File size: 6,332 Bytes
5ca4e86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from tqdm import tqdm
import random
from llava_utils import prompt_wrapper
from torchvision.utils import save_image

import copy
import numpy as np
import matplotlib.pyplot as plt
from matplotlib.widgets import MultiCursor
import seaborn as sns



def denormalize(images):
    mean = torch.tensor([0.48145466, 0.4578275, 0.40821073]).cuda()
    std = torch.tensor([0.26862954, 0.26130258, 0.27577711]).cuda()
    new_images = (images - mean[None, :, None, None])/ std[None, :, None, None]
    return new_images

def normalize(images):
    mean = torch.tensor([0.48145466, 0.4578275, 0.40821073]).cuda()
    std = torch.tensor([0.26862954, 0.26130258, 0.27577711]).cuda()
    new_images = (images * std[None, :, None, None])+ mean[None, :, None, None]
    return new_images


class Defender:

    def __init__(self, args, model, tokenizer, pos_targets, neg_targets, device='cuda:0', is_rtp=False, image_processor=None):

        self.args = args
        self.model = model
        self.tokenizer= tokenizer
        self.device = device
        self.is_rtp = is_rtp

        self.pos_targets = pos_targets
        self.neg_targets = neg_targets
        self.num_targets = len(pos_targets)

        self.loss_buffer = []

        # freeze and set to eval model:
        self.model.eval()
        self.model.requires_grad_(False)

        self.image_processor = image_processor

    def defense_constrained(self, text_prompt, img, batch_size = 4, num_iter=2000, alpha=1/255, epsilon = 128/255 ):

        print('>>> batch_size:', batch_size)

        adv_noise = torch.rand_like(img[0]).cuda() * 2 * epsilon - epsilon
        x = normalize(img).clone().cuda()
        adv_noise.data = (adv_noise.data + x.data).clamp(0, 1) - x.data
        adv_noise = adv_noise.cuda()

        neg_prompt = prompt_wrapper.Prompt(self.model, self.tokenizer, text_prompts=text_prompt, device=self.device)
        adv_noise.requires_grad_(True)
        adv_noise.retain_grad()

        for t in tqdm(range(num_iter + 1)):
        
            neg_batch_targets = random.sample(self.neg_targets, batch_size)

            target_loss = 0
            x_adv = x + adv_noise
            x_adv = denormalize(x_adv)
            target_loss -= self.attack_loss(neg_prompt,x_adv,neg_batch_targets)
            target_loss.backward()

            adv_noise.data = (adv_noise.data - alpha * adv_noise.grad.detach().sign()).clamp(-epsilon, epsilon)
            adv_noise.data = (adv_noise.data + x.data).clamp(0, 1) - x.data
            adv_noise.grad.zero_()
            self.model.zero_grad()
            self.loss_buffer.append(target_loss.item())

            print("target_loss: %f" % (
                target_loss.item())
                  )

            if t % 20 == 0:
                self.plot_loss()

            if t % 100 == 0:
                safety_patch = adv_noise.detach().cpu().squeeze(0)
                #if you want to save the image safety patch
#                save_image(safety_patch, '%s/safety_patch_temp_%d.bmp' % (self.args.save_dir, t))

        return safety_patch

    def plot_loss(self):

        sns.set_theme()
        num_iters = len(self.loss_buffer)

        x_ticks = list(range(0, num_iters))

        # Plot and label the training and validation loss values
        plt.plot(x_ticks, self.loss_buffer, label='Target Loss')

        # Add in a title and axes labels
        plt.title('Loss Plot')
        plt.xlabel('Iters')
        plt.ylabel('Loss')

        # Display the plot
        plt.legend(loc='best')
        plt.savefig('%s/loss_curve.png' % (self.args.save_dir))
        plt.clf()

        torch.save(self.loss_buffer, '%s/loss' % (self.args.save_dir))

    def attack_loss(self, prompts, images, targets):

        context_length = prompts.context_length
        context_input_ids = prompts.input_ids
        batch_size = len(targets)

        images = images.repeat(batch_size, 1, 1, 1)
 
        if len(context_input_ids) == 1:
            context_length = context_length * batch_size
            context_input_ids = context_input_ids * batch_size

        assert len(context_input_ids) == len(targets), f"Unmathced batch size of prompts and targets {len(context_input_ids)} != {len(targets)}"


        tokens = [ torch.as_tensor([item[1:]]).cuda() for item in self.tokenizer(targets).input_ids] # get rid of the default <bos> in targets tokenization.

        seq_tokens_length = []
        labels = []
        input_ids = []

        for i, item in enumerate(tokens):

            L = item.shape[1] + context_length[i]
            seq_tokens_length.append(L)

            context_mask = torch.full([1, context_length[i]], -100,
                                      dtype=tokens[0].dtype,
                                      device=tokens[0].device)
            labels.append( torch.cat( [context_mask, item], dim=1 ) )
            input_ids.append( torch.cat( [context_input_ids[i], item], dim=1 ) )

        # padding token
        pad = torch.full([1, 1], 0,
                         dtype=tokens[0].dtype,
                         device=tokens[0].device).cuda() # it does not matter ... Anyway will be masked out from attention...


        max_length = max(seq_tokens_length)
        attention_mask = []

        for i in range(batch_size):

            # padding to align the length
            num_to_pad = max_length - seq_tokens_length[i]

            padding_mask = (
                torch.full([1, num_to_pad], -100,
                       dtype=torch.long,
                       device=self.device)
            )
            labels[i] = torch.cat( [labels[i], padding_mask], dim=1 )

            input_ids[i] = torch.cat( [input_ids[i],
                                       pad.repeat(1, num_to_pad)], dim=1 )
            attention_mask.append( torch.LongTensor( [ [1]* (seq_tokens_length[i]) + [0]*num_to_pad ] ) )

        labels = torch.cat( labels, dim=0 ).cuda()
        input_ids = torch.cat( input_ids, dim=0 ).cuda()
        attention_mask = torch.cat(attention_mask, dim=0).cuda()
        
        outputs = self.model(
                input_ids=input_ids,
                attention_mask=attention_mask,
                return_dict=True,
                labels=labels,
                images=images.half(),
            )
        loss = outputs.loss
        return loss