File size: 7,011 Bytes
15fa80a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import numpy as np
# device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

class PGDAttacker():
    def __init__(self, radius, steps, step_size, random_start, norm_type, ascending=True):
        self.radius = radius  # attack radius
        self.steps = steps # how many step to conduct pgd
        self.step_size = step_size  # coefficient of PGD
        self.random_start = random_start
        self.norm_type = norm_type # which norm of your noise
        self.ascending = ascending # perform gradient ascending, i.e, to maximum the loss

    def output(self, x, model, tokens_lens, text_token):
        
        x = x + model.positional_embedding.type(model.dtype)

        x = x.permute(1, 0, 2)  # NLD -> LND
        x, weight = model.transformer(x)
        x = x.permute(1, 0, 2)  # LND -> NLD
        x = model.ln_final(x).type(model.dtype)
        x = x[torch.arange(x.shape[0]), text_token.argmax(dim=-1)] @ model.text_projection

        attention_weights_all = []
        for i in range(len(tokens_lens)):
            attention_weights = weight[-1][i][min(76, tokens_lens[i])][:1+min(75, max(tokens_lens))][1:][:-1]
            attention_weights_all.append(attention_weights)
        attention_weights = torch.stack(attention_weights_all, dim=0)

        return x, attention_weights

    def perturb(self, device, m_tokens_len, bs, criterion, x, y,a_indices,encoder, tokens_lens=None, model=None, text_token=None):
        if self.steps==0 or self.radius==0:
            return x.clone()

        adv_x = x.clone()

        if self.random_start:
            if self.norm_type == 'l-infty':
                adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius
            else:
                adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius / self.steps
            self._clip_(adv_x, x)

        ''' temporarily shutdown autograd of model to improve pgd efficiency '''
        # adv_x, attention_weights = self.output(adv_x, model, tokens_lens, text_token)

        # model.eval()
        encoder.eval()
        for pp in encoder.parameters():
            pp.requires_grad = False

        for step in range(self.steps):
            adv_x_o = adv_x.clone()
            adv_x.requires_grad_()
            _y = encoder(a_indices,adv_x)
            loss = criterion(y.to(device), _y, m_tokens_len, bs)
            grad = torch.autograd.grad(loss, [adv_x])[0]

            with torch.no_grad():
                if not self.ascending: grad.mul_(-1)

                if self.norm_type == 'l-infty':
                    adv_x.add_(torch.sign(grad), alpha=self.step_size)
                else:
                    if self.norm_type == 'l2':
                        grad_norm = (grad.reshape(grad.shape[0],-1)**2).sum(dim=1).sqrt()
                    elif self.norm_type == 'l1':
                        grad_norm = grad.reshape(grad.shape[0],-1).abs().sum(dim=1)
                    grad_norm = grad_norm.reshape( -1, *( [1] * (len(x.shape)-1) ) )
                    scaled_grad = grad / (grad_norm + 1e-10)
                    adv_x.add_(scaled_grad, alpha=self.step_size)

                self._clip_(adv_x, adv_x_o)

        ''' reopen autograd of model after pgd '''
        # decoder.trian()
        for pp in encoder.parameters():
            pp.requires_grad = True

        return adv_x  # , attention_weights
    
    def perturb_random(self, criterion, x, data, decoder,y,target_model,encoder=None):
        if self.steps==0 or self.radius==0:
            return x.clone()
        adv_x = x.clone()
        if self.norm_type == 'l-infty':
            adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius
        else:
            adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius / self.steps
        self._clip_(adv_x, x)
        return adv_x.data
    
    def perturb_iat(self, criterion, x, data, decoder,y,target_model,encoder=None):
        if self.steps==0 or self.radius==0:
            return x.clone()
        
        B = x.shape[0]
        L = x.shape[1]
        H = x.shape[2]
        nb_num = 8
        
        alpha = torch.rand(B,L,nb_num,1).to(device)
        
        A_1 = x.unsqueeze(2).expand(B,L,nb_num,H)
        A_2 = x.unsqueeze(1).expand(B,L,L,H)
        rand_idx = []
        for i in range(L):
            rand_idx.append(np.random.choice(L,nb_num,replace=False))
        rand_idx = np.array(rand_idx)
        rand_idx = torch.tensor(rand_idx).long().reshape(1,L,1,nb_num).expand(B,L,H,nb_num).to(device)
        # A_2 = A_2[:,np.arange(0,L), rand_idx,:]
        A_2 = torch.gather(A_2.reshape(B,L,H,L),-1,rand_idx).reshape(B,L,nb_num, H)
        A_e = A_1 - A_2
        # A_e
        # adv_x = (A_e * alpha).sum(dim=-1) + x.clone()
        
        adv_x = x.clone()

        if self.random_start:
            if self.norm_type == 'l-infty':
                adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius
            else:
                adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius / self.steps
        self._clip_(adv_x, x)

        # assert adv_x.shape[0] == 8

        ''' temporarily shutdown autograd of model to improve pgd efficiency '''
        # model.eval()
        decoder.eval()
        for pp in decoder.parameters():
            pp.requires_grad = False
            
        adv_x = x.clone()
        
        alpha.requires_grad_()

        for step in range(self.steps):
            alpha.requires_grad_()
            dot_Ae_alpha = (A_e * alpha).sum(dim=-2)
            # print("dot_Ae_alpha:", dot_Ae_alpha.shape)
            
            adv_x.add_(torch.sign(dot_Ae_alpha), alpha=self.step_size)
            
            self._clip_(adv_x, x)
            
            if encoder is None:
                adv_x_input = adv_x.squeeze(-1)
            else:
                adv_x_input = adv_x
            
            _y = target_model(adv_x_input, data,decoder,encoder)
            loss = criterion(y.to(device), _y)
            grad = torch.autograd.grad(loss, [alpha],retain_graph=True)[0]
            # with torch.no_grad():
            with torch.no_grad():
                if not self.ascending: grad.mul_(-1)
                assert self.norm_type == 'l-infty'
                alpha = alpha.detach()+ grad * 0.01

        ''' reopen autograd of model after pgd '''
        # decoder.trian()
        for pp in decoder.parameters():
            pp.requires_grad = True

        return adv_x.data

    def _clip_(self, adv_x, x):
        adv_x -= x
        if self.norm_type == 'l-infty':
            adv_x.clamp_(-self.radius, self.radius)
        else:
            if self.norm_type == 'l2':
                norm = (adv_x.reshape(adv_x.shape[0],-1)**2).sum(dim=1).sqrt()
            elif self.norm_type == 'l1':
                norm = adv_x.reshape(adv_x.shape[0],-1).abs().sum(dim=1)
            norm = norm.reshape( -1, *( [1] * (len(x.shape)-1) ) )
            adv_x /= (norm + 1e-10)
            adv_x *= norm.clamp(max=self.radius)
        adv_x += x
        adv_x.clamp_(0, 1)