - 2 * cur_loss + losses[i*2+2]) / (0.0001 * 0.0001)
# print("New epoch:")
# print(grad)
# print(hess)
# positive hessian, using newton's method
hess_indice = (hess >= 0)
# print(hess_indice)
# negative hessian, using ADAM
adam_indice = (hess < 0)
# print(adam_indice)
# print(sum(hess_indice), sum(adam_indice))
hess[hess < 0] = 1.0
hess[hess < 0.1] = 0.1
# hess[np.abs(hess) < 0.1] = sign(hess[np.abs(hess) < 0.1]) * 0.1
# print(adam_indice)
# Newton's Method
m = real_modifier.reshape(-1)
old_val = m[indice[hess_indice]]
old_val -= lr * grad[hess_indice] / hess[hess_indice]
# set it back to [-0.5, +0.5] region
if proj:
old_val = np.maximum(np.minimum(old_val, up[indice[hess_indice]]), down[indice[hess_indice]])
m[indice[hess_indice]] = old_val
# ADMM
mt = mt_arr[indice]
mt = beta1 * mt + (1 - beta1) * grad
mt_arr[indice] = mt
vt = vt_arr[indice]
vt = beta2 * vt + (1 - beta2) * (grad * grad)
vt_arr[indice] = vt
# epoch is an array; for each index we can have a different epoch number
epoch = adam_epoch[indice]
corr = (np.sqrt(1 - np.power(beta2,epoch[adam_indice]))) / (1 - np.power(beta1, epoch[adam_indice]))
old_val = m[indice[adam_indice]]
old_val -= lr * corr * mt[adam_indice] / (np.sqrt(vt[adam_indice]) + 1e-8)
# old_val -= lr * grad[adam_indice]
# set it back to [-0.5, +0.5] region
if proj:
old_val = np.maximum(np.minimum(old_val, up[indice[adam_indice]]), down[indice[adam_indice]])
m[indice[adam_indice]] = old_val
adam_epoch[indice] = epoch + 1
# print(m[indice])
class BlackBoxL2:
def __init__(self, sess, model, batch_size=1, confidence = CONFIDENCE,
targeted = TARGETED, learning_rate = LEARNING_RATE,
binary_search_steps = BINARY_SEARCH_STEPS, max_iterations = MAX_ITERATIONS, print_every = 100, early_stop_iters = 0,
abort_early = ABORT_EARLY,
initial_const = INITIAL_CONST,
use_log = False, use_tanh = True, use_resize = False, adam_beta1 = 0.9, adam_beta2 = 0.999, reset_adam_after_found = False,
solver = "adam", save_ckpts = "", load_checkpoint = "", start_iter = 0,
init_size = 32, use_importance = True):
"""
The L_2 optimized attack.
This attack is the most efficient and should be used as the primary
attack to evaluate potential defenses.
Returns adversarial examples for the supplied model.
confidence: Confidence of adversarial examples: higher produces examples
that are farther away, but more strongly classified as adversarial.
batch_size: Number of gradient evaluations to run simultaneously.
targeted: True if we should perform a targetted attack, False otherwise.
learning_rate: The learning rate for the attack algorithm. Smaller values
produce better results but are slower to converge.
binary_search_steps: The number of times we perform binary search to
find the optimal tradeoff-constant between distance and confidence.
max_iterations: The maximum number of iterations. Larger values are more
accurate; setting too small will require a large learning rate and will
produce poor results.
abort_early: If true, allows early aborts if gradient descent gets stuck.
initial_const: The initial tradeoff-constant to use to tune the relative
importance of distance and confidence. If binary_search_steps is large,
the initial constant is not important.
"""
image_size, num_channels, num_labels = model.image_size, model.num_channels, model.num_labels
self.model = model
self.sess = sess
self.TARGETED = targeted
self.LEARNING_RATE = learning_rate
self.MAX_ITERATIONS = max_iterations
self.print_every = print_every
self.early_stop_iters = early_stop_iters if early_stop_iters != 0 else max_iterations // 10
print("early stop:", self.early_stop_iters)
self.BINARY_SEARCH_STEPS = binary_search_steps
self.ABORT_EARLY = abort_early
self.CONFIDENCE = confidence
self.initial_const = initial_const
self.start_iter = start_iter
self.batch_size = batch_size
self.num_channels = num_channels
self.resize_init_size = init_size
self.use_importance = use_importance
if use_resize:
self.small_x = self.resize_init_size
self.small_y = self.resize_init_size
else:
self.small_x = image_size
self.small_y = image_size
self.use_tanh = use_tanh
self.use_resize = use_resize
self.save_ckpts = save_ckpts
if save_ckpts:
os.system("mkdir -p {}".format(save_ckpts))
self.repeat = binary_search_steps >= 10
# each batch has a different modifier value (see below) to evaluate
# small_shape = (None,self.small_x,self.small_y,num_channels)
shape = (None,image_size,image_size,num_channels)
single_shape = (image_size, image_size, num_channels)
small_single_shape = (self.small_x, self.small_y, num_channels)
# the variable we're going to optimize over
# support multiple batches
# support any size image, will be resized to model native size
if self.use_resize:
self.modifier = tf.placeholder(tf.float32, shape=(None, None, None, None))
# scaled up image
self.scaled_modifier = tf.image.resize_images(self.modifier, [image_size, image_size])
# operator used for resizing image
self.resize_size_x = tf.placeholder(tf.int32)
self.resize_size_y = tf.placeholder(tf.int32)
self.resize_input = tf.placeholder(tf.float32, shape=(1, None, None, None))
self.resize_op = tf.image.resize_images(self.resize_input, [self.resize_size_x, self.resize_size_y])
else:
self.modifier = tf.placeholder(tf.float32, shape=(None, image_size, image_size, num_channels))
# no resize
self.scaled_modifier = self.modifier
# the real variable, initialized to 0
self.load_checkpoint = load_checkpoint
if load_checkpoint:
# if checkpoint is incorrect reshape will fail
print("Using checkpint", load_checkpoint)
self.real_modifier = np.load(load_checkpoint).reshape((1,) + small_single_shape)
else:
self.real_modifier = np.zeros((1,) + small_single_shape, dtype=np.float32)
# self.real_modifier = np.random.randn(image_size * image_size * num_channels).astype(np.float32).reshape((1,) + single_shape)
# self.real_modifier /= np.linalg.norm(self.real_modifier)
# these are variables to be more efficient in sending data to tf
# we only work on 1 image at once; the batch is for evaluation loss at different modifiers
self.timg = tf.Variable(np.zeros(single_shape), dtype=tf.float32)
self.tlab = tf.Variable(np.zeros(num_labels), dtype=tf.float32)
self.const = tf.Variable(0.0, dtype=tf.float32)
# and here's what we use to assign them
self.assign_timg = tf.placeholder(tf.float32, single_shape)
self.assign_tlab = tf.placeholder(tf.float32, num_labels)
self.assign_const = tf.placeholder(tf.float32)
# the resulting image, tanh'd to keep bounded from -0.5 to 0.5
# broadcast self.timg to every dimension of modifier
if use_tanh:
self.newimg = tf.tanh(self.scaled_modifier + self.timg)/2
else:
self.newimg = self.scaled_modifier + self.timg
# prediction BEFORE-SOFTMAX of the model
# now we have output at #batch_size different modifiers
# the output should have shape (batch_size, num_labels)
self.output = model.predict(self.newimg)
# distance to the input data
if use_tanh:
self.l2dist = tf.reduce_sum(tf.square(self.newimg-tf.tanh(self.timg)/2), [1,2,3])
else:
self.l2dist = tf.reduce_sum(tf.square(self.newimg - self.timg), [1,2,3])
# compute the probability of the label class versus the maximum other
# self.tlab * self.output selects the Z value of real class
# because self.tlab is an one-hot vector
# the reduce_sum removes extra zeros, now get a vector of size #batch_size
self.real = tf.reduce_sum((self.tlab)*self.output,1)
# (1-self.tlab)*self.output gets all Z values for other classes
# Because soft Z values are negative, it is possible that all Z values are less than 0
# and we mistakenly select the real class as the max. So we minus 10000 for real class
self.other = tf.reduce_max((1-self.tlab)*self.output - (self.tlab*10000),1)
# If self.targeted is true, then the targets represents the target labels.
# If self.targeted is false, then targets are the original class labels.
if self.TARGETED:
if use_log:
# loss1 = - tf.log(self.real)
loss1 = tf.maximum(0.0, tf.log(self.other + 1e-30) - tf.log(self.real + 1e-30))
else:
# if targetted, optimize for making the other class (real) most likely
loss1 = tf.maximum(0.0, self.other-self.real+self.CONFIDENCE)
else:
if use_log:
# loss1 = tf.log(self.real)
loss1 = tf.maximum(0.0, tf.log(self.real + 1e-30) - tf.log(self.other + 1e-30))
else:
# if untargeted, optimize for making this class least likely.
loss1 = tf.maximum(0.0, self.real-self.other+self.CONFIDENCE)
# sum up the losses (output is a vector of #batch_size)
self.loss2 = self.l2dist
self.loss1 = self.const*loss1
self.loss = self.loss1+self.loss2
# these are the variables to initialize when we run
self.setup = []
self.setup.append(self.timg.assign(self.assign_timg))
self.setup.append(self.tlab.assign(self.assign_tlab))
self.setup.append(self.const.assign(self.assign_const))
# prepare the list of all valid variables
var_size = self.small_x * self.small_y * num_channels
self.use_var_len = var_size
self.var_list = np.array(range(0, self.use_var_len), dtype = np.int32)
self.used_var_list = np.zeros(var_size, dtype = np.int32)
self.sample_prob = np.ones(var_size, dtype = np.float32) / var_size
# upper and lower bounds for the modifier
self.modifier_up = np.zeros(var_size, dtype = np.float32)
self.modifier_down = np.zeros(var_size, dtype = np.float32)
# random permutation for coordinate update
self.perm = np.random.permutation(var_size)
self.perm_index = 0
# ADAM status
self.mt = np.zeros(var_size, dtype = np.float32)
self.vt = np.zeros(var_size, dtype = np.float32)
# self.beta1 = 0.8
# self.beta2 = 0.99
self.beta1 = adam_beta1
self.beta2 = adam_beta2
self.reset_adam_after_found = reset_adam_after_found
self.adam_epoch = np.ones(var_size, dtype = np.int32)
self.stage = 0
# variables used during optimization process
self.grad = np.zeros(batch_size, dtype = np.float32)
self.hess = np.zeros(batch_size, dtype = np.float32)
# for testing
self.grad_op = tf.gradients(self.loss, self.modifier)
# compile numba function
# self.coordinate_ADAM_numba = jit(coordinate_ADAM, nopython = True)
# self.coordinate_ADAM_numba.recompile()
# print(self.coordinate_ADAM_numba.inspect_llvm())
# np.set_printoptions(threshold=np.nan)
# set solver
solver = solver.lower()
self.solver_name = solver
if solver == "adam":
self.solver = coordinate_ADAM
elif solver == "newton":
self.solver = coordinate_Newton
elif solver == "adam_newton":
self.solver = coordinate_Newton_ADAM
elif solver != "fake_zero":
print("unknown solver", solver)
self.solver = coordinate_ADAM
print("Using", solver, "solver")
def max_pooling(self, image, size):
img_pool = np.copy(image)
img_x = image.shape[0]
img_y = image.shape[1]
for i in range(0, img_x, size):
for j in range(0, img_y, size):
img_pool[i:i+size, j:j+size] = np.max(image[i:i+size, j:j+size])
return img_pool
def get_new_prob(self, prev_modifier, gen_double = False):
prev_modifier = np.squeeze(prev_modifier)
old_shape = prev_modifier.shape
if gen_double:
new_shape = (old_shape[0]*2, old_shape[1]*2, old_shape[2])
else:
new_shape = old_shape
prob = np.empty(shape=new_shape, dtype = np.float32)
for i in range(prev_modifier.shape[2]):
image = np.abs(prev_modifier[:,:,i])
image_pool = self.max_pooling(image, old_shape[0] // 8)
if gen_double:
prob[:,:,i] = scipy.misc.imresize(image_pool, 2.0, 'nearest', mode = 'F')
else:
prob[:,:,i] = image_pool
prob /= np.sum(prob)
return prob
def resize_img(self, small_x, small_y, reset_only = False):
self.small_x = small_x
self.small_y = small_y
small_single_shape = (self.small_x, self.small_y, self.num_channels)
if reset_only:
self.real_modifier = np.zeros((1,) + small_single_shape, dtype=np.float32)
else:
# run the resize_op once to get the scaled image
prev_modifier = np.copy(self.real_modifier)
self.real_modifier = self.sess.run(self.resize_op, feed_dict={self.resize_size_x: self.small_x, self.resize_size_y: self.small_y, self.resize_input: self.real_modifier})
# prepare the list of all valid variables
var_size = self.small_x * self.small_y * self.num_channels
self.use_var_len = var_size
self.var_list = np.array(range(0, self.use_var_len), dtype = np.int32)
# ADAM status
self.mt = np.zeros(var_size, dtype = np.float32)
self.vt = np.zeros(var_size, dtype = np.float32)
self.adam_epoch = np.ones(var_size, dtype = np.int32)
# update sample probability
if reset_only:
self.sample_prob = np.ones(var_size, dtype = np.float32) / var_size
else:
self.sample_prob = self.get_new_prob(prev_modifier, True)
self.sample_prob = self.sample_prob.reshape(var_size)
def fake_blackbox_optimizer(self):
true_grads, losses, l2s, loss1, loss2, scores, nimgs = self.sess.run([self.grad_op, self.loss, self.l2dist, self.loss1, self.loss2, self.output, self.newimg], feed_dict={self.modifier: self.real_modifier})
# ADAM update
grad = true_grads[0].reshape(-1)
# print(true_grads[0])
epoch = self.adam_epoch[0]
mt = self.beta1 * self.mt + (1 - self.beta1) * grad
vt = self.beta2 * self.vt + (1 - self.beta2) * np.square(grad)
corr = (math.sqrt(1 - self.beta2 ** epoch)) / (1 - self.beta1 ** epoch)
# print(grad.shape, mt.shape, vt.shape, self.real_modifier.shape)
# m is a *view* of self.real_modifier
m = self.real_modifier.reshape(-1)
# this is in-place
m -= self.LEARNING_RATE * corr * (mt / (np.sqrt(vt) + 1e-8))
self.mt = mt
self.vt = vt
# m -= self.LEARNING_RATE * grad
if not self.use_tanh:
m_proj = np.maximum(np.minimum(m, self.modifier_up), self.modifier_down)
np.copyto(m, m_proj)
self.adam_epoch[0] = epoch + 1
return losses[0], l2s[0], loss1[0], loss2[0], scores[0], nimgs[0]
def blackbox_optimizer(self, iteration):
# build new inputs, based on current variable value
var = np.repeat(self.real_modifier, self.batch_size * 2 + 1, axis=0)
var_size = self.real_modifier.size
# print(s, "variables remaining")
# var_indice = np.random.randint(0, self.var_list.size, size=self.batch_size)
if self.use_importance:
var_indice = np.random.choice(self.var_list.size, self.batch_size, replace=False, p = self.sample_prob)
else:
var_indice = np.random.choice(self.var_list.size, self.batch_size, replace=False)
indice = self.var_list[var_indice]
# indice = self.var_list
# regenerate the permutations if we run out
# if self.perm_index + self.batch_size >= var_size:
# self.perm = np.random.permutation(var_size)
# self.perm_index = 0
# indice = self.perm[self.perm_index:self.perm_index + self.batch_size]
# b[0] has the original modifier, b[1] has one index added 0.0001
for i in range(self.batch_size):
var[i * 2 + 1].reshape(-1)[indice[i]] += 0.0001
var[i * 2 + 2].reshape(-1)[indice[i]] -= 0.0001
losses, l2s, loss1, loss2, scores, nimgs = self.sess.run([self.loss, self.l2dist, self.loss1, self.loss2, self.output, self.newimg], feed_dict={self.modifier: var})
# losses = self.sess.run(self.loss, feed_dict={self.modifier: var})
# t_grad = self.sess.run(self.grad_op, feed_dict={self.modifier: self.real_modifier})
# self.grad = t_grad[0].reshape(-1)
# true_grads = self.sess.run(self.grad_op, feed_dict={self.modifier: self.real_modifier})
# self.coordinate_ADAM_numba(losses, indice, self.grad, self.hess, self.batch_size, self.mt, self.vt, self.real_modifier, self.modifier_up, self.modifier_down, self.LEARNING_RATE, self.adam_epoch, self.beta1, self.beta2, not self.use_tanh)
# coordinate_ADAM(losses, indice, self.grad, self.hess, self.batch_size, self.mt, self.vt, self.real_modifier, self.modifier_up, self.modifier_down, self.LEARNING_RATE, self.adam_epoch, self.beta1, self.beta2, not self.use_tanh)
# coordinate_ADAM(losses, indice, self.grad, self.hess, self.batch_size, self.mt, self.vt, self.real_modifier, self.modifier_up, self.modifier_down, self.LEARNING_RATE, self.adam_epoch, self.beta1, self.beta2, not self.use_tanh, true_grads)
# coordinate_Newton(losses, indice, self.grad, self.hess, self.batch_size, self.mt, self.vt, self.real_modifier, self.modifier_up, self.modifier_down, self.LEARNING_RATE, self.adam_epoch, self.beta1, self.beta2, not self.use_tanh)
# coordinate_Newton_ADAM(losses, indice, self.grad, self.hess, self.batch_size, self.mt, self.vt, self.real_modifier, self.modifier_up, self.modifier_down, self.LEARNING_RATE, self.adam_epoch, self.beta1, self.beta2, not self.use_tanh)
self.solver(losses, indice, self.grad, self.hess, self.batch_size, self.mt, self.vt, self.real_modifier, self.modifier_up, self.modifier_down, self.LEARNING_RATE, self.adam_epoch, self.beta1, self.beta2, not self.use_tanh)
# adjust sample probability, sample around the points with large gradient
if self.save_ckpts:
np.save('{}/iter{}'.format(self.save_ckpts, iteration), self.real_modifier)
if self.real_modifier.shape[0] > self.resize_init_size:
self.sample_prob = self.get_new_prob(self.real_modifier)
# self.sample_prob = self.get_new_prob(tmp_mt.reshape(self.real_modifier.shape))
self.sample_prob = self.sample_prob.reshape(var_size)
# if the gradient is too small, do not optimize on this variable
# self.var_list = np.delete(self.var_list, indice[np.abs(self.grad) < 5e-3])
# reset the list every 10000 iterations
# if iteration%200 == 0:
# print("{} variables remained at last stage".format(self.var_list.size))
# var_size = self.real_modifier.size
# self.var_list = np.array(range(0, var_size))
return losses[0], l2s[0], loss1[0], loss2[0], scores[0], nimgs[0]
# return losses[0]
def attack(self, imgs, targets):
"""
Perform the L_2 attack on the given images for the given targets.
If self.targeted is true, then the targets represents the target labels.
If self.targeted is false, then targets are the original class labels.
"""
r = []
print('go up to',len(imgs))
# we can only run 1 image at a time, minibatches are used for gradient evaluation
for i in range(0,len(imgs)):
print('tick',i)
r.extend(self.attack_batch(imgs[i], targets[i]))
return np.array(r)
# only accepts 1 image at a time. Batch is used for gradient evaluations at different points
def attack_batch(self, img, lab):
"""
Run the attack on a batch of images and labels.
"""
def compare(x,y):
if not isinstance(x, (float, int, np.int64)):
x = np.copy(x)
if self.TARGETED:
x[y] -= self.CONFIDENCE
else:
x[y] += self.CONFIDENCE
x = np.argmax(x)
if self.TARGETED:
return x == y
else:
return x != y
# remove the extra batch dimension
if len(img.shape) == 4:
img = img[0]
if len(lab.shape) == 2:
lab = lab[0]
# convert to tanh-space
if self.use_tanh:
img = np.arctanh(img*1.999999)
# set the lower and upper bounds accordingly
lower_bound = 0.0
CONST = self.initial_const
upper_bound = 1e10
# convert img to float32 to avoid numba error
img = img.astype(np.float32)
# set the upper and lower bounds for the modifier
if not self.use_tanh:
self.modifier_up = 0.5 - img.reshape(-1)
self.modifier_down = -0.5 - img.reshape(-1)
# clear the modifier
if not self.load_checkpoint:
if self.use_resize:
self.resize_img(self.resize_init_size, self.resize_init_size, True)
else:
self.real_modifier.fill(0.0)
# the best l2, score, and image attack
o_best_const = CONST
o_bestl2 = 1e10
o_bestscore = -1
o_bestattack = img
for outer_step in range(self.BINARY_SEARCH_STEPS):
print(o_bestl2)
bestl2 = 1e10
bestscore = -1
# The last iteration (if we run many steps) repeat the search once.
if self.repeat == True and outer_step == self.BINARY_SEARCH_STEPS-1:
CONST = upper_bound
# set the variables so that we don't have to send them over again
self.sess.run(self.setup, {self.assign_timg: img,
self.assign_tlab: lab,
self.assign_const: CONST})
# use the current best model
# np.copyto(self.real_modifier, o_bestattack - img)
# use the model left by last constant change
prev = 1e6
train_timer = 0.0
last_loss1 = 1.0
if not self.load_checkpoint:
if self.use_resize:
self.resize_img(self.resize_init_size, self.resize_init_size, True)
else:
self.real_modifier.fill(0.0)
# reset ADAM status
self.mt.fill(0.0)
self.vt.fill(0.0)
self.adam_epoch.fill(1)
self.stage = 0
multiplier = 1
eval_costs = 0
if self.solver_name != "fake_zero":
multiplier = 24
for iteration in range(self.start_iter, self.MAX_ITERATIONS):
if self.use_resize:
if iteration == 2000:
# if iteration == 2000 // 24:
self.resize_img(64,64)
if iteration == 10000:
# if iteration == 2000 // 24 + (10000 - 2000) // 96:
self.resize_img(128,128)
# if iteration == 200*30:
# if iteration == 250 * multiplier:
# self.resize_img(256,256)
# print out the losses every 10%
if iteration%(self.print_every) == 0:
# print(iteration,self.sess.run((self.loss,self.real,self.other,self.loss1,self.loss2), feed_dict={self.modifier: self.real_modifier}))
loss, real, other, loss1, loss2 = self.sess.run((self.loss,self.real,self.other,self.loss1,self.loss2), feed_dict={self.modifier: self.real_modifier})
print("[STATS][L2] iter = {}, cost = {}, time = {:.3f}, size = {}, loss = {:.5g}, real = {:.5g}, other = {:.5g}, loss1 = {:.5g}, loss2 = {:.5g}".format(iteration, eval_costs, train_timer, self.real_modifier.shape, loss[0], real[0], other[0], loss1[0], loss2[0]))
sys.stdout.flush()
# np.save('black_iter_{}'.format(iteration), self.real_modifier)
attack_begin_time = time.time()
# perform the attack
if self.solver_name == "fake_zero":
l, l2, loss1, loss2, score, nimg = self.fake_blackbox_optimizer()
else:
l, l2, loss1, loss2, score, nimg = self.blackbox_optimizer(iteration)
# l = self.blackbox_optimizer(iteration)
if self.solver_name == "fake_zero":
eval_costs += np.prod(self.real_modifier.shape)
else:
eval_costs += self.batch_size
# reset ADAM states when a valid example has been found
if loss1 == 0.0 and last_loss1 != 0.0 and self.stage == 0:
# we have reached the fine tunning point
# reset ADAM to avoid overshoot
if self.reset_adam_after_found:
self.mt.fill(0.0)
self.vt.fill(0.0)
self.adam_epoch.fill(1)
self.stage = 1
last_loss1 = loss1
# check if we should abort search if we're getting nowhere.
# if self.ABORT_EARLY and iteration%(self.MAX_ITERATIONS//10) == 0:
if self.ABORT_EARLY and iteration % self.early_stop_iters == 0:
if l > prev*.9999:
print("Early stopping because there is no improvement")
break
prev = l
# adjust the best result found so far
# the best attack should have the target class with the largest value,
# and has smallest l2 distance
if l2 < bestl2 and compare(score, np.argmax(lab)):
bestl2 = l2
bestscore = np.argmax(score)
if l2 < o_bestl2 and compare(score, np.argmax(lab)):
# print a message if it is the first attack found
if o_bestl2 == 1e10:
print("[STATS][L3](First valid attack found!) iter = {}, cost = {}, time = {:.3f}, size = {}, loss = {:.5g}, loss1 = {:.5g}, loss2 = {:.5g}, l2 = {:.5g}".format(iteration, eval_costs, train_timer, self.real_modifier.shape, l, loss1, loss2, l2))
sys.stdout.flush()
o_bestl2 = l2
o_bestscore = np.argmax(score)
o_bestattack = nimg
o_best_const = CONST
train_timer += time.time() - attack_begin_time
# adjust the constant as needed
if compare(bestscore, np.argmax(lab)) and bestscore != -1:
# success, divide const by two
print('old constant: ', CONST)
upper_bound = min(upper_bound,CONST)
if upper_bound < 1e9:
CONST = (lower_bound + upper_bound)/2
print('new constant: ', CONST)
else:
# failure, either multiply by 10 if no solution found yet
# or do binary search with the known upper bound
print('old constant: ', CONST)
lower_bound = max(lower_bound,CONST)
if upper_bound < 1e9:
CONST = (lower_bound + upper_bound)/2
else:
CONST *= 10
print('new constant: ', CONST)
# return the best solution found
return o_bestattack, o_best_const