File size: 26,018 Bytes
e1aaaac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2020-present, Francesco Croce
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
#

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import torch
import time
import math
import torch.nn.functional as F

from autoattack.autopgd_base import L1_projection

class SquareAttack():
    """
    Square Attack
    https://arxiv.org/abs/1912.00049

    :param predict:       forward pass function
    :param norm:          Lp-norm of the attack ('Linf', 'L2' supported)
    :param n_restarts:    number of random restarts
    :param n_queries:     max number of queries (each restart)
    :param eps:           bound on the norm of perturbations
    :param seed:          random seed for the starting point
    :param p_init:        parameter to control size of squares
    :param loss:          loss function optimized ('margin', 'ce' supported)
    :param resc_schedule  adapt schedule of p to n_queries
    """

    def __init__(
            self,
            predict,
            norm='Linf',
            n_queries=5000,
            eps=None,
            p_init=.8,
            n_restarts=1,
            seed=0,
            verbose=False,
            targeted=False,
            loss='margin',
            resc_schedule=True,
            device=None):
        """
        Square Attack implementation in PyTorch
        """
        
        self.predict = predict
        self.norm = norm
        self.n_queries = n_queries
        self.eps = eps
        self.p_init = p_init
        self.n_restarts = n_restarts
        self.seed = seed
        self.verbose = verbose
        self.targeted = targeted
        self.loss = loss
        self.rescale_schedule = resc_schedule
        self.device = device
        self.return_all = False
    
    def margin_and_loss(self, x, y):
        """
        :param y:        correct labels if untargeted else target labels
        """

        logits = self.predict(x)
        xent = F.cross_entropy(logits, y, reduction='none')
        u = torch.arange(x.shape[0])
        y_corr = logits[u, y].clone()
        logits[u, y] = -float('inf')
        y_others = logits.max(dim=-1)[0]

        if not self.targeted:
            if self.loss == 'ce':
                return y_corr - y_others, -1. * xent
            elif self.loss == 'margin':
                return y_corr - y_others, y_corr - y_others
        else:
            return y_others - y_corr, xent

    def init_hyperparam(self, x):
        assert self.norm in ['Linf', 'L2', 'L1']
        assert not self.eps is None
        assert self.loss in ['ce', 'margin']

        if self.device is None:
            self.device = x.device
        self.orig_dim = list(x.shape[1:])
        self.ndims = len(self.orig_dim)
        if self.seed is None:
            self.seed = time.time()

    def random_target_classes(self, y_pred, n_classes):
        y = torch.zeros_like(y_pred)
        for counter in range(y_pred.shape[0]):
            l = list(range(n_classes))
            l.remove(y_pred[counter])
            t = self.random_int(0, len(l))
            y[counter] = l[t]

        return y.long().to(self.device)

    def check_shape(self, x):
        return x if len(x.shape) == (self.ndims + 1) else x.unsqueeze(0)

    def random_choice(self, shape):
        t = 2 * torch.rand(shape).to(self.device) - 1
        return torch.sign(t)

    def random_int(self, low=0, high=1, shape=[1]):
        t = low + (high - low) * torch.rand(shape).to(self.device)
        return t.long()

    def normalize(self, x):
        if self.norm == 'Linf':
            t = x.abs().view(x.shape[0], -1).max(1)[0]
            return x / (t.view(-1, *([1] * self.ndims)) + 1e-12)

        elif self.norm == 'L2':
            t = (x ** 2).view(x.shape[0], -1).sum(-1).sqrt()
            return x / (t.view(-1, *([1] * self.ndims)) + 1e-12)

        elif self.norm == 'L1':
            t = x.abs().view(x.shape[0], -1).sum(dim=-1)
            return x / (t.view(-1, *([1] * self.ndims)) + 1e-12)
    
    def lp_norm(self, x):
        if self.norm == 'L2':
            t = (x ** 2).view(x.shape[0], -1).sum(-1).sqrt()
            return t.view(-1, *([1] * self.ndims))

        elif self.norm == 'L1':
            t = x.abs().view(x.shape[0], -1).sum(dim=-1)
            return t.view(-1, *([1] * self.ndims))
    
    def eta_rectangles(self, x, y):
        delta = torch.zeros([x, y]).to(self.device)
        x_c, y_c = x // 2 + 1, y // 2 + 1

        counter2 = [x_c - 1, y_c - 1]
        if self.norm == 'L2':
            for counter in range(0, max(x_c, y_c)):
              delta[max(counter2[0], 0):min(counter2[0] + (2*counter + 1), x),
                  max(0, counter2[1]):min(counter2[1] + (2*counter + 1), y)
                  ] += 1.0/(torch.Tensor([counter + 1]).view(1, 1).to(
                  self.device) ** 2)
              counter2[0] -= 1
              counter2[1] -= 1
    
            delta /= (delta ** 2).sum(dim=(0, 1), keepdim=True).sqrt()
        
        elif self.norm == 'L1':
            for counter in range(0, max(x_c, y_c)):
              delta[max(counter2[0], 0):min(counter2[0] + (2*counter + 1), x),
                  max(0, counter2[1]):min(counter2[1] + (2*counter + 1), y)
                  ] += 1.0/(torch.Tensor([counter + 1]).view(1, 1).to(
                  self.device) ** 4)
              counter2[0] -= 1
              counter2[1] -= 1
    
            delta /= delta.abs().sum(dim=(), keepdim=True)
        
        return delta

    def eta(self, s):
        if self.norm == 'L2':
            delta = torch.zeros([s, s]).to(self.device)
            delta[:s // 2] = self.eta_rectangles(s // 2, s)
            delta[s // 2:] = -1. * self.eta_rectangles(s - s // 2, s)
            delta /= (delta ** 2).sum(dim=(0, 1), keepdim=True).sqrt()
        
        elif self.norm == 'L1':
            delta = torch.zeros([s, s]).to(self.device)
            delta[:s // 2] = self.eta_rectangles(s // 2, s)
            delta[s // 2:] = -1. * self.eta_rectangles(s - s // 2, s)
            #delta = self.eta_rectangles(s, s)
            delta /= delta.abs().sum(dim=(), keepdim=True)
            #delta *= (torch.rand([1]) - .5).sign().to(self.device)
        
        if torch.rand([1]) > 0.5:
            delta = delta.permute([1, 0])

        return delta

    def p_selection(self, it):
        """ schedule to decrease the parameter p """

        if self.rescale_schedule:
            it = int(it / self.n_queries * 10000)

        if 10 < it <= 50:
            p = self.p_init / 2
        elif 50 < it <= 200:
            p = self.p_init / 4
        elif 200 < it <= 500:
            p = self.p_init / 8
        elif 500 < it <= 1000:
            p = self.p_init / 16
        elif 1000 < it <= 2000:
            p = self.p_init / 32
        elif 2000 < it <= 4000:
            p = self.p_init / 64
        elif 4000 < it <= 6000:
            p = self.p_init / 128
        elif 6000 < it <= 8000:
            p = self.p_init / 256
        elif 8000 < it:
            p = self.p_init / 512
        else:
            p = self.p_init

        return p

    def attack_single_run(self, x, y):
        with torch.no_grad():
            adv = x.clone()
            c, h, w = x.shape[1:]
            n_features = c * h * w
            n_ex_total = x.shape[0]

            if self.verbose and h != w:
                print('square attack may not work properly for non-square image.')
                print('for details please refer to https://github.com/fra31/auto-attack/issues/95')

            
            if self.norm == 'Linf':
                x_best = torch.clamp(x + self.eps * self.random_choice(
                    [x.shape[0], c, 1, w]), 0., 1.)
                margin_min, loss_min = self.margin_and_loss(x_best, y)
                n_queries = torch.ones(x.shape[0]).to(self.device)
                s_init = int(math.sqrt(self.p_init * n_features / c))
                
                if (margin_min < 0.0).all():
                    return n_queries, x_best
                
                for i_iter in range(self.n_queries):
                    idx_to_fool = (margin_min > 0.0).nonzero().squeeze()
                    
                    x_curr = self.check_shape(x[idx_to_fool])
                    x_best_curr = self.check_shape(x_best[idx_to_fool])
                    y_curr = y[idx_to_fool]
                    if len(y_curr.shape) == 0:
                        y_curr = y_curr.unsqueeze(0)
                    margin_min_curr = margin_min[idx_to_fool]
                    loss_min_curr = loss_min[idx_to_fool]
                    
                    p = self.p_selection(i_iter)
                    s = max(int(round(math.sqrt(p * n_features / c))), 1)
                    s = min(s, min(h, w))
                    vh = self.random_int(0, h - s)
                    vw = self.random_int(0, w - s)
                    new_deltas = torch.zeros([c, h, w]).to(self.device)
                    new_deltas[:, vh:vh + s, vw:vw + s
                        ] = 2. * self.eps * self.random_choice([c, 1, 1])
                    
                    x_new = x_best_curr + new_deltas
                    x_new = torch.min(torch.max(x_new, x_curr - self.eps),
                        x_curr + self.eps)
                    x_new = torch.clamp(x_new, 0., 1.)
                    x_new = self.check_shape(x_new)
                    
                    margin, loss = self.margin_and_loss(x_new, y_curr)

                    # update loss if new loss is better
                    idx_improved = (loss < loss_min_curr).float()

                    loss_min[idx_to_fool] = idx_improved * loss + (
                        1. - idx_improved) * loss_min_curr

                    # update margin and x_best if new loss is better
                    # or misclassification
                    idx_miscl = (margin <= 0.).float()
                    idx_improved = torch.max(idx_improved, idx_miscl)

                    margin_min[idx_to_fool] = idx_improved * margin + (
                        1. - idx_improved) * margin_min_curr
                    idx_improved = idx_improved.reshape([-1,
                        *[1]*len(x.shape[:-1])])
                    x_best[idx_to_fool] = idx_improved * x_new + (
                        1. - idx_improved) * x_best_curr
                    n_queries[idx_to_fool] += 1.

                    ind_succ = (margin_min <= 0.).nonzero().squeeze()
                    if self.verbose and ind_succ.numel() != 0:
                        print('{}'.format(i_iter + 1),
                            '- success rate={}/{} ({:.2%})'.format(
                            ind_succ.numel(), n_ex_total,
                            float(ind_succ.numel()) / n_ex_total),
                            '- avg # queries={:.1f}'.format(
                            n_queries[ind_succ].mean().item()),
                            '- med # queries={:.1f}'.format(
                            n_queries[ind_succ].median().item()),
                            '- loss={:.3f}'.format(loss_min.mean()))

                    if ind_succ.numel() == n_ex_total:
                        break
              
            elif self.norm == 'L2':
                delta_init = torch.zeros_like(x)
                s = h // 5
                sp_init = (h - s * 5) // 2
                vh = sp_init + 0
                for _ in range(h // s):
                    vw = sp_init + 0
                    for _ in range(w // s):
                        delta_init[:, :, vh:vh + s, vw:vw + s] += self.eta(
                            s).view(1, 1, s, s) * self.random_choice(
                            [x.shape[0], c, 1, 1])
                        vw += s
                    vh += s

                x_best = torch.clamp(x + self.normalize(delta_init
                    ) * self.eps, 0., 1.)
                margin_min, loss_min = self.margin_and_loss(x_best, y)
                n_queries = torch.ones(x.shape[0]).to(self.device)
                s_init = int(math.sqrt(self.p_init * n_features / c))
                
                if (margin_min < 0.0).all():
                    return n_queries, x_best

                for i_iter in range(self.n_queries):
                    idx_to_fool = (margin_min > 0.0).nonzero().squeeze()

                    x_curr = self.check_shape(x[idx_to_fool])
                    x_best_curr = self.check_shape(x_best[idx_to_fool])
                    y_curr = y[idx_to_fool]
                    if len(y_curr.shape) == 0:
                        y_curr = y_curr.unsqueeze(0)
                    margin_min_curr = margin_min[idx_to_fool]
                    loss_min_curr = loss_min[idx_to_fool]

                    delta_curr = x_best_curr - x_curr
                    p = self.p_selection(i_iter)
                    s = max(int(round(math.sqrt(p * n_features / c))), 3)
                    if s % 2 == 0:
                        s += 1
                    s = min(s, min(h, w))

                    vh = self.random_int(0, h - s)
                    vw = self.random_int(0, w - s)
                    new_deltas_mask = torch.zeros_like(x_curr)
                    new_deltas_mask[:, :, vh:vh + s, vw:vw + s] = 1.0
                    norms_window_1 = (delta_curr[:, :, vh:vh + s, vw:vw + s
                        ] ** 2).sum(dim=(-2, -1), keepdim=True).sqrt()

                    vh2 = self.random_int(0, h - s)
                    vw2 = self.random_int(0, w - s)
                    new_deltas_mask_2 = torch.zeros_like(x_curr)
                    new_deltas_mask_2[:, :, vh2:vh2 + s, vw2:vw2 + s] = 1.

                    norms_image = self.lp_norm(x_best_curr - x_curr)
                    mask_image = torch.max(new_deltas_mask, new_deltas_mask_2)
                    norms_windows = ((delta_curr * mask_image) ** 2).sum(dim=(
                        -2, -1), keepdim=True).sqrt()

                    new_deltas = torch.ones([x_curr.shape[0], c, s, s]
                        ).to(self.device)
                    new_deltas *= (self.eta(s).view(1, 1, s, s) *
                        self.random_choice([x_curr.shape[0], c, 1, 1]))
                    old_deltas = delta_curr[:, :, vh:vh + s, vw:vw + s] / (
                        1e-12 + norms_window_1)
                    new_deltas += old_deltas
                    new_deltas = new_deltas / (1e-12 + (new_deltas ** 2).sum(
                        dim=(-2, -1), keepdim=True).sqrt()) * (torch.max(
                        (self.eps * torch.ones_like(new_deltas)) ** 2 -
                        norms_image ** 2, torch.zeros_like(new_deltas)) /
                        c + norms_windows ** 2).sqrt()
                    delta_curr[:, :, vh2:vh2 + s, vw2:vw2 + s] = 0.
                    delta_curr[:, :, vh:vh + s, vw:vw + s] = new_deltas + 0

                    x_new = torch.clamp(x_curr + self.normalize(delta_curr
                        ) * self.eps, 0. ,1.)
                    x_new = self.check_shape(x_new)
                    norms_image = self.lp_norm(x_new - x_curr)

                    margin, loss = self.margin_and_loss(x_new, y_curr)

                    # update loss if new loss is better
                    idx_improved = (loss < loss_min_curr).float()

                    loss_min[idx_to_fool] = idx_improved * loss + (
                        1. - idx_improved) * loss_min_curr

                    # update margin and x_best if new loss is better
                    # or misclassification
                    idx_miscl = (margin <= 0.).float()
                    idx_improved = torch.max(idx_improved, idx_miscl)

                    margin_min[idx_to_fool] = idx_improved * margin + (
                        1. - idx_improved) * margin_min_curr
                    idx_improved = idx_improved.reshape([-1,
                        *[1]*len(x.shape[:-1])])
                    x_best[idx_to_fool] = idx_improved * x_new + (
                        1. - idx_improved) * x_best_curr
                    n_queries[idx_to_fool] += 1.

                    ind_succ = (margin_min <= 0.).nonzero().squeeze()
                    if self.verbose and ind_succ.numel() != 0:
                        print('{}'.format(i_iter + 1),
                            '- success rate={}/{} ({:.2%})'.format(
                            ind_succ.numel(), n_ex_total, float(
                            ind_succ.numel()) / n_ex_total),
                            '- avg # queries={:.1f}'.format(
                            n_queries[ind_succ].mean().item()),
                            '- med # queries={:.1f}'.format(
                            n_queries[ind_succ].median().item()),
                            '- loss={:.3f}'.format(loss_min.mean()))

                    assert (x_new != x_new).sum() == 0
                    assert (x_best != x_best).sum() == 0
                    
                    if ind_succ.numel() == n_ex_total:
                        break

            elif self.norm == 'L1':
                delta_init = torch.zeros_like(x)
                s = h // 5
                sp_init = (h - s * 5) // 2
                vh = sp_init + 0
                for _ in range(h // s):
                    vw = sp_init + 0
                    for _ in range(w // s):
                        delta_init[:, :, vh:vh + s, vw:vw + s] += self.eta(
                            s).view(1, 1, s, s) * self.random_choice(
                            [x.shape[0], c, 1, 1])
                        vw += s
                    vh += s

                #x_best = torch.clamp(x + self.normalize(delta_init
                #    ) * self.eps, 0., 1.)
                r_best = L1_projection(x, delta_init, self.eps * (1. - 1e-6))
                x_best = x + delta_init + r_best
                margin_min, loss_min = self.margin_and_loss(x_best, y)
                n_queries = torch.ones(x.shape[0]).to(self.device)
                s_init = int(math.sqrt(self.p_init * n_features / c))
                
                if (margin_min < 0.0).all():
                    return n_queries, x_best

                for i_iter in range(self.n_queries):
                    idx_to_fool = (margin_min > 0.0).nonzero().squeeze()

                    x_curr = self.check_shape(x[idx_to_fool])
                    x_best_curr = self.check_shape(x_best[idx_to_fool])
                    y_curr = y[idx_to_fool]
                    if len(y_curr.shape) == 0:
                        y_curr = y_curr.unsqueeze(0)
                    margin_min_curr = margin_min[idx_to_fool]
                    loss_min_curr = loss_min[idx_to_fool]

                    delta_curr = x_best_curr - x_curr
                    p = self.p_selection(i_iter)
                    s = max(int(round(math.sqrt(p * n_features / c))), 3)
                    if s % 2 == 0:
                        s += 1
                        #pass
                    s = min(s, min(h, w))
                    
                    vh = self.random_int(0, h - s)
                    vw = self.random_int(0, w - s)
                    new_deltas_mask = torch.zeros_like(x_curr)
                    new_deltas_mask[:, :, vh:vh + s, vw:vw + s] = 1.0
                    norms_window_1 = delta_curr[:, :, vh:vh + s, vw:vw + s
                        ].abs().sum(dim=(-2, -1), keepdim=True)

                    vh2 = self.random_int(0, h - s)
                    vw2 = self.random_int(0, w - s)
                    new_deltas_mask_2 = torch.zeros_like(x_curr)
                    new_deltas_mask_2[:, :, vh2:vh2 + s, vw2:vw2 + s] = 1.

                    norms_image = self.lp_norm(x_best_curr - x_curr)
                    mask_image = torch.max(new_deltas_mask, new_deltas_mask_2)
                    norms_windows = (delta_curr * mask_image).abs().sum(dim=(
                        -2, -1), keepdim=True)

                    new_deltas = torch.ones([x_curr.shape[0], c, s, s]
                        ).to(self.device)
                    new_deltas *= (self.eta(s).view(1, 1, s, s) *
                        self.random_choice([x_curr.shape[0], c, 1, 1]))
                    old_deltas = delta_curr[:, :, vh:vh + s, vw:vw + s] / (
                        1e-12 + norms_window_1)
                    new_deltas += old_deltas
                    new_deltas = new_deltas / (1e-12 + new_deltas.abs().sum(
                        dim=(-2, -1), keepdim=True)) * (torch.max(
                        self.eps * torch.ones_like(norms_image) -
                        norms_image, torch.zeros_like(norms_image)) /
                        c + norms_windows) * c
                    delta_curr[:, :, vh2:vh2 + s, vw2:vw2 + s] = 0.
                    delta_curr[:, :, vh:vh + s, vw:vw + s] = new_deltas + 0

                    #
                    #norms_image_old = self.lp_norm(delta_curr)
                    r_curr = L1_projection(x_curr, delta_curr, self.eps * (1. - 1e-6))
                    x_new = x_curr + delta_curr + r_curr
                    x_new = self.check_shape(x_new)
                    norms_image = self.lp_norm(x_new - x_curr)

                    margin, loss = self.margin_and_loss(x_new, y_curr)

                    # update loss if new loss is better
                    idx_improved = (loss < loss_min_curr).float()

                    loss_min[idx_to_fool] = idx_improved * loss + (
                        1. - idx_improved) * loss_min_curr

                    # update margin and x_best if new loss is better
                    # or misclassification
                    idx_miscl = (margin <= 0.).float()
                    idx_improved = torch.max(idx_improved, idx_miscl)

                    margin_min[idx_to_fool] = idx_improved * margin + (
                        1. - idx_improved) * margin_min_curr
                    idx_improved = idx_improved.reshape([-1,
                        *[1]*len(x.shape[:-1])])
                    x_best[idx_to_fool] = idx_improved * x_new + (
                        1. - idx_improved) * x_best_curr
                    n_queries[idx_to_fool] += 1.

                    ind_succ = (margin_min <= 0.).nonzero().squeeze()
                    if self.verbose and ind_succ.numel() != 0:
                        print('{}'.format(i_iter + 1),
                            '- success rate={}/{} ({:.2%})'.format(
                            ind_succ.numel(), n_ex_total, float(
                            ind_succ.numel()) / n_ex_total),
                            '- avg # queries={:.1f}'.format(
                            n_queries[ind_succ].mean().item()),
                            '- med # queries={:.1f}'.format(
                            n_queries[ind_succ].median().item()),
                            '- loss={:.3f}'.format(loss_min.mean()),
                            '- max pert={:.3f}'.format(norms_image.max().item()),
                            #'- old pert={:.3f}'.format(norms_image_old.max().item())
                            )
                    
                    assert (x_new != x_new).sum() == 0
                    assert (x_best != x_best).sum() == 0
        
                    if ind_succ.numel() == n_ex_total:
                        break
        
        return n_queries, x_best

    def perturb(self, x, y=None):
        """
        :param x:           clean images
        :param y:           untargeted attack -> clean labels,
                            if None we use the predicted labels
                            targeted attack -> target labels, if None random classes,
                            different from the predicted ones, are sampled
        """

        self.init_hyperparam(x)

        adv = x.clone()
        #adv_all = x.clone()
        if y is None:
            if not self.targeted:
                with torch.no_grad():
                    output = self.predict(x)
                    y_pred = output.max(1)[1]
                    y = y_pred.detach().clone().long().to(self.device)
            else:
                with torch.no_grad():
                    output = self.predict(x)
                    n_classes = output.shape[-1]
                    y_pred = output.max(1)[1]
                    y = self.random_target_classes(y_pred, n_classes)
        else:
            y = y.detach().clone().long().to(self.device)

        if not self.targeted:
            acc = self.predict(x).max(1)[1] == y
        else:
            acc = self.predict(x).max(1)[1] != y

        startt = time.time()

        torch.random.manual_seed(self.seed)
        torch.cuda.random.manual_seed(self.seed)

        for counter in range(self.n_restarts):
            ind_to_fool = acc.nonzero().squeeze()
            if len(ind_to_fool.shape) == 0:
                ind_to_fool = ind_to_fool.unsqueeze(0)
            if ind_to_fool.numel() != 0:
                x_to_fool = x[ind_to_fool].clone()
                y_to_fool = y[ind_to_fool].clone()

                _, adv_curr = self.attack_single_run(x_to_fool, y_to_fool)

                output_curr = self.predict(adv_curr)
                if not self.targeted:
                    acc_curr = output_curr.max(1)[1] == y_to_fool
                else:
                    acc_curr = output_curr.max(1)[1] != y_to_fool
                ind_curr = (acc_curr == 0).nonzero().squeeze()

                acc[ind_to_fool[ind_curr]] = 0
                adv[ind_to_fool[ind_curr]] = adv_curr[ind_curr].clone()
                #adv_all[ind_to_fool] = adv_curr.clone()
                if self.verbose:
                    print('restart {} - robust accuracy: {:.2%}'.format(
                        counter, acc.float().mean()),
                        '- cum. time: {:.1f} s'.format(
                        time.time() - startt))

        if not self.return_all:
            return adv
        else:
            print('returning final points')
            return adv_all