-
Notifications
You must be signed in to change notification settings - Fork 0
/
Extractor.py
220 lines (188 loc) · 9.85 KB
/
Extractor.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
211
212
213
214
215
216
217
218
219
220
import os
import math
from diffusers import StableDiffusionImg2ImgPipeline
from diffusers import UNet2DConditionModel
from transformers import CLIPTextModel
import torch
import torch.nn as nn
import numpy as np
from torchvision import transforms
from PIL import Image
try:
from diffusers.utils import randn_tensor
except ImportError:
from diffusers.utils.torch_utils import randn_tensor
import random
from tqdm import tqdm
from sdm_utils import tokenize_prompt, attack_model, preprocess, identity_loss
from PGD import L2PGDAttack as L2Attack
mask_tensor = None
global_steps = 0
transform = transforms.Compose([
transforms.Resize(512, interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop(512)
])
class attack_noise_model(nn.Module):
def __init__(self, net1, net2, ori_latents=None):
super().__init__()
self.net1 = net1
self.net2 = net2
self.fn = nn.MSELoss(reduction="mean")
self.ori_latents = ori_latents
def forward(self, src_latents, current_t = None):
latents =self.ori_latents * mask_tensor + src_latents * (1-mask_tensor)
if current_t is None:
t = random.randint(1,999)
else:
t = current_t
self.net1.latents = latents
loss1 = self.net1.forward_test(token_possibility=None, noise=None, t=t, source=True, latents=latents)
self.net2.latents = latents
loss2 =self.net2.forward_test(token_possibility=None, noise=None, t=t, source=True, latents=latents)
loss = loss1 - loss2
loss = loss
global global_steps
global_steps += 1
if global_steps%100 == 0:
print(t, torch.norm(latents).item(), torch.max(latents).item(), torch.min(latents).item(), loss.item())
return loss
def main(
training_mode = 'db_prior',
num_imgs = 10,
per_step = 50,
src_dataset_mode = 0,
version = "1.4",
epsilon = 70.0,
kernel_size = 16,
steps = 1000,
adaptive=False,
alpha= 2.0
):
global mask_tensor
checkpoint = num_imgs * per_step
if src_dataset_mode == 0:
data_src = "wikiart_vangogh"
data_type = "style"
elif src_dataset_mode == 1:
data_src = "object_dog"
data_type = "dog"
assert version == "1.4"
training_parameters = f"{num_imgs}_{per_step}_v{version}"
src_style_dirs = f"{training_mode}/{data_src}/{training_parameters}"
low_pass_rate = 2
low_pass = True
clip_min, clip_max = -100.0, 100.0
if training_mode =='db_prior':
given_prompt = f"sks {data_type}"
else:
given_prompt = f"a figure"
atk_type = f"{low_pass}_{low_pass_rate}{kernel_size}_{epsilon}{adaptive}_{alpha}_{steps}"
dir_name = f"Recovered_Samples"
for style_name in os.listdir(src_style_dirs):
if style_name == '.ipynb_checkpoints':
continue
model_id = os.path.join(src_style_dirs, style_name)
type_list = ['membership', 'hold_out']
saving_dirname = f"{dir_name}/{src_style_dirs}/{atk_type}/{style_name}/{type_list[0]}"
if os.path.exists(saving_dirname):
print(f"Existing {saving_dirname}. Skipping")
continue
unet = UNet2DConditionModel.from_pretrained(f"{model_id}/checkpoint-{checkpoint}/unet", torch_dtype=torch.float16).to("cuda")
text_encoder = CLIPTextModel.from_pretrained(f"{model_id}/checkpoint-{checkpoint}/text_encoder", torch_dtype=torch.float16).to("cuda")
pip = StableDiffusionImg2ImgPipeline.from_pretrained('CompVis/stable-diffusion-v1-4', safety_checker=None, unet=unet, text_encoder=text_encoder, torch_dtype=torch.float16).to("cuda")
ori_pip = StableDiffusionImg2ImgPipeline.from_pretrained('CompVis/stable-diffusion-v1-4', torch_dtype=torch.float16).to("cuda")
device = pip._execution_device
try:
prompt_embeds_reference = pip._encode_prompt(
given_prompt,
device,
1,
do_classifier_free_guidance=False,
negative_prompt=None,
)
except TypeError:
prompt_embeds_reference = pip.encode_prompt(
given_prompt,
device,
1,
do_classifier_free_guidance=False,
negative_prompt=None,
)[0]
print(prompt_embeds_reference)
for class_id in tqdm(range(num_imgs)):
for type_name in type_list:
data_dirname = f"./datasets/{data_src}/{num_imgs}/{style_name}/{type_name}"
if not os.path.exists(data_dirname):
os.makedirs(data_dirname)
img_path = os.path.join(data_dirname, os.listdir(data_dirname)[class_id])
saving_dirname = f"{dir_name}/{src_style_dirs}/{atk_type}/{style_name}/{type_name}"
if not os.path.exists(saving_dirname):
os.makedirs(saving_dirname)
init_image = Image.open(img_path).convert("RGB")
with torch.no_grad():
init_image.save(f"{saving_dirname}/{os.listdir(data_dirname)[class_id]}_src_img.png")
init_image = transform(init_image)
ori_init_image = init_image
# lowpass_image = init_image.filter(ImageFilter.GaussianBlur(radius=low_pass_rate))
if low_pass:
with torch.no_grad():
ori_latents = pip.vae.encode(preprocess(init_image).half().cuda()).latent_dist.sample()* pip.vae.config.scaling_factor
current_latents = ori_latents.clone()
mask_tensor = torch.zeros([1, 4, 64, 64]).to(torch.float16).cuda()
per_block = 64//kernel_size
passing_kernel = per_block//low_pass_rate
for i in range(kernel_size):
for j in range(kernel_size):
mask_tensor[:, :, i*per_block:i*per_block+per_block - passing_kernel:, j*per_block:j*per_block+per_block - passing_kernel] += 1
mask_tensor[:, :, i*per_block+passing_kernel:i*per_block+per_block:, j*per_block+passing_kernel:j*per_block+per_block] += 1
current_latents *= mask_tensor
blurred_latents = current_latents
lowpass_img = pip.decode_latents(blurred_latents)
lowpass_image = pip.numpy_to_pil(lowpass_img)[0]
init_image = lowpass_image
lowpass_image.save(f"{saving_dirname}/low_pass_{os.listdir(data_dirname)[class_id]}.png")
latents = pip.vae.encode(preprocess(init_image).half().cuda()).latent_dist.sample()* pip.vae.config.scaling_factor
print(f"ori latent dist {torch.norm(latents - ori_latents)}, max and min ori is {torch.max(ori_latents), torch.min(ori_latents)}")
if adaptive:
epsilon = torch.norm(latents - ori_latents).item()
_, _ , input_ids, ori_embedding = tokenize_prompt(pip, "a figure", pip._execution_device, randn_init=True)
net = attack_model(pip, latents=latents, input_ids=input_ids, mode=1, prompt_embeds_reference=prompt_embeds_reference, token_possibility=None, possibility_embedding=None)
given_prompt2 = given_prompt
try:
prompt_embeds_reference_2 = ori_pip._encode_prompt(
given_prompt2,
device,
1,
do_classifier_free_guidance=False,
negative_prompt=None,
)
except TypeError:
prompt_embeds_reference_2 = ori_pip.encode_prompt(
given_prompt2,
device,
1,
do_classifier_free_guidance=False,
negative_prompt=None,
)[0]
net2 = attack_model(ori_pip, latents=latents, input_ids=input_ids, mode=1, prompt_embeds_reference=prompt_embeds_reference_2, token_possibility=None, possibility_embedding=None)
noise = randn_tensor(net.latents.shape, generator=None, device=net.pip._execution_device, dtype=net.prompt_embeds_reference.dtype)
with torch.no_grad():
for i in range(10):
t = i * 100 +1
net.latents = latents
loss_value = net.forward_test(token_possibility=None, noise=noise, t=t, source=True)
net2.latents = latents
loss_value_2 = net2.forward_test(token_possibility=None, noise=noise, t=t, source=True)
print(t, loss_value.item(), loss_value_2.item())
noise_net = attack_noise_model(net, net2, ori_latents)
fn = identity_loss()
noise_net_attack = L2Attack(noise_net, fn, epsilon, steps, eps_iter=alpha, clip_min=clip_min, clip_max=clip_max, targeted=True, rand_init=False)
attacked_latents = noise_net_attack.perturb(latents, torch.zeros(1).cuda())
attacked_latents = ori_latents * mask_tensor + attacked_latents * (1-mask_tensor)
noise = randn_tensor(net.latents.shape, generator=None, device=net.pip._execution_device, dtype=net.prompt_embeds_reference.dtype)
with torch.no_grad():
image = pip.numpy_to_pil(pip.decode_latents(attacked_latents))
image[0].save(f"{saving_dirname}/{os.listdir(data_dirname)[class_id]}.png")
if __name__ == '__main__':
main(src_dataset_mode = 0, num_imgs = 10)
main(src_dataset_mode = 1, num_imgs = 2)