Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -3,114 +3,184 @@ import torch
|
|
3 |
import torch.nn as nn
|
4 |
import torch.optim as optim
|
5 |
from torchvision import transforms, models
|
6 |
-
from art.attacks.evasion import
|
|
|
|
|
|
|
|
|
7 |
from art.estimators.classification import PyTorchClassifier
|
8 |
-
from PIL import Image
|
9 |
import numpy as np
|
10 |
import os
|
11 |
-
import io
|
12 |
from blind_watermark import WaterMark
|
|
|
|
|
13 |
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
num_ftrs = model.fc.in_features
|
19 |
-
model.fc = nn.Linear(num_ftrs, 10)
|
20 |
|
21 |
-
|
22 |
-
|
23 |
-
|
|
|
24 |
|
25 |
-
# ์์ค ํจ์์ ์ตํฐ๋ง์ด์ ์ค์
|
26 |
criterion = nn.CrossEntropyLoss()
|
27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
|
29 |
-
|
30 |
-
|
31 |
-
model=model,
|
32 |
loss=criterion,
|
33 |
-
optimizer=
|
34 |
-
input_shape=(3,
|
35 |
nb_classes=10,
|
36 |
)
|
37 |
|
38 |
-
|
|
|
|
|
|
|
|
|
39 |
def preprocess_image(image):
|
40 |
transform = transforms.Compose([
|
|
|
41 |
transforms.ToTensor(),
|
42 |
-
transforms.Normalize(mean=[0.485, 0.456, 0.406],
|
43 |
-
std=[0.229, 0.224, 0.225])
|
44 |
])
|
45 |
-
return transform(image).unsqueeze(0).to(
|
46 |
-
|
47 |
-
# FGSM ๊ณต๊ฒฉ ์ ์ฉ ๋ฐ ์ด๋ฏธ์ง ์ฒ๋ฆฌ ํจ์
|
48 |
-
def generate_adversarial_image(image, eps_value):
|
49 |
-
img_tensor = preprocess_image(image)
|
50 |
-
|
51 |
-
# FGSM ๊ณต๊ฒฉ ์ค์
|
52 |
-
attack = FastGradientMethod(estimator=classifier, eps=eps_value)
|
53 |
-
|
54 |
-
# ์ ๋์ ์์ ์์ฑ
|
55 |
-
adv_img_tensor = attack.generate(x=img_tensor.cpu().numpy())
|
56 |
-
adv_img_tensor = torch.tensor(adv_img_tensor).to(device)
|
57 |
|
58 |
-
|
59 |
-
adv_img_np =
|
60 |
mean = np.array([0.485, 0.456, 0.406])
|
61 |
std = np.array([0.229, 0.224, 0.225])
|
62 |
adv_img_np = (adv_img_np * std[:, None, None]) + mean[:, None, None]
|
63 |
adv_img_np = np.clip(adv_img_np, 0, 1)
|
64 |
adv_img_np = adv_img_np.transpose(1, 2, 0)
|
65 |
-
|
66 |
-
# PIL ์ด๋ฏธ์ง๋ก ๋ณํ
|
67 |
-
adv_image_pil = Image.fromarray((adv_img_np * 255).astype(np.uint8))
|
68 |
-
|
69 |
return adv_image_pil
|
70 |
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
# ์ด๋ฏธ์ง ๋ฐ์ดํธ ๋ฐ์ดํฐ๋ฅผ ์์ ํ์ผ๋ก ์ ์ฅ
|
76 |
-
temp_image_path = "temp_image.png"
|
77 |
-
image_pil.save(temp_image_path)
|
78 |
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
|
87 |
-
|
88 |
-
|
|
|
|
|
|
|
89 |
|
90 |
-
|
91 |
-
|
92 |
-
|
|
|
93 |
|
94 |
-
|
|
|
|
|
95 |
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
# ์ ๋์ ์ด๋ฏธ์ง์ ์ํฐ๋งํฌ ์ฝ์
|
102 |
-
watermarked_image = apply_watermark(adv_image, wm_text, int(password_img), int(password_wm))
|
103 |
|
104 |
-
|
|
|
|
|
|
|
105 |
|
106 |
-
|
107 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
fn=process_image,
|
109 |
-
inputs=[
|
110 |
-
gr.
|
111 |
-
gr.
|
112 |
-
gr.
|
113 |
-
gr.
|
|
|
|
|
|
|
114 |
],
|
115 |
-
outputs=
|
116 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
import torch.nn as nn
|
4 |
import torch.optim as optim
|
5 |
from torchvision import transforms, models
|
6 |
+
from art.attacks.evasion import (
|
7 |
+
FastGradientMethod, CarliniL2Method, DeepFool, AutoAttack,
|
8 |
+
ProjectedGradientDescent, BasicIterativeMethod, SpatialTransformation,
|
9 |
+
MomentumIterativeMethod, SaliencyMapMethod, NewtonFool
|
10 |
+
)
|
11 |
from art.estimators.classification import PyTorchClassifier
|
12 |
+
from PIL import Image, ImageOps
|
13 |
import numpy as np
|
14 |
import os
|
|
|
15 |
from blind_watermark import WaterMark
|
16 |
+
from torchvision.models import resnet50, vgg16, ResNet50_Weights, VGG16_Weights
|
17 |
+
import tempfile
|
18 |
|
19 |
+
resnet_model = resnet50(weights=ResNet50_Weights.DEFAULT)
|
20 |
+
num_ftrs_resnet = resnet_model.fc.in_features
|
21 |
+
resnet_model.fc = nn.Linear(num_ftrs_resnet, 10)
|
22 |
+
resnet_model = resnet_model.to("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
|
23 |
|
24 |
+
vgg_model = vgg16(weights=VGG16_Weights.DEFAULT)
|
25 |
+
num_ftrs_vgg = vgg_model.classifier[6].in_features
|
26 |
+
vgg_model.classifier[6] = nn.Linear(num_ftrs_vgg, 10)
|
27 |
+
vgg_model = vgg_model.to("cuda" if torch.cuda.is_available() else "cpu")
|
28 |
|
|
|
29 |
criterion = nn.CrossEntropyLoss()
|
30 |
+
optimizer_resnet = optim.Adam(resnet_model.parameters(), lr=0.001)
|
31 |
+
optimizer_vgg = optim.Adam(vgg_model.parameters(), lr=0.001)
|
32 |
+
|
33 |
+
resnet_classifier = PyTorchClassifier(
|
34 |
+
model=resnet_model,
|
35 |
+
loss=criterion,
|
36 |
+
optimizer=optimizer_resnet,
|
37 |
+
input_shape=(3, 224, 224),
|
38 |
+
nb_classes=10,
|
39 |
+
)
|
40 |
|
41 |
+
vgg_classifier = PyTorchClassifier(
|
42 |
+
model=vgg_model,
|
|
|
43 |
loss=criterion,
|
44 |
+
optimizer=optimizer_vgg,
|
45 |
+
input_shape=(3, 224, 224),
|
46 |
nb_classes=10,
|
47 |
)
|
48 |
|
49 |
+
models_dict = {
|
50 |
+
"ResNet50": resnet_classifier,
|
51 |
+
"VGG16": vgg_classifier
|
52 |
+
}
|
53 |
+
|
54 |
def preprocess_image(image):
|
55 |
transform = transforms.Compose([
|
56 |
+
transforms.Resize((224, 224)),
|
57 |
transforms.ToTensor(),
|
58 |
+
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
|
|
|
59 |
])
|
60 |
+
return transform(image).unsqueeze(0).to("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
|
62 |
+
def postprocess_image(tensor, original_size):
|
63 |
+
adv_img_np = tensor.squeeze(0).cpu().numpy()
|
64 |
mean = np.array([0.485, 0.456, 0.406])
|
65 |
std = np.array([0.229, 0.224, 0.225])
|
66 |
adv_img_np = (adv_img_np * std[:, None, None]) + mean[:, None, None]
|
67 |
adv_img_np = np.clip(adv_img_np, 0, 1)
|
68 |
adv_img_np = adv_img_np.transpose(1, 2, 0)
|
69 |
+
adv_image_pil = Image.fromarray((adv_img_np * 255).astype(np.uint8)).resize(original_size)
|
|
|
|
|
|
|
70 |
return adv_image_pil
|
71 |
|
72 |
+
def generate_adversarial_image(image, model_name, attack_types, eps_value):
|
73 |
+
original_size = image.size
|
74 |
+
img_tensor = preprocess_image(image)
|
|
|
|
|
|
|
|
|
75 |
|
76 |
+
classifier = models_dict[model_name]
|
77 |
+
|
78 |
+
try:
|
79 |
+
for attack_type in attack_types:
|
80 |
+
if attack_type == "FGSM":
|
81 |
+
attack = FastGradientMethod(estimator=classifier, eps=eps_value)
|
82 |
+
elif attack_type == "C&W":
|
83 |
+
attack = CarliniL2Method(classifier=classifier, confidence=0.05)
|
84 |
+
elif attack_type == "DeepFool":
|
85 |
+
attack = DeepFool(classifier=classifier, max_iter=20)
|
86 |
+
elif attack_type == "AutoAttack":
|
87 |
+
attack = AutoAttack(estimator=classifier, eps=eps_value, batch_size=1)
|
88 |
+
elif attack_type == "PGD":
|
89 |
+
attack = ProjectedGradientDescent(estimator=classifier, eps=eps_value, eps_step=eps_value / 10, max_iter=40)
|
90 |
+
elif attack_type == "BIM":
|
91 |
+
attack = BasicIterativeMethod(estimator=classifier, eps=eps_value, eps_step=eps_value / 10, max_iter=10)
|
92 |
+
elif attack_type == "STA":
|
93 |
+
attack = SpatialTransformation(estimator=classifier, max_translation=0.2)
|
94 |
+
elif attack_type == "MIM":
|
95 |
+
attack = MomentumIterativeMethod(estimator=classifier, eps=eps_value, eps_step=eps_value / 10, max_iter=10)
|
96 |
+
elif attack_type == "JSMA":
|
97 |
+
attack = SaliencyMapMethod(classifier=classifier, theta=0.1, gamma=0.1)
|
98 |
+
elif attack_type == "NewtonFool":
|
99 |
+
attack = NewtonFool(classifier=classifier, max_iter=20)
|
100 |
+
|
101 |
+
adv_img_np = attack.generate(x=img_tensor.cpu().numpy())
|
102 |
+
img_tensor = torch.tensor(adv_img_np).to("cuda" if torch.cuda.is_available() else "cpu")
|
103 |
+
except Exception as e:
|
104 |
+
print(f"Error in adversarial generation: {e}")
|
105 |
+
return image
|
106 |
+
|
107 |
+
adv_image_pil = postprocess_image(img_tensor, original_size)
|
108 |
+
return adv_image_pil
|
109 |
|
110 |
+
def apply_watermark(image_pil, wm_text="ํ
์คํธ ์ฝ์
", password_img=0, password_wm=0):
|
111 |
+
try:
|
112 |
+
bwm = WaterMark(password_img=password_img, password_wm=password_wm)
|
113 |
+
temp_image_path = tempfile.mktemp(suffix=".png")
|
114 |
+
image_pil.save(temp_image_path, format="PNG")
|
115 |
|
116 |
+
bwm.read_img(temp_image_path)
|
117 |
+
bwm.read_wm(wm_text, mode='str')
|
118 |
+
output_path = tempfile.mktemp(suffix=".png")
|
119 |
+
bwm.embed(output_path)
|
120 |
|
121 |
+
result_image = Image.open(output_path).convert("RGB")
|
122 |
+
os.remove(temp_image_path)
|
123 |
+
os.remove(output_path)
|
124 |
|
125 |
+
return result_image
|
126 |
+
except Exception as e:
|
127 |
+
print(f"Error in apply_watermark: {str(e)}")
|
128 |
+
return image_pil
|
|
|
|
|
|
|
129 |
|
130 |
+
def extract_watermark(image_pil, password_img=0, password_wm=0):
|
131 |
+
bwm = WaterMark(password_img=password_img, password_wm=password_wm)
|
132 |
+
temp_image_path = tempfile.mktemp(suffix=".png")
|
133 |
+
image_pil.save(temp_image_path, format="PNG")
|
134 |
|
135 |
+
extracted_wm_text = bwm.extract(temp_image_path, wm_shape=(32, 32), mode='str')
|
136 |
+
os.remove(temp_image_path)
|
137 |
+
return extracted_wm_text
|
138 |
+
|
139 |
+
def process_image(image, model_name, attack_types, eps_value, wm_text, password_img, password_wm):
|
140 |
+
try:
|
141 |
+
adv_image = generate_adversarial_image(image, model_name, attack_types, eps_value)
|
142 |
+
except Exception as e:
|
143 |
+
error_message = f"Error in adversarial generation: {str(e)}"
|
144 |
+
return image, error_message, None, None, None
|
145 |
+
|
146 |
+
try:
|
147 |
+
watermarked_image = apply_watermark(adv_image, wm_text, int(password_img), int(password_wm))
|
148 |
+
except Exception as e:
|
149 |
+
error_message = f"Error in watermarking: {str(e)}"
|
150 |
+
return image, adv_image, error_message, None, None
|
151 |
+
|
152 |
+
try:
|
153 |
+
extracted_wm_text = extract_watermark(watermarked_image, int(password_img), int(password_wm))
|
154 |
+
except Exception as e:
|
155 |
+
error_message = f"Error in watermark extraction: {str(e)}"
|
156 |
+
return image, adv_image, watermarked_image, error_message, None
|
157 |
+
|
158 |
+
output_path = tempfile.mktemp(suffix=".png")
|
159 |
+
watermarked_image.save(output_path, format="PNG")
|
160 |
+
return image, adv_image, watermarked_image, extracted_wm_text, output_path
|
161 |
+
|
162 |
+
def download_image_as_png(image_path):
|
163 |
+
with open(image_path, "rb") as file:
|
164 |
+
return file.read(), "image/png"
|
165 |
+
|
166 |
+
interface = gr.Interface(
|
167 |
fn=process_image,
|
168 |
+
inputs=[
|
169 |
+
gr.Image(type="pil", label="์ด๋ฏธ์ง๋ฅผ ์
๋ก๋ํ์ธ์"),
|
170 |
+
gr.Dropdown(choices=["ResNet50", "VGG16"], label="๋ชจ๋ธ ์ ํ"),
|
171 |
+
gr.CheckboxGroup(choices=["FGSM", "C&W", "DeepFool", "AutoAttack", "PGD", "BIM", "STA", "MIM", "JSMA", "NewtonFool"], label="๊ณต๊ฒฉ ์ ํ ์ ํ"),
|
172 |
+
gr.Slider(0.001, 0.9, step=0.001, value=0.005, label="Epsilon ๊ฐ ์ค์ (๋
ธ์ด์ฆ ๊ฐ๋)"),
|
173 |
+
gr.Textbox(label="์ํฐ๋งํฌ ํ
์คํธ ์
๋ ฅ", value="ํ
์คํธ ์ฝ์
"),
|
174 |
+
gr.Number(label="์ด๋ฏธ์ง ๋น๋ฐ๋ฒํธ", value=0),
|
175 |
+
gr.Number(label="์ํฐ๋งํฌ ๋น๋ฐ๋ฒํธ", value=0)
|
176 |
],
|
177 |
+
outputs=[
|
178 |
+
gr.Image(type="numpy", label="์๋ณธ ์ด๋ฏธ์ง"),
|
179 |
+
gr.Image(type="numpy", label="์ ๋์ ์ด๋ฏธ์ง ์์ฑ ๋จ๊ณ"),
|
180 |
+
gr.Image(type="numpy", label="์ํฐ๋งํฌ๊ฐ ์ฝ์
๋ ์ต์ข
์ด๋ฏธ์ง"),
|
181 |
+
gr.Textbox(label="์ถ์ถ๋ ์ํฐ๋งํฌ ํ
์คํธ"),
|
182 |
+
gr.File(label="PNG๋ก ๋ค์ด๋ก๋")
|
183 |
+
]
|
184 |
+
)
|
185 |
+
|
186 |
+
interface.launch(debug=True, share=True)
|