Spaces:
Sleeping
Sleeping
root
commited on
Commit
β’
0663fac
1
Parent(s):
55150dc
add random noise
Browse files
README.md
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
---
|
2 |
-
title:
|
3 |
emoji: π
|
4 |
colorFrom: gray
|
5 |
colorTo: green
|
@@ -7,4 +7,16 @@ sdk: docker
|
|
7 |
pinned: false
|
8 |
---
|
9 |
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
title: FGSM Project
|
3 |
emoji: π
|
4 |
colorFrom: gray
|
5 |
colorTo: green
|
|
|
7 |
pinned: false
|
8 |
---
|
9 |
|
10 |
+
This repository was developed inside a [devcontainer](https://containers.dev/).
|
11 |
+
|
12 |
+
If you are after speed, you can run this application locally.
|
13 |
+
|
14 |
+
1. Clone the repository
|
15 |
+
|
16 |
+
`git clone https://huggingface.co/spaces/niniack/fgsm-project`
|
17 |
+
|
18 |
+
2. Open up the project inside a devcontainer. Check [this](https://code.visualstudio.com/docs/devcontainers/containers) for instructions with VS Code.
|
19 |
+
|
20 |
+
3. Start the application
|
21 |
+
|
22 |
+
`panel serve /path/to/app.py/ --dev`
|
app.py
CHANGED
@@ -52,6 +52,7 @@ def run_forward_backward(image: Image, epsilon):
|
|
52 |
)
|
53 |
|
54 |
# Grab input
|
|
|
55 |
input_tensor = processor(image, return_tensors="pt")["pixel_values"]
|
56 |
input_tensor.requires_grad_(True)
|
57 |
|
@@ -70,11 +71,18 @@ def run_forward_backward(image: Image, epsilon):
|
|
70 |
# Denormalize input
|
71 |
mean = torch.tensor(processor.image_mean).view(1, -1, 1, 1)
|
72 |
std = torch.tensor(processor.image_std).view(1, -1, 1, 1)
|
73 |
-
input_tensor_denorm = input_tensor.detach() * std + mean
|
|
|
|
|
|
|
|
|
|
|
74 |
|
75 |
# FGSM attack
|
76 |
adv_input_tensor_denorm = fgsm_attack(
|
77 |
-
image=
|
|
|
|
|
78 |
)
|
79 |
|
80 |
# Normalize adversarial input tensor back to the input range
|
@@ -84,7 +92,6 @@ def run_forward_backward(image: Image, epsilon):
|
|
84 |
adv_output = model(adv_input_tensor)
|
85 |
adv_output = adv_output.logits
|
86 |
|
87 |
-
|
88 |
return (
|
89 |
output,
|
90 |
adv_output,
|
@@ -109,10 +116,10 @@ async def process_inputs(button_event, image_data: bytes, epsilon: float):
|
|
109 |
try:
|
110 |
# Open the image using PIL
|
111 |
pil_img = Image.open(BytesIO(image_data))
|
112 |
-
|
113 |
# Run forward + FGSM
|
114 |
-
clean_logits, adv_logits, input_tensor, adv_input_tensor =
|
115 |
-
image=pil_img, epsilon=epsilon
|
116 |
)
|
117 |
|
118 |
except Exception as e:
|
@@ -121,21 +128,37 @@ async def process_inputs(button_event, image_data: bytes, epsilon: float):
|
|
121 |
|
122 |
img = pn.pane.Image(
|
123 |
to_pil_image(input_tensor, do_rescale=True),
|
124 |
-
height=
|
125 |
align="center",
|
126 |
)
|
127 |
|
128 |
# Convert image for visualizing
|
|
|
129 |
adv_img = pn.pane.Image(
|
130 |
-
|
131 |
-
height=
|
132 |
align="center",
|
133 |
)
|
134 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
135 |
# Build the results column
|
136 |
k_val = 5
|
137 |
results = pn.Column(
|
138 |
-
pn.Row("###### Uploaded", "###### Adversarial"),
|
|
|
|
|
|
|
139 |
)
|
140 |
|
141 |
# Get likelihoods
|
@@ -150,43 +173,43 @@ async def process_inputs(button_event, image_data: bytes, epsilon: float):
|
|
150 |
# Get top k values and indices
|
151 |
vals_topk_clean, idx_topk_clean = torch.topk(likelihood_tensor, k=k_val)
|
152 |
label_bars = pn.Column()
|
153 |
-
|
154 |
for idx, val in zip(idx_topk_clean, vals_topk_clean):
|
155 |
prob = val.item()
|
156 |
row_label = pn.widgets.StaticText(
|
157 |
-
name=f"{classes[idx]}",
|
158 |
-
value=f"{prob:.2%}",
|
159 |
-
align="center"
|
160 |
)
|
161 |
row_bar = pn.indicators.Progress(
|
162 |
value=int(prob * 100),
|
163 |
sizing_mode="stretch_width",
|
164 |
-
bar_color="success"
|
|
|
|
|
165 |
margin=(0, 10),
|
166 |
design=pn.theme.Material,
|
167 |
)
|
168 |
label_bars.append(pn.Column(row_label, row_bar))
|
169 |
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
|
188 |
label_bars_rows.append(label_bars)
|
189 |
-
|
190 |
results.append(label_bars_rows)
|
191 |
|
192 |
yield results
|
@@ -194,7 +217,7 @@ async def process_inputs(button_event, image_data: bytes, epsilon: float):
|
|
194 |
except Exception as e:
|
195 |
yield f"##### Something went wrong! \n {e}"
|
196 |
return
|
197 |
-
|
198 |
finally:
|
199 |
main.disabled = False
|
200 |
|
@@ -213,37 +236,68 @@ file_input = pn.widgets.FileInput(name="Upload a PNG image", accept=".png,.jpg")
|
|
213 |
|
214 |
# Epsilon
|
215 |
epsilon_slider = pn.widgets.FloatSlider(
|
216 |
-
name=r"$$\epsilon$$
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
217 |
)
|
218 |
|
219 |
-
#
|
220 |
-
|
221 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
222 |
############################################
|
223 |
|
224 |
# Organize widgets in a column
|
225 |
input_widgets = pn.Column(
|
226 |
"""
|
227 |
-
###### Classify an image with a pre-trained [ResNet18](https://huggingface.co/microsoft/resnet-18) and generate an adversarial example.\n
|
|
|
|
|
228 |
|
229 |
-
Please be patient with the application, it is running on a low-resource device
|
230 |
""",
|
231 |
file_input,
|
232 |
-
epsilon_slider,
|
233 |
)
|
234 |
|
235 |
# Add interactivity
|
236 |
interactive_result = pn.panel(
|
237 |
pn.bind(
|
238 |
-
process_inputs,
|
|
|
|
|
|
|
239 |
),
|
240 |
height=600,
|
241 |
)
|
242 |
|
243 |
footer = pn.pane.Markdown(
|
244 |
"""
|
245 |
-
<br><br
|
246 |
-
|
|
|
247 |
"""
|
248 |
)
|
249 |
|
@@ -259,7 +313,7 @@ title = "Adversarial Sample Generation"
|
|
259 |
pn.template.BootstrapTemplate(
|
260 |
title=title,
|
261 |
main=main,
|
262 |
-
main_max_width="min(
|
263 |
header_background="#101820",
|
264 |
).servable(title=title)
|
265 |
|
|
|
52 |
)
|
53 |
|
54 |
# Grab input
|
55 |
+
processor.crop_pct = 1
|
56 |
input_tensor = processor(image, return_tensors="pt")["pixel_values"]
|
57 |
input_tensor.requires_grad_(True)
|
58 |
|
|
|
71 |
# Denormalize input
|
72 |
mean = torch.tensor(processor.image_mean).view(1, -1, 1, 1)
|
73 |
std = torch.tensor(processor.image_std).view(1, -1, 1, 1)
|
74 |
+
input_tensor_denorm = input_tensor.clone().detach() * std + mean
|
75 |
+
|
76 |
+
# Add noise to input
|
77 |
+
random_noise = torch.sign(torch.randn_like(input_tensor)) * 0.02
|
78 |
+
input_tensor_denorm_noised = torch.clamp(input_tensor_denorm + random_noise, 0, 1)
|
79 |
+
# input_tensor_denorm_noised = input_tensor_denorm
|
80 |
|
81 |
# FGSM attack
|
82 |
adv_input_tensor_denorm = fgsm_attack(
|
83 |
+
image=input_tensor_denorm_noised,
|
84 |
+
epsilon=epsilon,
|
85 |
+
data_grad=input_tensor.grad.data,
|
86 |
)
|
87 |
|
88 |
# Normalize adversarial input tensor back to the input range
|
|
|
92 |
adv_output = model(adv_input_tensor)
|
93 |
adv_output = adv_output.logits
|
94 |
|
|
|
95 |
return (
|
96 |
output,
|
97 |
adv_output,
|
|
|
116 |
try:
|
117 |
# Open the image using PIL
|
118 |
pil_img = Image.open(BytesIO(image_data))
|
119 |
+
|
120 |
# Run forward + FGSM
|
121 |
+
clean_logits, adv_logits, input_tensor, adv_input_tensor = (
|
122 |
+
run_forward_backward(image=pil_img, epsilon=epsilon)
|
123 |
)
|
124 |
|
125 |
except Exception as e:
|
|
|
128 |
|
129 |
img = pn.pane.Image(
|
130 |
to_pil_image(input_tensor, do_rescale=True),
|
131 |
+
height=300,
|
132 |
align="center",
|
133 |
)
|
134 |
|
135 |
# Convert image for visualizing
|
136 |
+
adv_img_pil = to_pil_image(adv_input_tensor, do_rescale=True)
|
137 |
adv_img = pn.pane.Image(
|
138 |
+
adv_img_pil,
|
139 |
+
height=300,
|
140 |
align="center",
|
141 |
)
|
142 |
|
143 |
+
# Download image button
|
144 |
+
adv_img_bytes = io.BytesIO()
|
145 |
+
adv_img_pil.save(adv_img_bytes, format="PNG")
|
146 |
+
# download = pn.widgets.FileDownload(
|
147 |
+
# to_pil_image(adv_img_bytes, do_rescale=True),
|
148 |
+
# embed=True,
|
149 |
+
# filename="adv_img.png",
|
150 |
+
# button_type="primary",
|
151 |
+
# button_style="outline",
|
152 |
+
# width_policy="min",
|
153 |
+
# )
|
154 |
+
|
155 |
# Build the results column
|
156 |
k_val = 5
|
157 |
results = pn.Column(
|
158 |
+
pn.Row("###### Uploaded", "###### Adversarial"),
|
159 |
+
pn.Row(img, adv_img),
|
160 |
+
# pn.Row(pn.Spacer(), download),
|
161 |
+
f" ###### Top {k_val} class predictions",
|
162 |
)
|
163 |
|
164 |
# Get likelihoods
|
|
|
173 |
# Get top k values and indices
|
174 |
vals_topk_clean, idx_topk_clean = torch.topk(likelihood_tensor, k=k_val)
|
175 |
label_bars = pn.Column()
|
176 |
+
|
177 |
for idx, val in zip(idx_topk_clean, vals_topk_clean):
|
178 |
prob = val.item()
|
179 |
row_label = pn.widgets.StaticText(
|
180 |
+
name=f"{classes[idx]}", value=f"{prob:.2%}", align="center"
|
|
|
|
|
181 |
)
|
182 |
row_bar = pn.indicators.Progress(
|
183 |
value=int(prob * 100),
|
184 |
sizing_mode="stretch_width",
|
185 |
+
bar_color="success"
|
186 |
+
if prob > 0.7
|
187 |
+
else "warning", # Dynamic color based on value
|
188 |
margin=(0, 10),
|
189 |
design=pn.theme.Material,
|
190 |
)
|
191 |
label_bars.append(pn.Column(row_label, row_bar))
|
192 |
|
193 |
+
# for likelihood_tensor in likelihoods:
|
194 |
+
# # Get top
|
195 |
+
# vals_topk_clean, idx_topk_clean = torch.topk(likelihood_tensor, k=k_val)
|
196 |
+
# label_bars = pn.Column()
|
197 |
+
# for idx, val in zip(idx_topk_clean, vals_topk_clean):
|
198 |
+
# prob = val.item()
|
199 |
+
# row_label = pn.widgets.StaticText(
|
200 |
+
# name=f"{classes[idx]}", value=f"{prob:.2%}", align="center"
|
201 |
+
# )
|
202 |
+
# row_bar = pn.indicators.Progress(
|
203 |
+
# value=int(prob * 100),
|
204 |
+
# sizing_mode="stretch_width",
|
205 |
+
# bar_color="secondary",
|
206 |
+
# margin=(0, 10),
|
207 |
+
# design=pn.theme.Material,
|
208 |
+
# )
|
209 |
+
# label_bars.append(pn.Column(row_label, row_bar))
|
210 |
|
211 |
label_bars_rows.append(label_bars)
|
212 |
+
|
213 |
results.append(label_bars_rows)
|
214 |
|
215 |
yield results
|
|
|
217 |
except Exception as e:
|
218 |
yield f"##### Something went wrong! \n {e}"
|
219 |
return
|
220 |
+
|
221 |
finally:
|
222 |
main.disabled = False
|
223 |
|
|
|
236 |
|
237 |
# Epsilon
|
238 |
epsilon_slider = pn.widgets.FloatSlider(
|
239 |
+
name=r"$$\epsilon$$ parameter for FGSM",
|
240 |
+
start=0,
|
241 |
+
end=0.1,
|
242 |
+
step=0.005,
|
243 |
+
value=0.000,
|
244 |
+
format="1[.]000",
|
245 |
+
align="center",
|
246 |
+
max_width=500,
|
247 |
+
width_policy="max",
|
248 |
)
|
249 |
|
250 |
+
# alpha_slider = pn.widgets.FloatSlider(
|
251 |
+
# name=r"$$\alpha$$ parameter for Gaussian noise",
|
252 |
+
# start=0,
|
253 |
+
# end=0.1,
|
254 |
+
# step=0.005,
|
255 |
+
# value=0.000,
|
256 |
+
# format="1[.]000",
|
257 |
+
# align="center",
|
258 |
+
# max_width=500,
|
259 |
+
# width_policy="max"
|
260 |
+
|
261 |
+
# )
|
262 |
+
|
263 |
+
# Regenerate button
|
264 |
+
regenerate = pn.widgets.Button(
|
265 |
+
name="Regenerate",
|
266 |
+
button_type="primary",
|
267 |
+
width_policy="min",
|
268 |
+
max_width=105,
|
269 |
+
)
|
270 |
############################################
|
271 |
|
272 |
# Organize widgets in a column
|
273 |
input_widgets = pn.Column(
|
274 |
"""
|
275 |
+
###### Classify an image (png/jpeg) with a pre-trained [ResNet18](https://huggingface.co/microsoft/resnet-18) and generate an adversarial example.\n
|
276 |
+
|
277 |
+
Wondering where the class names come from? Find the list of ImageNet-1K classes [here.](https://deeplearning.cms.waikato.ac.nz/user-guide/class-maps/IMAGENET/)
|
278 |
|
279 |
+
*Please be patient with the application, it is running on a low-resource device.*
|
280 |
""",
|
281 |
file_input,
|
282 |
+
pn.Row(epsilon_slider, pn.Spacer(width_policy="min", max_width=25), regenerate),
|
283 |
)
|
284 |
|
285 |
# Add interactivity
|
286 |
interactive_result = pn.panel(
|
287 |
pn.bind(
|
288 |
+
process_inputs,
|
289 |
+
regenerate,
|
290 |
+
file_input.param.value,
|
291 |
+
epsilon_slider.param.value,
|
292 |
),
|
293 |
height=600,
|
294 |
)
|
295 |
|
296 |
footer = pn.pane.Markdown(
|
297 |
"""
|
298 |
+
<br><br>
|
299 |
+
|
300 |
+
If the application is too slow for you, head over to the README to get this running locally.
|
301 |
"""
|
302 |
)
|
303 |
|
|
|
313 |
pn.template.BootstrapTemplate(
|
314 |
title=title,
|
315 |
main=main,
|
316 |
+
main_max_width="min(75%, 698px)",
|
317 |
header_background="#101820",
|
318 |
).servable(title=title)
|
319 |
|