-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathAIP.py
210 lines (154 loc) · 8 KB
/
AIP.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
import os
import time
import argparse
import numpy as np
import random
from tqdm import tqdm
import torch
import torch.nn as nn
import torch.optim as optim
import torch.utils.data as data
import torch.backends.cudnn as cudnn
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data import Dataset, DataLoader
from torchvision import models
import torchvision.transforms as transforms
import torch.nn.functional as F
import matplotlib.pyplot as plt
from PIL import Image
from io import StringIO, BytesIO
class trainset(Dataset):
def __init__(self, train_ls):
self.target = train_ls
def __getitem__(self, index):
target = self.target[index]
return target
def __len__(self):
return len(self.target)
orginal_transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize((0.6949, 0.6748, 0.6676), (0.3102, 0.3220, 0.3252))
])
tensor_transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),])
orginal_transform_alex = transforms.Compose([
transforms.Resize((256, 256)),
transforms.CenterCrop((224, 224)),
transforms.ToTensor(),])
partial_transform = transforms.Compose([
transforms.ToTensor(),])
def INSA_DVBPR(save_root, org_img_num, User_content_embedding, epsilon, Item, device, model, norm):
image_o = tensor_transform(Image.open(BytesIO(Item[org_img_num][b'imgs']))).to(device).unsqueeze(0)
delta = torch.rand([1, 3, 224, 224], requires_grad=True, device=device)
optimizer = torch.optim.Adam([delta], lr=1e-3)
train_ls = [User_content_embedding[u_idx] for u_idx in range(len(User_content_embedding))]
train_data = trainset(train_ls)
for epoch in range(10):
for data in DataLoader(train_data, batch_size = 512, shuffle = False, num_workers = 4):
loss = -1 * torch.sum(model(norm(image_o + delta)) * data.to(device), axis = 1).exp().log().mean()
# print(loss.item())
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
delta.data = torch.clamp(delta.data, -epsilon/255, epsilon/255)
delta.data = (torch.clamp(image_o + delta.data, 0 + 1e-6, 1 - 1e-6) - image_o)
X_new = image_o + delta.data
x_np = transforms.ToPILImage()((torch.round(X_new[0]*255)/255).detach().cpu())
x_np.save(save_root + str(org_img_num) +'.png')
def EXPA_DVBPR(org_img_path, target_number, adv_images_root, epsilon, Item, device, model, norm):
image_t = tensor_transform(Image.open(BytesIO(Item[org_img_path][b'imgs']))).to(device)
target_feature = model(orginal_transform(Image.open(BytesIO(Item[target_number][b'imgs']))).unsqueeze(0).to(device))
v = torch.zeros_like(image_t, requires_grad=True, device = device)
loss_fn = torch.nn.MSELoss(reduction='sum')
learning_rate = 1e-3
optimizer = torch.optim.Adam([v], lr=learning_rate)
for t in tqdm(range(5000)):
y_pred = model(norm(image_t + v))
loss = loss_fn(y_pred, target_feature)
# print(loss.item())
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
v.data = torch.clamp(v.data, -epsilon/255, epsilon/255)
v.data = (torch.clamp(image_t + v.data, 0 + 1e-6, 1 - 1e-6) - image_t)
X_new = image_t + v.data
x_np = transforms.ToPILImage()((torch.round(X_new*255)/255).detach().cpu())
x_np.save(adv_images_root + str(org_img_path) +'.png')
def EXPA_DVBPR_new(org_img_path, target_number, adv_images_root, epsilon, Item, device, model, norm, alpha=1/255):
image_t = tensor_transform(Image.open(BytesIO(Item[org_img_path][b'imgs']))).to(device)
target_feature = model(orginal_transform(Image.open(BytesIO(Item[target_number][b'imgs']))).unsqueeze(0).to(device))
v = torch.zeros_like(image_t, requires_grad=True, device = device)
loss_fn = torch.nn.MSELoss(reduction='mean')
# learning_rate = 1e-2
# optimizer = torch.optim.Adam([v], lr=learning_rate)
for t in tqdm(range(1000)):
y_pred = model(norm(image_t + v))
loss = loss_fn(y_pred, target_feature)
# print(loss.item())
# optimizer.zero_grad()
loss.backward(retain_graph=True)
# optimizer.step()
adv_images = image_t - alpha*torch.sign(v.grad)
v.data = torch.clamp(adv_images - image_t, -epsilon/255, epsilon/255)
# v.data = torch.clamp(v.data, -epsilon/255, epsilon/255)
v.data = (torch.clamp(image_t + v.data, 0 + 1e-6, 1 - 1e-6) - image_t)
X_new = image_t + v.data
x_np = transforms.ToPILImage()((torch.round(X_new*255)/255).detach().cpu())
x_np.save(adv_images_root + str(org_img_path) +'.png')
def INSA_VBPR(save_root, org_img_num, usernum, epsilon, Item, device, VBPRmodel, feature_model, norm):
delta = torch.rand([1, 3, 224, 224], requires_grad=True, device=device)
optimizer = torch.optim.Adam([delta], lr=1e-4)
train_ls = [list((u_idx, org_img_num)) for u_idx in range(usernum)]
train_data = trainset(train_ls)
for epoch in range(5):
for data in DataLoader(train_data, batch_size = 256, shuffle = False, num_workers = 4):
ui, xj = data
image_o = orginal_transform_alex(Image.open(BytesIO(Item[int(xj[0].numpy())][b'imgs']))).unsqueeze(0).to(device)
loss = -1 *(VBPRmodel(ui.to(device), xj.to(device), feature_model(norm((image_o + delta))))).exp().sum()
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
delta.data = torch.clamp(delta.data, -epsilon/255, epsilon/255)
delta.data = (torch.clamp(image_o + delta.data, 0 + 1e-6, 1 - 1e-6) - image_o)
X_new = image_o + delta.data
x_np = transforms.ToPILImage()((torch.round(X_new[0]*255)/255).detach().cpu())
x_np.save(save_root + str(org_img_num) +'.png')
def Alex_EXPA(save_root, org_img_num, target_item, usernum, epsilon, Item, device, feature_model, norm):
delta = torch.rand([1, 3, 224, 224], requires_grad=True, device=device)
optimizer = torch.optim.Adam([delta], lr=1e-2)
for epoch in range(5000):
image_o = orginal_transform_alex(Image.open(BytesIO(Item[org_img_num][b'imgs']))).unsqueeze(0).to(device)
loss = torch.norm(feature_model(norm((image_o + delta))) - feature_model(norm(orginal_transform_alex(Image.open(BytesIO(Item[target_item][b'imgs']))).to(device))))
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
delta.data = torch.clamp(delta.data, -epsilon/255, epsilon/255)
delta.data = (torch.clamp(image_o + delta.data, 0 + 1e-6, 1 - 1e-6) - image_o)
X_new = image_o + delta.data
x_np = transforms.ToPILImage()((torch.round(X_new[0]*255)/255).detach().cpu())
x_np.save(save_root + str(org_img_num) +'.png')
def INSA_AlexRank(save_root, org_img_num, alexnet_feature, usernum, epsilon, Item, device, model, norm, user_train):
item_dict = {}
for u in tqdm(range(usernum)):
for j in user_train[u]:
item_id = j[b'productid']
if u not in item_dict:
item_dict[u] = [item_id]
else:
item_dict[u].append(item_id)
delta = torch.rand([1, 3, 224, 224], requires_grad=True, device=device)
image_o = orginal_transform_alex(Image.open(BytesIO(Item[org_img_num][b'imgs']))).unsqueeze(0).to(device)
optimizer = torch.optim.Adam([delta], lr=1e-3)
for epoch in range(1):
for i in tqdm(range(usernum)):
loss = torch.norm(model(norm(image_o + delta)) - torch.tensor(alexnet_feature[item_dict[i]]).to(device), dim = 1).mean()
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
delta.data = torch.clamp(delta.data, -epsilon/255, epsilon/255)
delta.data = (torch.clamp(image_o + delta.data, 0 + 1e-6, 1 - 1e-6) - image_o)
X_new = image_o + delta.data
x_np = transforms.ToPILImage()((torch.round(X_new[0]*255)/255).detach().cpu())
x_np.save(save_root + str(org_img_num) +'.png')