Spaces:
Running
on
CPU Upgrade
Running
on
CPU Upgrade
Upload app.py
Browse files
app.py
CHANGED
@@ -1,14 +1,198 @@
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
|
4 |
def update(name):
|
5 |
return f"Welcome to Gradio, {name}!"
|
6 |
|
7 |
|
8 |
-
|
9 |
-
gr.
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
+
import time
|
3 |
+
import os
|
4 |
+
from pathlib import Path
|
5 |
+
|
6 |
+
import subprocess
|
7 |
+
from concrete.ml.deployment import FHEModelClient
|
8 |
+
from requests import head
|
9 |
+
import numpy
|
10 |
+
import os
|
11 |
+
from pathlib import Path
|
12 |
+
import requests
|
13 |
+
import json
|
14 |
+
import base64
|
15 |
+
import subprocess
|
16 |
+
import shutil
|
17 |
+
import time
|
18 |
+
import pandas as pd
|
19 |
+
import pickle
|
20 |
+
import numpy as np
|
21 |
+
|
22 |
+
# This repository's directory
|
23 |
+
REPO_DIR = Path(__file__).parent
|
24 |
+
subprocess.Popen(["uvicorn", "server:app"], cwd=REPO_DIR)
|
25 |
+
|
26 |
+
# if not exists, create a directory for the FHE keys called .fhe_keys
|
27 |
+
if not os.path.exists(".fhe_keys"):
|
28 |
+
os.mkdir(".fhe_keys")
|
29 |
+
# if not exists, create a directory for the tmp files called tmp
|
30 |
+
if not os.path.exists("tmp"):
|
31 |
+
os.mkdir("tmp")
|
32 |
+
|
33 |
+
|
34 |
+
# Wait 4 sec for the server to start
|
35 |
+
time.sleep(4)
|
36 |
+
|
37 |
+
|
38 |
+
# Encrypted data limit for the browser to display
|
39 |
+
# (encrypted data is too large to display in the browser)
|
40 |
+
ENCRYPTED_DATA_BROWSER_LIMIT = 500
|
41 |
+
N_USER_KEY_STORED = 20
|
42 |
+
|
43 |
+
|
44 |
+
def clean_tmp_directory():
|
45 |
+
# Allow 20 user keys to be stored.
|
46 |
+
# Once that limitation is reached, deleted the oldest.
|
47 |
+
path_sub_directories = sorted(
|
48 |
+
[f for f in Path(".fhe_keys/").iterdir() if f.is_dir()], key=os.path.getmtime
|
49 |
+
)
|
50 |
+
|
51 |
+
user_ids = []
|
52 |
+
if len(path_sub_directories) > N_USER_KEY_STORED:
|
53 |
+
n_files_to_delete = len(path_sub_directories) - N_USER_KEY_STORED
|
54 |
+
for p in path_sub_directories[:n_files_to_delete]:
|
55 |
+
user_ids.append(p.name)
|
56 |
+
shutil.rmtree(p)
|
57 |
+
|
58 |
+
list_files_tmp = Path("tmp/").iterdir()
|
59 |
+
# Delete all files related to user_id
|
60 |
+
for file in list_files_tmp:
|
61 |
+
for user_id in user_ids:
|
62 |
+
if file.name.endswith(f"{user_id}.npy"):
|
63 |
+
file.unlink()
|
64 |
+
|
65 |
+
|
66 |
+
def keygen():
|
67 |
+
# Clean tmp directory if needed
|
68 |
+
clean_tmp_directory()
|
69 |
+
|
70 |
+
print("Initializing FHEModelClient...")
|
71 |
+
# Let's create a user_id
|
72 |
+
user_id = numpy.random.randint(0, 2**32)
|
73 |
+
fhe_api = FHEModelClient(f"deployment/deployment_{task}", f".fhe_keys/{user_id}")
|
74 |
+
fhe_api.load()
|
75 |
+
|
76 |
+
# Generate a fresh key
|
77 |
+
fhe_api.generate_private_and_evaluation_keys(force=True)
|
78 |
+
evaluation_key = fhe_api.get_serialized_evaluation_keys()
|
79 |
+
|
80 |
+
numpy.save(f"tmp/tmp_evaluation_key_{user_id}.npy", evaluation_key)
|
81 |
+
|
82 |
+
return [list(evaluation_key)[:ENCRYPTED_DATA_BROWSER_LIMIT], user_id]
|
83 |
+
|
84 |
+
|
85 |
+
def encode_quantize_encrypt(test_file, user_id):
|
86 |
+
|
87 |
+
fhe_api = FHEModelClient(f"fhe_model", f".fhe_keys/{user_id}")
|
88 |
+
fhe_api.load()
|
89 |
+
from PE_main import extract_infos
|
90 |
+
|
91 |
+
features = pickle.loads(open(os.path.join("features.pkl"), "rb").read())
|
92 |
+
encodings = extract_infos(test_file)
|
93 |
+
encodings = list(map(lambda x: encodings[x], features))
|
94 |
+
|
95 |
+
quantized_encodings = fhe_api.model.quantize_input(encodings).astype(numpy.uint8)
|
96 |
+
encrypted_quantized_encoding = fhe_api.quantize_encrypt_serialize(encodings)
|
97 |
+
|
98 |
+
# Save encrypted_quantized_encoding in a file, since too large to pass through regular Gradio
|
99 |
+
# buttons, https://github.com/gradio-app/gradio/issues/1877
|
100 |
+
numpy.save(
|
101 |
+
f"tmp/tmp_encrypted_quantized_encoding_{user_id}.npy",
|
102 |
+
encrypted_quantized_encoding,
|
103 |
+
)
|
104 |
+
|
105 |
+
# Compute size
|
106 |
+
encrypted_quantized_encoding_shorten = list(encrypted_quantized_encoding)[
|
107 |
+
:ENCRYPTED_DATA_BROWSER_LIMIT
|
108 |
+
]
|
109 |
+
encrypted_quantized_encoding_shorten_hex = "".join(
|
110 |
+
f"{i:02x}" for i in encrypted_quantized_encoding_shorten
|
111 |
+
)
|
112 |
+
return (
|
113 |
+
encodings[0],
|
114 |
+
quantized_encodings[0],
|
115 |
+
encrypted_quantized_encoding_shorten_hex,
|
116 |
+
)
|
117 |
+
|
118 |
+
|
119 |
+
def run_fhe(user_id):
|
120 |
+
encoded_data_path = Path(f"tmp/tmp_encrypted_quantized_encoding_{user_id}.npy")
|
121 |
+
encrypted_quantized_encoding = numpy.load(encoded_data_path)
|
122 |
+
|
123 |
+
# Read evaluation_key from the file
|
124 |
+
evaluation_key = numpy.load(f"tmp/tmp_evaluation_key_{user_id}.npy")
|
125 |
+
|
126 |
+
# Use base64 to encode the encodings and evaluation key
|
127 |
+
encrypted_quantized_encoding = base64.b64encode(
|
128 |
+
encrypted_quantized_encoding
|
129 |
+
).decode()
|
130 |
+
encoded_evaluation_key = base64.b64encode(evaluation_key).decode()
|
131 |
+
|
132 |
+
query = {}
|
133 |
+
query["evaluation_key"] = encoded_evaluation_key
|
134 |
+
query["encrypted_encoding"] = encrypted_quantized_encoding
|
135 |
+
headers = {"Content-type": "application/json"}
|
136 |
+
|
137 |
+
|
138 |
+
response = requests.post(
|
139 |
+
"http://localhost:8000/predict",
|
140 |
+
data=json.dumps(query),
|
141 |
+
headers=headers,
|
142 |
+
)
|
143 |
+
|
144 |
+
|
145 |
+
encrypted_prediction = base64.b64decode(response.json()["encrypted_prediction"])
|
146 |
+
|
147 |
+
numpy.save(f"tmp/tmp_encrypted_prediction_{user_id}.npy", encrypted_prediction)
|
148 |
+
encrypted_prediction_shorten = list(encrypted_prediction)[
|
149 |
+
:ENCRYPTED_DATA_BROWSER_LIMIT
|
150 |
+
]
|
151 |
+
encrypted_prediction_shorten_hex = "".join(
|
152 |
+
f"{i:02x}" for i in encrypted_prediction_shorten
|
153 |
+
)
|
154 |
+
|
155 |
+
|
156 |
+
def decrypt_prediction(user_id):
|
157 |
+
encoded_data_path = Path(f"tmp/tmp_encrypted_prediction_{user_id}.npy")
|
158 |
+
|
159 |
+
# Read encrypted_prediction from the file
|
160 |
+
|
161 |
+
encrypted_prediction = numpy.load(encoded_data_path).tobytes()
|
162 |
+
|
163 |
+
fhe_api = FHEModelClient(f"fhe_model", f".fhe_keys/{user_id}")
|
164 |
+
fhe_api.load()
|
165 |
+
|
166 |
+
# We need to retrieve the private key that matches the client specs (see issue #18)
|
167 |
+
fhe_api.generate_private_and_evaluation_keys(force=False)
|
168 |
+
|
169 |
+
predictions = fhe_api.deserialize_decrypt_dequantize(encrypted_prediction)
|
170 |
|
171 |
|
172 |
def update(name):
|
173 |
return f"Welcome to Gradio, {name}!"
|
174 |
|
175 |
|
176 |
+
if __name__ == "__main__":
|
177 |
+
app = gr.Interface(
|
178 |
+
[
|
179 |
+
keygen,
|
180 |
+
encode_quantize_encrypt,
|
181 |
+
run_fhe,
|
182 |
+
decrypt_prediction,
|
183 |
+
],
|
184 |
+
[
|
185 |
+
gr.inputs.Textbox(label="Task", default="malware"),
|
186 |
+
gr.inputs.File(label="Test File"),
|
187 |
+
gr.inputs.Textbox(label="User ID"),
|
188 |
+
],
|
189 |
+
[
|
190 |
+
gr.outputs.Textbox(label="Evaluation Key"),
|
191 |
+
gr.outputs.Textbox(label="Encodings"),
|
192 |
+
gr.outputs.Textbox(label="Encrypted Quantized Encoding"),
|
193 |
+
gr.outputs.Textbox(label="Encrypted Prediction"),
|
194 |
+
],
|
195 |
+
title="FHE Model",
|
196 |
+
description="This is a FHE Model",
|
197 |
+
)
|
198 |
+
app.launch()
|