File size: 14,703 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
# Copyright (c) 2019-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 time

import torch

from autoattack.fab_projections import projection_linf, projection_l2,\
    projection_l1

DEFAULT_EPS_DICT_BY_NORM = {'Linf': .3, 'L2': 1., 'L1': 5.0}


class FABAttack():
    """
    Fast Adaptive Boundary Attack (Linf, L2, L1)
    https://arxiv.org/abs/1907.02044
    
    :param norm:          Lp-norm to minimize ('Linf', 'L2', 'L1' supported)
    :param n_restarts:    number of random restarts
    :param n_iter:        number of iterations
    :param eps:           epsilon for the random restarts
    :param alpha_max:     alpha_max
    :param eta:           overshooting
    :param beta:          backward step
    """

    def __init__(
            self,
            norm='Linf',
            n_restarts=1,
            n_iter=100,
            eps=None,
            alpha_max=0.1,
            eta=1.05,
            beta=0.9,
            loss_fn=None,
            verbose=False,
            seed=0,
            targeted=False,
            device=None,
            n_target_classes=9):
        """ FAB-attack implementation in pytorch """

        self.norm = norm
        self.n_restarts = n_restarts
        self.n_iter = n_iter
        self.eps = eps if eps is not None else DEFAULT_EPS_DICT_BY_NORM[norm]
        self.alpha_max = alpha_max
        self.eta = eta
        self.beta = beta
        self.targeted = targeted
        self.verbose = verbose
        self.seed = seed
        self.target_class = None
        self.device = device
        self.n_target_classes = n_target_classes

    def check_shape(self, x):
        return x if len(x.shape) > 0 else x.unsqueeze(0)

    def _predict_fn(self, x):
        raise NotImplementedError("Virtual function.")

    def _get_predicted_label(self, x):
        raise NotImplementedError("Virtual function.")

    def get_diff_logits_grads_batch(self, imgs, la):
        raise NotImplementedError("Virtual function.")

    def get_diff_logits_grads_batch_targeted(self, imgs, la, la_target):
       raise NotImplementedError("Virtual function.")

    def attack_single_run(self, x, y=None, use_rand_start=False, is_targeted=False):
        """
        :param x:             clean images
        :param y:             clean labels, if None we use the predicted labels
        :param is_targeted    True if we ise targeted version. Targeted class is assigned by `self.target_class`
        """

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

        x = x.detach().clone().float().to(self.device)
        #assert next(self.predict.parameters()).device == x.device

        y_pred = self._get_predicted_label(x)
        if y is None:
            y = y_pred.detach().clone().long().to(self.device)
        else:
            y = y.detach().clone().long().to(self.device)
        pred = y_pred == y
        corr_classified = pred.float().sum()
        if self.verbose:
            print('Clean accuracy: {:.2%}'.format(pred.float().mean()))
        if pred.sum() == 0:
            return x
        pred = self.check_shape(pred.nonzero().squeeze())

        if is_targeted:
            output = self._predict_fn(x)
            la_target = output.sort(dim=-1)[1][:, -self.target_class]
            la_target2 = la_target[pred].detach().clone()

        startt = time.time()
        # runs the attack only on correctly classified points
        im2 = x[pred].detach().clone()
        la2 = y[pred].detach().clone()
        if len(im2.shape) == self.ndims:
            im2 = im2.unsqueeze(0)
        bs = im2.shape[0]
        u1 = torch.arange(bs)
        adv = im2.clone()
        adv_c = x.clone()
        res2 = 1e10 * torch.ones([bs]).to(self.device)
        x1 = im2.clone()
        x0 = im2.clone().reshape([bs, -1])

        if use_rand_start:
            if self.norm == 'Linf':
                t = 2 * torch.rand(x1.shape).to(self.device) - 1
                x1 = im2 + (torch.min(res2,
                                        self.eps * torch.ones(res2.shape)
                                        .to(self.device)
                                        ).reshape([-1, *[1]*self.ndims])
                            ) * t / (t.reshape([t.shape[0], -1]).abs()
                                        .max(dim=1, keepdim=True)[0]
                                        .reshape([-1, *[1]*self.ndims])) * .5
            elif self.norm == 'L2':
                t = torch.randn(x1.shape).to(self.device)
                x1 = im2 + (torch.min(res2,
                                        self.eps * torch.ones(res2.shape)
                                        .to(self.device)
                                        ).reshape([-1, *[1]*self.ndims])
                            ) * t / ((t ** 2)
                                        .view(t.shape[0], -1)
                                        .sum(dim=-1)
                                        .sqrt()
                                        .view(t.shape[0], *[1]*self.ndims)) * .5
            elif self.norm == 'L1':
                t = torch.randn(x1.shape).to(self.device)
                x1 = im2 + (torch.min(res2,
                                        self.eps * torch.ones(res2.shape)
                                        .to(self.device)
                                        ).reshape([-1, *[1]*self.ndims])
                            ) * t / (t.abs().view(t.shape[0], -1)
                                        .sum(dim=-1)
                                        .view(t.shape[0], *[1]*self.ndims)) / 2

            x1 = x1.clamp(0.0, 1.0)

        counter_iter = 0
        while counter_iter < self.n_iter:
            with torch.no_grad():
                if is_targeted:
                    df, dg = self.get_diff_logits_grads_batch_targeted(x1, la2, la_target2)
                else:
                    df, dg = self.get_diff_logits_grads_batch(x1, la2)
                if self.norm == 'Linf':
                    dist1 = df.abs() / (1e-12 +
                                        dg.abs()
                                        .reshape(dg.shape[0], dg.shape[1], -1)
                                        .sum(dim=-1))
                elif self.norm == 'L2':
                    dist1 = df.abs() / (1e-12 + (dg ** 2)
                                        .reshape(dg.shape[0], dg.shape[1], -1)
                                        .sum(dim=-1).sqrt())
                elif self.norm == 'L1':
                    dist1 = df.abs() / (1e-12 + dg.abs().reshape(
                        [df.shape[0], df.shape[1], -1]).max(dim=2)[0])
                else:
                    raise ValueError('norm not supported')
                ind = dist1.min(dim=1)[1]
                dg2 = dg[u1, ind]
                b = (- df[u1, ind] + (dg2 * x1).reshape(x1.shape[0], -1)
                                        .sum(dim=-1))
                w = dg2.reshape([bs, -1])

                if self.norm == 'Linf':
                    d3 = projection_linf(
                        torch.cat((x1.reshape([bs, -1]), x0), 0),
                        torch.cat((w, w), 0),
                        torch.cat((b, b), 0))
                elif self.norm == 'L2':
                    d3 = projection_l2(
                        torch.cat((x1.reshape([bs, -1]), x0), 0),
                        torch.cat((w, w), 0),
                        torch.cat((b, b), 0))
                elif self.norm == 'L1':
                    d3 = projection_l1(
                        torch.cat((x1.reshape([bs, -1]), x0), 0),
                        torch.cat((w, w), 0),
                        torch.cat((b, b), 0))
                d1 = torch.reshape(d3[:bs], x1.shape)
                d2 = torch.reshape(d3[-bs:], x1.shape)
                if self.norm == 'Linf':
                    a0 = d3.abs().max(dim=1, keepdim=True)[0]\
                        .view(-1, *[1]*self.ndims)
                elif self.norm == 'L2':
                    a0 = (d3 ** 2).sum(dim=1, keepdim=True).sqrt()\
                        .view(-1, *[1]*self.ndims)
                elif self.norm == 'L1':
                    a0 = d3.abs().sum(dim=1, keepdim=True)\
                        .view(-1, *[1]*self.ndims)
                a0 = torch.max(a0, 1e-8 * torch.ones(
                    a0.shape).to(self.device))
                a1 = a0[:bs]
                a2 = a0[-bs:]
                alpha = torch.min(torch.max(a1 / (a1 + a2),
                                            torch.zeros(a1.shape)
                                            .to(self.device)),
                                    self.alpha_max * torch.ones(a1.shape)
                                    .to(self.device))
                x1 = ((x1 + self.eta * d1) * (1 - alpha) +
                        (im2 + d2 * self.eta) * alpha).clamp(0.0, 1.0)

                is_adv = self._get_predicted_label(x1) != la2

                if is_adv.sum() > 0:
                    ind_adv = is_adv.nonzero().squeeze()
                    ind_adv = self.check_shape(ind_adv)
                    if self.norm == 'Linf':
                        t = (x1[ind_adv] - im2[ind_adv]).reshape(
                            [ind_adv.shape[0], -1]).abs().max(dim=1)[0]
                    elif self.norm == 'L2':
                        t = ((x1[ind_adv] - im2[ind_adv]) ** 2)\
                            .reshape(ind_adv.shape[0], -1).sum(dim=-1).sqrt()
                    elif self.norm == 'L1':
                        t = (x1[ind_adv] - im2[ind_adv])\
                            .abs().reshape(ind_adv.shape[0], -1).sum(dim=-1)
                    adv[ind_adv] = x1[ind_adv] * (t < res2[ind_adv]).\
                        float().reshape([-1, *[1]*self.ndims]) + adv[ind_adv]\
                        * (t >= res2[ind_adv]).float().reshape(
                        [-1, *[1]*self.ndims])
                    res2[ind_adv] = t * (t < res2[ind_adv]).float()\
                        + res2[ind_adv] * (t >= res2[ind_adv]).float()
                    x1[ind_adv] = im2[ind_adv] + (
                        x1[ind_adv] - im2[ind_adv]) * self.beta

                counter_iter += 1

        ind_succ = res2 < 1e10
        if self.verbose:
            print('success rate: {:.0f}/{:.0f}'
                  .format(ind_succ.float().sum(), corr_classified) +
                  ' (on correctly classified points) in {:.1f} s'
                  .format(time.time() - startt))

        ind_succ = self.check_shape(ind_succ.nonzero().squeeze())
        adv_c[pred[ind_succ]] = adv[ind_succ].clone()

        return adv_c

    def perturb(self, x, y):
        if self.device is None:
            self.device = x.device
        adv = x.clone()
        with torch.no_grad():
            acc = self._predict_fn(x).max(1)[1] == y

            startt = time.time()

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

            if not self.targeted:
                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, y_to_fool = x[ind_to_fool].clone(), y[ind_to_fool].clone()
                        adv_curr = self.attack_single_run(x_to_fool, y_to_fool, use_rand_start=(counter > 0), is_targeted=False)

                        acc_curr = self._predict_fn(adv_curr).max(1)[1] == y_to_fool
                        if self.norm == 'Linf':
                            res = (x_to_fool - adv_curr).abs().reshape(x_to_fool.shape[0], -1).max(1)[0]
                        elif self.norm == 'L2':
                            res = ((x_to_fool - adv_curr) ** 2).reshape(x_to_fool.shape[0], -1).sum(dim=-1).sqrt()
                        elif self.norm == 'L1':
                            res = (x_to_fool - adv_curr).abs().reshape(x_to_fool.shape[0], -1).sum(-1)
                        acc_curr = torch.max(acc_curr, res > self.eps)

                        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()

                        if self.verbose:
                            print('restart {} - robust accuracy: {:.2%} at eps = {:.5f} - cum. time: {:.1f} s'.format(
                                counter, acc.float().mean(), self.eps, time.time() - startt))

            else:
                for target_class in range(2, self.n_target_classes + 2):
                    self.target_class = target_class
                    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, y_to_fool = x[ind_to_fool].clone(), y[ind_to_fool].clone()
                            adv_curr = self.attack_single_run(x_to_fool, y_to_fool, use_rand_start=(counter > 0), is_targeted=True)

                            acc_curr = self._predict_fn(adv_curr).max(1)[1] == y_to_fool
                            if self.norm == 'Linf':
                                res = (x_to_fool - adv_curr).abs().reshape(x_to_fool.shape[0], -1).max(1)[0]
                            elif self.norm == 'L2':
                                res = ((x_to_fool - adv_curr) ** 2).reshape(x_to_fool.shape[0], -1).sum(dim=-1).sqrt()
                            elif self.norm == 'L1':
                                res = (x_to_fool - adv_curr).abs().reshape(x_to_fool.shape[0], -1).sum(-1)
                            acc_curr = torch.max(acc_curr, res > self.eps)

                            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()

                            if self.verbose:
                                print('restart {} - target_class {} - robust accuracy: {:.2%} at eps = {:.5f} - cum. time: {:.1f} s'.format(
                                    counter, self.target_class, acc.float().mean(), self.eps, time.time() - startt))

        return adv