Spaces:
Sleeping
Sleeping
Thomas Chardonnens
commited on
Commit
•
127130c
1
Parent(s):
cbb90a5
baseline, wip
Browse files- app.py +434 -4
- client_server_interface.py +150 -0
- common.py +29 -0
- input_examples/eeg-1.png +0 -0
- input_examples/eeg-2.png +0 -0
- requirements.txt +2 -0
- seizure_detection.py +0 -0
- server.py +97 -0
app.py
CHANGED
@@ -1,7 +1,437 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
2 |
|
3 |
-
|
4 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
-
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""A local gradio app that detects seizures with EEG using FHE."""
|
2 |
+
from PIL import Image
|
3 |
+
import os
|
4 |
+
import shutil
|
5 |
+
import subprocess
|
6 |
+
import time
|
7 |
import gradio as gr
|
8 |
+
import numpy
|
9 |
+
import requests
|
10 |
+
from itertools import chain
|
11 |
|
12 |
+
from common import (
|
13 |
+
CLIENT_TMP_PATH,
|
14 |
+
SERVER_TMP_PATH,
|
15 |
+
EXAMPLES,
|
16 |
+
INPUT_SHAPE,
|
17 |
+
KEYS_PATH,
|
18 |
+
REPO_DIR,
|
19 |
+
SERVER_URL,
|
20 |
+
)
|
21 |
+
from client_server_interface import FHEClient
|
22 |
|
23 |
+
# Uncomment here to have both the server and client in the same terminal
|
24 |
+
subprocess.Popen(["uvicorn", "server:app"], cwd=REPO_DIR)
|
25 |
+
time.sleep(3)
|
26 |
+
|
27 |
+
def shorten_bytes_object(bytes_object, limit=500):
|
28 |
+
"""Shorten the input bytes object to a given length.
|
29 |
+
|
30 |
+
Encrypted data is too large for displaying it in the browser using Gradio. This function
|
31 |
+
provides a shorten representation of it.
|
32 |
+
|
33 |
+
Args:
|
34 |
+
bytes_object (bytes): The input to shorten
|
35 |
+
limit (int): The length to consider. Default to 500.
|
36 |
+
|
37 |
+
Returns:
|
38 |
+
str: Hexadecimal string shorten representation of the input byte object.
|
39 |
+
|
40 |
+
"""
|
41 |
+
# Define a shift for better display
|
42 |
+
shift = 100
|
43 |
+
return bytes_object[shift : limit + shift].hex()
|
44 |
+
|
45 |
+
def get_client(user_id):
|
46 |
+
"""Get the client API.
|
47 |
+
|
48 |
+
Args:
|
49 |
+
user_id (int): The current user's ID.
|
50 |
+
|
51 |
+
Returns:
|
52 |
+
FHEClient: The client API.
|
53 |
+
"""
|
54 |
+
return FHEClient(
|
55 |
+
key_dir=KEYS_PATH / f"seizure_detection_{user_id}",
|
56 |
+
)
|
57 |
+
|
58 |
+
def get_client_file_path(name, user_id):
|
59 |
+
"""Get the correct temporary file path for the client.
|
60 |
+
|
61 |
+
Args:
|
62 |
+
name (str): The desired file name.
|
63 |
+
user_id (int): The current user's ID.
|
64 |
+
|
65 |
+
Returns:
|
66 |
+
pathlib.Path: The file path.
|
67 |
+
"""
|
68 |
+
return CLIENT_TMP_PATH / f"{name}_seizure_detection_{user_id}"
|
69 |
+
|
70 |
+
def clean_temporary_files(n_keys=20):
|
71 |
+
"""Clean keys and encrypted images.
|
72 |
+
|
73 |
+
A maximum of n_keys keys and associated temporary files are allowed to be stored. Once this
|
74 |
+
limit is reached, the oldest files are deleted.
|
75 |
+
|
76 |
+
Args:
|
77 |
+
n_keys (int): The maximum number of keys and associated files to be stored. Default to 20.
|
78 |
+
|
79 |
+
"""
|
80 |
+
# Get the oldest key files in the key directory
|
81 |
+
key_dirs = sorted(KEYS_PATH.iterdir(), key=os.path.getmtime)
|
82 |
+
|
83 |
+
# If more than n_keys keys are found, remove the oldest
|
84 |
+
user_ids = []
|
85 |
+
if len(key_dirs) > n_keys:
|
86 |
+
n_keys_to_delete = len(key_dirs) - n_keys
|
87 |
+
for key_dir in key_dirs[:n_keys_to_delete]:
|
88 |
+
user_ids.append(key_dir.name)
|
89 |
+
shutil.rmtree(key_dir)
|
90 |
+
|
91 |
+
# Get all the encrypted objects in the temporary folder
|
92 |
+
client_files = CLIENT_TMP_PATH.iterdir()
|
93 |
+
server_files = SERVER_TMP_PATH.iterdir()
|
94 |
+
|
95 |
+
# Delete all files related to the ids whose keys were deleted
|
96 |
+
for file in chain(client_files, server_files):
|
97 |
+
for user_id in user_ids:
|
98 |
+
if user_id in file.name:
|
99 |
+
file.unlink()
|
100 |
+
|
101 |
+
def keygen():
|
102 |
+
"""Generate the private key for seizure detection.
|
103 |
+
|
104 |
+
Returns:
|
105 |
+
(user_id, True) (Tuple[int, bool]): The current user's ID and a boolean used for visual display.
|
106 |
+
|
107 |
+
"""
|
108 |
+
# Clean temporary files
|
109 |
+
clean_temporary_files()
|
110 |
+
|
111 |
+
# Create an ID for the current user
|
112 |
+
user_id = numpy.random.randint(0, 2**32)
|
113 |
+
|
114 |
+
# Retrieve the client API
|
115 |
+
client = get_client(user_id)
|
116 |
+
|
117 |
+
# Generate a private key
|
118 |
+
client.generate_private_and_evaluation_keys(force=True)
|
119 |
+
|
120 |
+
# Retrieve the serialized evaluation key
|
121 |
+
evaluation_key = client.get_serialized_evaluation_keys()
|
122 |
+
|
123 |
+
# Save evaluation_key as bytes in a file as it is too large to pass through regular Gradio
|
124 |
+
# buttons (see https://github.com/gradio-app/gradio/issues/1877)
|
125 |
+
evaluation_key_path = get_client_file_path("evaluation_key", user_id)
|
126 |
+
|
127 |
+
with evaluation_key_path.open("wb") as evaluation_key_file:
|
128 |
+
evaluation_key_file.write(evaluation_key)
|
129 |
+
|
130 |
+
return (user_id, True)
|
131 |
+
|
132 |
+
def encrypt(user_id, input_image):
|
133 |
+
"""Encrypt the given image for seizure detection.
|
134 |
+
|
135 |
+
Args:
|
136 |
+
user_id (int): The current user's ID.
|
137 |
+
input_image (numpy.ndarray): The image to encrypt.
|
138 |
+
|
139 |
+
Returns:
|
140 |
+
(input_image, encrypted_image_short) (Tuple[bytes]): The encrypted image and one of its
|
141 |
+
representation.
|
142 |
+
|
143 |
+
"""
|
144 |
+
if user_id == "":
|
145 |
+
raise gr.Error("Please generate the private key first.")
|
146 |
+
|
147 |
+
if input_image is None:
|
148 |
+
raise gr.Error("Please choose an image first.")
|
149 |
+
|
150 |
+
if input_image.shape[-1] != 3:
|
151 |
+
raise ValueError(f"Input image must have 3 channels (RGB). Current shape: {input_image.shape}")
|
152 |
+
|
153 |
+
# Resize the image if it hasn't the shape (224, 224, 3)
|
154 |
+
if input_image.shape != (224, 224, 3):
|
155 |
+
input_image_pil = Image.fromarray(input_image)
|
156 |
+
input_image_pil = input_image_pil.resize((224, 224))
|
157 |
+
input_image = numpy.array(input_image_pil)
|
158 |
+
|
159 |
+
# Retrieve the client API
|
160 |
+
client = get_client(user_id)
|
161 |
+
|
162 |
+
# Pre-process, encrypt and serialize the image
|
163 |
+
encrypted_image = client.encrypt_serialize(input_image)
|
164 |
+
|
165 |
+
# Save encrypted_image to bytes in a file, since too large to pass through regular Gradio
|
166 |
+
# buttons, https://github.com/gradio-app/gradio/issues/1877
|
167 |
+
encrypted_image_path = get_client_file_path("encrypted_image", user_id)
|
168 |
+
|
169 |
+
with encrypted_image_path.open("wb") as encrypted_image_file:
|
170 |
+
encrypted_image_file.write(encrypted_image)
|
171 |
+
|
172 |
+
# Create a truncated version of the encrypted image for display
|
173 |
+
encrypted_image_short = shorten_bytes_object(encrypted_image)
|
174 |
+
|
175 |
+
return (resize_img(input_image), encrypted_image_short)
|
176 |
+
|
177 |
+
def send_input(user_id):
|
178 |
+
"""Send the encrypted input image as well as the evaluation key to the server.
|
179 |
+
|
180 |
+
Args:
|
181 |
+
user_id (int): The current user's ID.
|
182 |
+
"""
|
183 |
+
# Get the evaluation key path
|
184 |
+
evaluation_key_path = get_client_file_path("evaluation_key", user_id)
|
185 |
+
|
186 |
+
if user_id == "" or not evaluation_key_path.is_file():
|
187 |
+
raise gr.Error("Please generate the private key first.")
|
188 |
+
|
189 |
+
encrypted_input_path = get_client_file_path("encrypted_image", user_id)
|
190 |
+
|
191 |
+
if not encrypted_input_path.is_file():
|
192 |
+
raise gr.Error("Please generate the private key and then encrypt an image first.")
|
193 |
+
|
194 |
+
# Define the data and files to post
|
195 |
+
data = {
|
196 |
+
"user_id": user_id,
|
197 |
+
}
|
198 |
+
|
199 |
+
files = [
|
200 |
+
("files", open(encrypted_input_path, "rb")),
|
201 |
+
("files", open(evaluation_key_path, "rb")),
|
202 |
+
]
|
203 |
+
|
204 |
+
# Send the encrypted input image and evaluation key to the server
|
205 |
+
url = SERVER_URL + "send_input"
|
206 |
+
with requests.post(
|
207 |
+
url=url,
|
208 |
+
data=data,
|
209 |
+
files=files,
|
210 |
+
) as response:
|
211 |
+
return response.ok
|
212 |
+
|
213 |
+
def run_fhe(user_id):
|
214 |
+
"""Apply the seizure detection model on the encrypted image previously sent using FHE.
|
215 |
+
|
216 |
+
Args:
|
217 |
+
user_id (int): The current user's ID.
|
218 |
+
"""
|
219 |
+
data = {
|
220 |
+
"user_id": user_id,
|
221 |
+
}
|
222 |
+
|
223 |
+
# Trigger the FHE execution on the encrypted image previously sent
|
224 |
+
url = SERVER_URL + "run_fhe"
|
225 |
+
with requests.post(
|
226 |
+
url=url,
|
227 |
+
data=data,
|
228 |
+
) as response:
|
229 |
+
if response.ok:
|
230 |
+
return response.json()
|
231 |
+
else:
|
232 |
+
raise gr.Error("Please wait for the input image to be sent to the server.")
|
233 |
+
|
234 |
+
def get_output(user_id):
|
235 |
+
"""Retrieve the encrypted output (boolean).
|
236 |
+
|
237 |
+
Args:
|
238 |
+
user_id (int): The current user's ID.
|
239 |
+
|
240 |
+
Returns:
|
241 |
+
encrypted_output_short (bytes): A representation of the encrypted result.
|
242 |
+
|
243 |
+
"""
|
244 |
+
data = {
|
245 |
+
"user_id": user_id,
|
246 |
+
}
|
247 |
+
|
248 |
+
# Retrieve the encrypted output
|
249 |
+
url = SERVER_URL + "get_output"
|
250 |
+
with requests.post(
|
251 |
+
url=url,
|
252 |
+
data=data,
|
253 |
+
) as response:
|
254 |
+
if response.ok:
|
255 |
+
encrypted_output = response.content
|
256 |
+
|
257 |
+
# Save the encrypted output to bytes in a file as it is too large to pass through regular
|
258 |
+
# Gradio buttons (see https://github.com/gradio-app/gradio/issues/1877)
|
259 |
+
encrypted_output_path = get_client_file_path("encrypted_output", user_id)
|
260 |
+
|
261 |
+
with encrypted_output_path.open("wb") as encrypted_output_file:
|
262 |
+
encrypted_output_file.write(encrypted_output)
|
263 |
+
|
264 |
+
# Create a truncated version of the encrypted output for display
|
265 |
+
encrypted_output_short = shorten_bytes_object(encrypted_output)
|
266 |
+
|
267 |
+
return encrypted_output_short
|
268 |
+
else:
|
269 |
+
raise gr.Error("Please wait for the FHE execution to be completed.")
|
270 |
+
|
271 |
+
def decrypt_output(user_id):
|
272 |
+
"""Decrypt the result.
|
273 |
+
|
274 |
+
Args:
|
275 |
+
user_id (int): The current user's ID.
|
276 |
+
|
277 |
+
Returns:
|
278 |
+
bool: The decrypted output (True if seizure detected, False otherwise)
|
279 |
+
|
280 |
+
"""
|
281 |
+
if user_id == "":
|
282 |
+
raise gr.Error("Please generate the private key first.")
|
283 |
+
|
284 |
+
# Get the encrypted output path
|
285 |
+
encrypted_output_path = get_client_file_path("encrypted_output", user_id)
|
286 |
+
|
287 |
+
if not encrypted_output_path.is_file():
|
288 |
+
raise gr.Error("Please run the FHE execution first.")
|
289 |
+
|
290 |
+
# Load the encrypted output as bytes
|
291 |
+
with encrypted_output_path.open("rb") as encrypted_output_file:
|
292 |
+
encrypted_output = encrypted_output_file.read()
|
293 |
+
|
294 |
+
# Retrieve the client API
|
295 |
+
client = get_client(user_id)
|
296 |
+
|
297 |
+
# Deserialize, decrypt and post-process the encrypted output
|
298 |
+
decrypted_output = client.deserialize_decrypt_post_process(encrypted_output)
|
299 |
+
|
300 |
+
return "Seizure detected" if decrypted_output else "No seizure detected"
|
301 |
+
|
302 |
+
def resize_img(img, width=256, height=256):
|
303 |
+
"""Resize the image."""
|
304 |
+
if img.dtype != numpy.uint8:
|
305 |
+
img = img.astype(numpy.uint8)
|
306 |
+
img_pil = Image.fromarray(img)
|
307 |
+
# Resize the image
|
308 |
+
resized_img_pil = img_pil.resize((width, height))
|
309 |
+
# Convert back to a NumPy array
|
310 |
+
return numpy.array(resized_img_pil)
|
311 |
+
|
312 |
+
demo = gr.Blocks()
|
313 |
+
|
314 |
+
print("Starting the demo...")
|
315 |
+
with demo:
|
316 |
+
gr.Markdown(
|
317 |
+
"""
|
318 |
+
<h1 align="center">Seizure Detection on Encrypted EEG Data Using Fully Homomorphic Encryption</h1>
|
319 |
+
"""
|
320 |
+
)
|
321 |
+
|
322 |
+
gr.Markdown("## Client side")
|
323 |
+
gr.Markdown("### Step 1: Upload an EEG image. ")
|
324 |
+
gr.Markdown(
|
325 |
+
f"The image will automatically be resized to shape (224x224). "
|
326 |
+
"The image here, however, is displayed in its original resolution."
|
327 |
+
)
|
328 |
+
with gr.Row():
|
329 |
+
input_image = gr.Image(
|
330 |
+
value=None, label="Upload an EEG image here.", height=256,
|
331 |
+
width=256, sources="upload", interactive=True,
|
332 |
+
)
|
333 |
+
|
334 |
+
examples = gr.Examples(
|
335 |
+
examples=EXAMPLES, inputs=[input_image], examples_per_page=5, label="Examples to use."
|
336 |
+
)
|
337 |
+
|
338 |
+
gr.Markdown("### Step 2: Generate the private key.")
|
339 |
+
keygen_button = gr.Button("Generate the private key.")
|
340 |
+
|
341 |
+
with gr.Row():
|
342 |
+
keygen_checkbox = gr.Checkbox(label="Private key generated:", interactive=False)
|
343 |
+
|
344 |
+
user_id = gr.Textbox(label="", max_lines=2, interactive=False, visible=False)
|
345 |
+
|
346 |
+
gr.Markdown("### Step 3: Encrypt the image using FHE.")
|
347 |
+
encrypt_button = gr.Button("Encrypt the image using FHE.")
|
348 |
+
|
349 |
+
with gr.Row():
|
350 |
+
encrypted_input = gr.Textbox(
|
351 |
+
label="Encrypted input representation:", max_lines=2, interactive=False
|
352 |
+
)
|
353 |
+
|
354 |
+
gr.Markdown("## Server side")
|
355 |
+
gr.Markdown(
|
356 |
+
"The encrypted value is received by the server. The server can then compute the seizure "
|
357 |
+
"detection directly over encrypted values. Once the computation is finished, the server returns "
|
358 |
+
"the encrypted results to the client."
|
359 |
+
)
|
360 |
+
|
361 |
+
gr.Markdown("### Step 4: Send the encrypted image to the server.")
|
362 |
+
send_input_button = gr.Button("Send the encrypted image to the server.")
|
363 |
+
send_input_checkbox = gr.Checkbox(label="Encrypted image sent.", interactive=False)
|
364 |
+
|
365 |
+
gr.Markdown("### Step 5: Run FHE execution.")
|
366 |
+
execute_fhe_button = gr.Button("Run FHE execution.")
|
367 |
+
fhe_execution_time = gr.Textbox(
|
368 |
+
label="Total FHE execution time (in seconds):", max_lines=1, interactive=False
|
369 |
+
)
|
370 |
+
|
371 |
+
gr.Markdown("### Step 6: Receive the encrypted output from the server.")
|
372 |
+
get_output_button = gr.Button("Receive the encrypted output from the server.")
|
373 |
+
|
374 |
+
with gr.Row():
|
375 |
+
encrypted_output = gr.Textbox(
|
376 |
+
label="Encrypted output representation:",
|
377 |
+
max_lines=2,
|
378 |
+
interactive=False
|
379 |
+
)
|
380 |
+
|
381 |
+
gr.Markdown("## Client side")
|
382 |
+
gr.Markdown(
|
383 |
+
"The encrypted output is sent back to the client, who can finally decrypt it with the "
|
384 |
+
"private key. Only the client is aware of the original image and the detection result."
|
385 |
+
)
|
386 |
+
|
387 |
+
gr.Markdown("### Step 7: Decrypt the output.")
|
388 |
+
decrypt_button = gr.Button("Decrypt the output")
|
389 |
+
|
390 |
+
with gr.Row():
|
391 |
+
decrypted_output = gr.Textbox(
|
392 |
+
label="Seizure detection result:",
|
393 |
+
interactive=False
|
394 |
+
)
|
395 |
+
|
396 |
+
# Button to generate the private key
|
397 |
+
keygen_button.click(
|
398 |
+
keygen,
|
399 |
+
outputs=[user_id, keygen_checkbox],
|
400 |
+
)
|
401 |
+
|
402 |
+
# Button to encrypt inputs on the client side
|
403 |
+
encrypt_button.click(
|
404 |
+
encrypt,
|
405 |
+
inputs=[user_id, input_image],
|
406 |
+
outputs=[input_image, encrypted_input],
|
407 |
+
)
|
408 |
+
|
409 |
+
# Button to send the encodings to the server using post method
|
410 |
+
send_input_button.click(
|
411 |
+
send_input, inputs=[user_id], outputs=[send_input_checkbox]
|
412 |
+
)
|
413 |
+
|
414 |
+
# Button to send the encodings to the server using post method
|
415 |
+
execute_fhe_button.click(run_fhe, inputs=[user_id], outputs=[fhe_execution_time])
|
416 |
+
|
417 |
+
# Button to send the encodings to the server using post method
|
418 |
+
get_output_button.click(
|
419 |
+
get_output,
|
420 |
+
inputs=[user_id],
|
421 |
+
outputs=[encrypted_output]
|
422 |
+
)
|
423 |
+
|
424 |
+
# Button to decrypt the output on the client side
|
425 |
+
decrypt_button.click(
|
426 |
+
decrypt_output,
|
427 |
+
inputs=[user_id],
|
428 |
+
outputs=[decrypted_output],
|
429 |
+
)
|
430 |
+
|
431 |
+
gr.Markdown(
|
432 |
+
"The app was built with [Concrete-ML](https://github.com/zama-ai/concrete-ml), a "
|
433 |
+
"Privacy-Preserving Machine Learning (PPML) open-source set of tools by [Zama](https://zama.ai/). "
|
434 |
+
"Try it yourself and don't forget to star on Github ⭐."
|
435 |
+
)
|
436 |
+
|
437 |
+
demo.launch(share=False)
|
client_server_interface.py
ADDED
@@ -0,0 +1,150 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"Client-server interface custom implementation for seizure detection models."
|
2 |
+
|
3 |
+
from concrete import fhe
|
4 |
+
|
5 |
+
from seizure_detection import SeizureDetector
|
6 |
+
|
7 |
+
|
8 |
+
class FHEServer:
|
9 |
+
"""Server interface to run a FHE circuit for seizure detection."""
|
10 |
+
|
11 |
+
def __init__(self, model_path):
|
12 |
+
"""Initialize the FHE interface.
|
13 |
+
|
14 |
+
Args:
|
15 |
+
model_path (Path): The path to the directory where the circuit is saved.
|
16 |
+
"""
|
17 |
+
self.model_path = model_path
|
18 |
+
|
19 |
+
# Load the FHE circuit
|
20 |
+
self.server = fhe.Server.load(self.model_path / "server.zip")
|
21 |
+
|
22 |
+
def run(self, serialized_encrypted_image, serialized_evaluation_keys):
|
23 |
+
"""Run seizure detection on the server over an encrypted image.
|
24 |
+
|
25 |
+
Args:
|
26 |
+
serialized_encrypted_image (bytes): The encrypted and serialized image.
|
27 |
+
serialized_evaluation_keys (bytes): The serialized evaluation keys.
|
28 |
+
|
29 |
+
Returns:
|
30 |
+
bytes: The encrypted boolean output indicating seizure detection.
|
31 |
+
"""
|
32 |
+
# Deserialize the encrypted input image and the evaluation keys
|
33 |
+
encrypted_image = fhe.Value.deserialize(serialized_encrypted_image)
|
34 |
+
evaluation_keys = fhe.EvaluationKeys.deserialize(serialized_evaluation_keys)
|
35 |
+
|
36 |
+
# Execute the seizure detection in FHE
|
37 |
+
encrypted_output = self.server.run(encrypted_image, evaluation_keys=evaluation_keys)
|
38 |
+
|
39 |
+
# Serialize the encrypted output
|
40 |
+
serialized_encrypted_output = encrypted_output.serialize()
|
41 |
+
|
42 |
+
return serialized_encrypted_output
|
43 |
+
|
44 |
+
|
45 |
+
class FHEDev:
|
46 |
+
"""Development interface to save and load the seizure detection model."""
|
47 |
+
|
48 |
+
def __init__(self, seizure_detector, model_path):
|
49 |
+
"""Initialize the FHE interface.
|
50 |
+
|
51 |
+
Args:
|
52 |
+
seizure_detector (SeizureDetector): The seizure detection model to use in the FHE interface.
|
53 |
+
model_path (str): The path to the directory where the circuit is saved.
|
54 |
+
"""
|
55 |
+
|
56 |
+
self.seizure_detector = seizure_detector
|
57 |
+
self.model_path = model_path
|
58 |
+
|
59 |
+
self.model_path.mkdir(parents=True, exist_ok=True)
|
60 |
+
|
61 |
+
def save(self):
|
62 |
+
"""Export all needed artifacts for the client and server interfaces."""
|
63 |
+
|
64 |
+
assert self.seizure_detector.fhe_circuit is not None, (
|
65 |
+
"The model must be compiled before saving it."
|
66 |
+
)
|
67 |
+
|
68 |
+
# Save the circuit for the server, using the via_mlir in order to handle cross-platform
|
69 |
+
# execution
|
70 |
+
path_circuit_server = self.model_path / "server.zip"
|
71 |
+
self.seizure_detector.fhe_circuit.server.save(path_circuit_server, via_mlir=True)
|
72 |
+
|
73 |
+
# Save the circuit for the client
|
74 |
+
path_circuit_client = self.model_path / "client.zip"
|
75 |
+
self.seizure_detector.fhe_circuit.client.save(path_circuit_client)
|
76 |
+
|
77 |
+
|
78 |
+
class FHEClient:
|
79 |
+
"""Client interface to encrypt and decrypt FHE data associated to a SeizureDetector."""
|
80 |
+
|
81 |
+
def __init__(self, model_path, key_dir=None):
|
82 |
+
"""Initialize the FHE interface.
|
83 |
+
|
84 |
+
Args:
|
85 |
+
model_path (Path): The path to the directory where the circuit is saved.
|
86 |
+
key_dir (Path): The path to the directory where the keys are stored. Default to None.
|
87 |
+
"""
|
88 |
+
self.model_path = model_path
|
89 |
+
self.key_dir = key_dir
|
90 |
+
|
91 |
+
# If model_path does not exist raise
|
92 |
+
assert model_path.exists(), f"{model_path} does not exist. Please specify a valid path."
|
93 |
+
|
94 |
+
# Load the client
|
95 |
+
self.client = fhe.Client.load(self.model_path / "client.zip", self.key_dir)
|
96 |
+
|
97 |
+
# Instantiate the seizure detector
|
98 |
+
self.seizure_detector = SeizureDetector()
|
99 |
+
|
100 |
+
def generate_private_and_evaluation_keys(self, force=False):
|
101 |
+
"""Generate the private and evaluation keys.
|
102 |
+
|
103 |
+
Args:
|
104 |
+
force (bool): If True, regenerate the keys even if they already exist.
|
105 |
+
"""
|
106 |
+
self.client.keygen(force)
|
107 |
+
|
108 |
+
def get_serialized_evaluation_keys(self):
|
109 |
+
"""Get the serialized evaluation keys.
|
110 |
+
|
111 |
+
Returns:
|
112 |
+
bytes: The evaluation keys.
|
113 |
+
"""
|
114 |
+
return self.client.evaluation_keys.serialize()
|
115 |
+
|
116 |
+
def encrypt_serialize(self, input_image):
|
117 |
+
"""Encrypt and serialize the input image in the clear.
|
118 |
+
|
119 |
+
Args:
|
120 |
+
input_image (numpy.ndarray): The image to encrypt and serialize.
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
bytes: The pre-processed, encrypted and serialized image.
|
124 |
+
"""
|
125 |
+
# Encrypt the image
|
126 |
+
encrypted_image = self.client.encrypt(input_image)
|
127 |
+
|
128 |
+
# Serialize the encrypted image to be sent to the server
|
129 |
+
serialized_encrypted_image = encrypted_image.serialize()
|
130 |
+
return serialized_encrypted_image
|
131 |
+
|
132 |
+
def deserialize_decrypt_post_process(self, serialized_encrypted_output):
|
133 |
+
"""Deserialize, decrypt and post-process the output in the clear.
|
134 |
+
|
135 |
+
Args:
|
136 |
+
serialized_encrypted_output (bytes): The serialized and encrypted output.
|
137 |
+
|
138 |
+
Returns:
|
139 |
+
bool: The decrypted and deserialized boolean indicating seizure detection.
|
140 |
+
"""
|
141 |
+
# Deserialize the encrypted output
|
142 |
+
encrypted_output = fhe.Value.deserialize(serialized_encrypted_output)
|
143 |
+
|
144 |
+
# Decrypt the output
|
145 |
+
output = self.client.decrypt(encrypted_output)
|
146 |
+
|
147 |
+
# Post-process the output (if needed)
|
148 |
+
seizure_detected = self.seizure_detector.post_processing(output)
|
149 |
+
|
150 |
+
return seizure_detected
|
common.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"All the constants used in this repo."
|
2 |
+
|
3 |
+
from pathlib import Path
|
4 |
+
|
5 |
+
# This repository's directory
|
6 |
+
REPO_DIR = Path(__file__).parent
|
7 |
+
|
8 |
+
# This repository's main necessary folders
|
9 |
+
FILTERS_PATH = REPO_DIR / "filters"
|
10 |
+
KEYS_PATH = REPO_DIR / ".fhe_keys"
|
11 |
+
CLIENT_TMP_PATH = REPO_DIR / "client_tmp"
|
12 |
+
SERVER_TMP_PATH = REPO_DIR / "server_tmp"
|
13 |
+
|
14 |
+
# Create the necessary folders
|
15 |
+
KEYS_PATH.mkdir(exist_ok=True)
|
16 |
+
CLIENT_TMP_PATH.mkdir(exist_ok=True)
|
17 |
+
SERVER_TMP_PATH.mkdir(exist_ok=True)
|
18 |
+
|
19 |
+
# The input images' shape. Images with different input shapes will be cropped and resized by Gradio
|
20 |
+
INPUT_SHAPE = (224, 224)
|
21 |
+
|
22 |
+
# Retrieve the input examples directory
|
23 |
+
INPUT_EXAMPLES_DIR = REPO_DIR / "input_examples"
|
24 |
+
|
25 |
+
# List of all image examples suggested in the demo
|
26 |
+
EXAMPLES = [str(image) for image in INPUT_EXAMPLES_DIR.glob("**/*")]
|
27 |
+
|
28 |
+
# Store the server's URL
|
29 |
+
SERVER_URL = "http://localhost:8000/"
|
input_examples/eeg-1.png
ADDED
input_examples/eeg-2.png
ADDED
requirements.txt
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
concrete-ml==1.1.0
|
2 |
+
gradio
|
seizure_detection.py
ADDED
File without changes
|
server.py
ADDED
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""Server that will listen for GET and POST requests from the client."""
|
2 |
+
|
3 |
+
import time
|
4 |
+
from typing import List
|
5 |
+
from fastapi import FastAPI, File, Form, UploadFile
|
6 |
+
from fastapi.responses import JSONResponse, Response
|
7 |
+
|
8 |
+
from common import SERVER_TMP_PATH
|
9 |
+
from client_server_interface import FHEServer
|
10 |
+
|
11 |
+
# Load the server object for seizure detection
|
12 |
+
FHE_SERVER = FHEServer(model_path="path/to/seizure_detection_model")
|
13 |
+
|
14 |
+
def get_server_file_path(name, user_id):
|
15 |
+
"""Get the correct temporary file path for the server.
|
16 |
+
|
17 |
+
Args:
|
18 |
+
name (str): The desired file name.
|
19 |
+
user_id (int): The current user's ID.
|
20 |
+
|
21 |
+
Returns:
|
22 |
+
pathlib.Path: The file path.
|
23 |
+
"""
|
24 |
+
return SERVER_TMP_PATH / f"{name}_seizure_detection_{user_id}"
|
25 |
+
|
26 |
+
|
27 |
+
# Initialize an instance of FastAPI
|
28 |
+
app = FastAPI()
|
29 |
+
|
30 |
+
# Define the default route
|
31 |
+
@app.get("/")
|
32 |
+
def root():
|
33 |
+
return {"message": "Welcome to Your Seizure Detection FHE Server!"}
|
34 |
+
|
35 |
+
|
36 |
+
@app.post("/send_input")
|
37 |
+
def send_input(
|
38 |
+
user_id: str = Form(),
|
39 |
+
files: List[UploadFile] = File(),
|
40 |
+
):
|
41 |
+
"""Send the inputs to the server."""
|
42 |
+
# Retrieve the encrypted input image and the evaluation key paths
|
43 |
+
encrypted_image_path = get_server_file_path("encrypted_image", user_id)
|
44 |
+
evaluation_key_path = get_server_file_path("evaluation_key", user_id)
|
45 |
+
|
46 |
+
# Write the files using the above paths
|
47 |
+
with encrypted_image_path.open("wb") as encrypted_image, evaluation_key_path.open(
|
48 |
+
"wb"
|
49 |
+
) as evaluation_key:
|
50 |
+
encrypted_image.write(files[0].file.read())
|
51 |
+
evaluation_key.write(files[1].file.read())
|
52 |
+
|
53 |
+
|
54 |
+
@app.post("/run_fhe")
|
55 |
+
def run_fhe(
|
56 |
+
user_id: str = Form(),
|
57 |
+
):
|
58 |
+
"""Execute seizure detection on the encrypted input image using FHE."""
|
59 |
+
# Retrieve the encrypted input image and the evaluation key paths
|
60 |
+
encrypted_image_path = get_server_file_path("encrypted_image", user_id)
|
61 |
+
evaluation_key_path = get_server_file_path("evaluation_key", user_id)
|
62 |
+
|
63 |
+
# Read the files using the above paths
|
64 |
+
with encrypted_image_path.open("rb") as encrypted_image_file, evaluation_key_path.open(
|
65 |
+
"rb"
|
66 |
+
) as evaluation_key_file:
|
67 |
+
encrypted_image = encrypted_image_file.read()
|
68 |
+
evaluation_key = evaluation_key_file.read()
|
69 |
+
|
70 |
+
# Run the FHE execution
|
71 |
+
start = time.time()
|
72 |
+
encrypted_output = FHE_SERVER.run(encrypted_image, evaluation_key)
|
73 |
+
fhe_execution_time = round(time.time() - start, 2)
|
74 |
+
|
75 |
+
# Retrieve the encrypted output path
|
76 |
+
encrypted_output_path = get_server_file_path("encrypted_output", user_id)
|
77 |
+
|
78 |
+
# Write the file using the above path
|
79 |
+
with encrypted_output_path.open("wb") as encrypted_output_file:
|
80 |
+
encrypted_output_file.write(encrypted_output)
|
81 |
+
|
82 |
+
return JSONResponse(content=fhe_execution_time)
|
83 |
+
|
84 |
+
|
85 |
+
@app.post("/get_output")
|
86 |
+
def get_output(
|
87 |
+
user_id: str = Form(),
|
88 |
+
):
|
89 |
+
"""Retrieve the encrypted output."""
|
90 |
+
# Retrieve the encrypted output path
|
91 |
+
encrypted_output_path = get_server_file_path("encrypted_output", user_id)
|
92 |
+
|
93 |
+
# Read the file using the above path
|
94 |
+
with encrypted_output_path.open("rb") as encrypted_output_file:
|
95 |
+
encrypted_output = encrypted_output_file.read()
|
96 |
+
|
97 |
+
return Response(encrypted_output)
|