Spaces:
Sleeping
Sleeping
Add statistical analysis
Browse files- analyse.py +408 -0
- api.py +3 -0
- config.ini +1 -0
- model_factory.py +42 -1
- requirements.txt +2 -0
- schemes.py +5 -0
- stegno.py +22 -11
- utils.py +1 -0
analyse.py
ADDED
@@ -0,0 +1,408 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import json
|
3 |
+
import base64
|
4 |
+
from argparse import ArgumentParser
|
5 |
+
|
6 |
+
import numpy as np
|
7 |
+
from matplotlib import pyplot as plt
|
8 |
+
|
9 |
+
import torch
|
10 |
+
from datasets import load_dataset
|
11 |
+
from model_factory import ModelFactory
|
12 |
+
from stegno import generate
|
13 |
+
|
14 |
+
rng = torch.Generator(device="cpu")
|
15 |
+
rng.manual_seed(0)
|
16 |
+
|
17 |
+
|
18 |
+
def load_msgs(msg_lens: list[int], file: str | None = None):
|
19 |
+
msgs = None
|
20 |
+
if file is not None and os.path.isfile(file):
|
21 |
+
with open(file, "r") as f:
|
22 |
+
msgs = json.load(f)
|
23 |
+
if "readable" not in msgs and "random" not in msgs:
|
24 |
+
msgs = None
|
25 |
+
else:
|
26 |
+
return msgs
|
27 |
+
|
28 |
+
msgs = {
|
29 |
+
"readable": [],
|
30 |
+
"random": [],
|
31 |
+
}
|
32 |
+
|
33 |
+
c4_en = load_dataset("allenai/c4", "en", split="validation", streaming=True)
|
34 |
+
iterator = iter(c4_en)
|
35 |
+
|
36 |
+
for length in msg_lens:
|
37 |
+
random_msg = torch.randint(256, (length,), generator=rng)
|
38 |
+
base64_msg = base64.b64encode(bytes(random_msg.tolist())).decode(
|
39 |
+
"ascii"
|
40 |
+
)
|
41 |
+
msgs["random"].append(base64_msg)
|
42 |
+
|
43 |
+
readable_msg = next(iterator)["text"]
|
44 |
+
while len(readable_msg) < length:
|
45 |
+
readable_msg = next(iterator)["text"]
|
46 |
+
msgs["readable"].append(readable_msg[:length])
|
47 |
+
|
48 |
+
return msgs
|
49 |
+
|
50 |
+
|
51 |
+
def load_prompts(n: int, min_length: int, file: str | None = None):
|
52 |
+
prompts = None
|
53 |
+
if file is not None and os.path.isfile(file):
|
54 |
+
with open(file, "r") as f:
|
55 |
+
prompts = json.load(f)
|
56 |
+
return prompts
|
57 |
+
|
58 |
+
prompts = []
|
59 |
+
|
60 |
+
c4_en = load_dataset("allenai/c4", "en", split="train", streaming=True)
|
61 |
+
iterator = iter(c4_en)
|
62 |
+
|
63 |
+
while len(prompts) < n:
|
64 |
+
text = next(iterator)["text"]
|
65 |
+
if len(text) < min_length:
|
66 |
+
continue
|
67 |
+
prompts.append(text)
|
68 |
+
|
69 |
+
return prompts
|
70 |
+
|
71 |
+
|
72 |
+
def create_args():
|
73 |
+
parser = ArgumentParser()
|
74 |
+
|
75 |
+
# messages
|
76 |
+
parser.add_argument(
|
77 |
+
"--msgs-file", type=str, default=None, help="Where messages are stored"
|
78 |
+
)
|
79 |
+
parser.add_argument(
|
80 |
+
"--msgs-lengths",
|
81 |
+
nargs=3,
|
82 |
+
type=int,
|
83 |
+
help="Range of messages' lengths. This is parsed in form: <start> <end> <step>",
|
84 |
+
)
|
85 |
+
parser.add_argument(
|
86 |
+
"--msgs-per-length",
|
87 |
+
type=int,
|
88 |
+
default=5,
|
89 |
+
help="Number of messages per length",
|
90 |
+
)
|
91 |
+
# prompts
|
92 |
+
parser.add_argument(
|
93 |
+
"--prompts-file",
|
94 |
+
type=str,
|
95 |
+
default=None,
|
96 |
+
help="Where prompts are stored",
|
97 |
+
)
|
98 |
+
parser.add_argument(
|
99 |
+
"--num-prompts",
|
100 |
+
type=int,
|
101 |
+
default=500,
|
102 |
+
help="Number of prompts",
|
103 |
+
)
|
104 |
+
parser.add_argument(
|
105 |
+
"--prompt-size",
|
106 |
+
type=int,
|
107 |
+
default=50,
|
108 |
+
help="Size of prompts",
|
109 |
+
)
|
110 |
+
parser.add_argument(
|
111 |
+
"--prompts-min-length",
|
112 |
+
type=int,
|
113 |
+
default=100,
|
114 |
+
help="Min length of prompts",
|
115 |
+
)
|
116 |
+
# Others
|
117 |
+
parser.add_argument(
|
118 |
+
"--overwrite",
|
119 |
+
action="store_true",
|
120 |
+
help="Whether to overwrite prompts and messages files",
|
121 |
+
)
|
122 |
+
|
123 |
+
# Hyperparameters
|
124 |
+
parser.add_argument(
|
125 |
+
"--gen-model",
|
126 |
+
type=str,
|
127 |
+
default="gpt2",
|
128 |
+
help="Model used to generate",
|
129 |
+
)
|
130 |
+
parser.add_argument(
|
131 |
+
"--deltas",
|
132 |
+
nargs=3,
|
133 |
+
type=float,
|
134 |
+
help="Range of delta. This is parsed in form: <start> <end> <step>",
|
135 |
+
)
|
136 |
+
parser.add_argument(
|
137 |
+
"--bases",
|
138 |
+
nargs=3,
|
139 |
+
type=int,
|
140 |
+
help="Range of base. This is parsed in form: <start> <end> <step>",
|
141 |
+
)
|
142 |
+
parser.add_argument(
|
143 |
+
"--judge-model",
|
144 |
+
type=str,
|
145 |
+
default="gpt2",
|
146 |
+
help="Model used to compute score perplexity of generated text",
|
147 |
+
)
|
148 |
+
# Results
|
149 |
+
parser.add_argument(
|
150 |
+
"--repeat",
|
151 |
+
type=int,
|
152 |
+
default=1,
|
153 |
+
help="How many times to repeat for each set of parameters, prompts and messages",
|
154 |
+
)
|
155 |
+
parser.add_argument(
|
156 |
+
"--results-load-file",
|
157 |
+
type=str,
|
158 |
+
default=None,
|
159 |
+
help="Where to load results",
|
160 |
+
)
|
161 |
+
parser.add_argument(
|
162 |
+
"--results-save-file",
|
163 |
+
type=str,
|
164 |
+
default=None,
|
165 |
+
help="Where to save results",
|
166 |
+
)
|
167 |
+
parser.add_argument(
|
168 |
+
"--figs-dir",
|
169 |
+
type=str,
|
170 |
+
default=None,
|
171 |
+
help="Where to save figures",
|
172 |
+
)
|
173 |
+
|
174 |
+
return parser.parse_args()
|
175 |
+
|
176 |
+
|
177 |
+
def get_results(args, prompts, msgs):
|
178 |
+
model, tokenizer = ModelFactory.load_model(args.gen_model)
|
179 |
+
results = []
|
180 |
+
|
181 |
+
for prompt in prompts[:1]:
|
182 |
+
for delta in np.arange(
|
183 |
+
args.deltas[0], args.deltas[1] + args.deltas[2], args.deltas[2]
|
184 |
+
):
|
185 |
+
for base in np.arange(
|
186 |
+
args.bases[0],
|
187 |
+
args.bases[1] + args.bases[2],
|
188 |
+
args.bases[2],
|
189 |
+
dtype=np.int32,
|
190 |
+
):
|
191 |
+
for k in msgs:
|
192 |
+
msg_type = k
|
193 |
+
for msg in msgs[k]:
|
194 |
+
msg_bytes = (
|
195 |
+
msg.encode("ascii")
|
196 |
+
if k == "readable"
|
197 |
+
else base64.b64decode(msg)
|
198 |
+
)
|
199 |
+
for _ in range(args.repeat):
|
200 |
+
text, msg_rate, tokens_info = generate(
|
201 |
+
tokenizer=tokenizer,
|
202 |
+
model=model,
|
203 |
+
prompt=prompt,
|
204 |
+
msg=msg_bytes,
|
205 |
+
start_pos_p=[0],
|
206 |
+
delta=delta,
|
207 |
+
msg_base=base,
|
208 |
+
seed_scheme="sha_left_hash",
|
209 |
+
window_length=1,
|
210 |
+
private_key=0,
|
211 |
+
min_new_tokens_ratio=1,
|
212 |
+
max_new_tokens_ratio=2,
|
213 |
+
num_beams=4,
|
214 |
+
repetition_penalty=1.5,
|
215 |
+
prompt_size=args.prompt_size,
|
216 |
+
)
|
217 |
+
results.append(
|
218 |
+
{
|
219 |
+
"msg_type": msg_type,
|
220 |
+
"delta": delta.item(),
|
221 |
+
"base": base.item(),
|
222 |
+
"perplexity": ModelFactory.compute_perplexity(
|
223 |
+
args.judge_model, text
|
224 |
+
),
|
225 |
+
"msg_rate": msg_rate,
|
226 |
+
}
|
227 |
+
)
|
228 |
+
return results
|
229 |
+
|
230 |
+
|
231 |
+
def process_results(results, save_dir):
|
232 |
+
data = {
|
233 |
+
"perplexities": {
|
234 |
+
"random": {},
|
235 |
+
"readable": {},
|
236 |
+
},
|
237 |
+
"msg_rates": {
|
238 |
+
"random": {},
|
239 |
+
"readable": {},
|
240 |
+
},
|
241 |
+
}
|
242 |
+
for r in results:
|
243 |
+
msg_type = r["msg_type"]
|
244 |
+
base = r["base"]
|
245 |
+
delta = r["delta"]
|
246 |
+
msg_rate = r["msg_rate"]
|
247 |
+
perplexity = r["perplexity"]
|
248 |
+
|
249 |
+
if (base, delta) not in data["msg_rates"][msg_type]:
|
250 |
+
data["msg_rates"][msg_type][(base, delta)] = []
|
251 |
+
data["msg_rates"][msg_type][(base, delta)].append(msg_rate)
|
252 |
+
|
253 |
+
if (base, delta) not in data["perplexities"][msg_type]:
|
254 |
+
data["perplexities"][msg_type][(base, delta)] = []
|
255 |
+
data["perplexities"][msg_type][(base, delta)].append(perplexity)
|
256 |
+
|
257 |
+
bases = {
|
258 |
+
"perplexities": {
|
259 |
+
"random": [],
|
260 |
+
"readable": [],
|
261 |
+
},
|
262 |
+
"msg_rates": {
|
263 |
+
"random": [],
|
264 |
+
"readable": [],
|
265 |
+
},
|
266 |
+
}
|
267 |
+
deltas = {
|
268 |
+
"perplexities": {
|
269 |
+
"random": [],
|
270 |
+
"readable": [],
|
271 |
+
},
|
272 |
+
"msg_rates": {
|
273 |
+
"random": [],
|
274 |
+
"readable": [],
|
275 |
+
},
|
276 |
+
}
|
277 |
+
values = {
|
278 |
+
"perplexities": {
|
279 |
+
"random": [],
|
280 |
+
"readable": [],
|
281 |
+
},
|
282 |
+
"msg_rates": {
|
283 |
+
"random": [],
|
284 |
+
"readable": [],
|
285 |
+
},
|
286 |
+
}
|
287 |
+
base_set = set()
|
288 |
+
delta_set = set()
|
289 |
+
for metric in data:
|
290 |
+
for msg_type in data[metric]:
|
291 |
+
for k in data[metric][msg_type]:
|
292 |
+
s = sum(data[metric][msg_type][k])
|
293 |
+
cnt = len(data[metric][msg_type][k])
|
294 |
+
data[metric][msg_type][k] = s / cnt
|
295 |
+
|
296 |
+
bases[metric][msg_type].append(k[0])
|
297 |
+
deltas[metric][msg_type].append(k[1])
|
298 |
+
values[metric][msg_type].append(s / cnt)
|
299 |
+
base_set.add(k[0])
|
300 |
+
delta_set.add(k[1])
|
301 |
+
for metric in data:
|
302 |
+
for msg_type in data[metric]:
|
303 |
+
bases[metric][msg_type] = np.array(bases[metric][msg_type], dtype=np.int32)
|
304 |
+
deltas[metric][msg_type] = np.array(deltas[metric][msg_type], dtype=np.int32)
|
305 |
+
values[metric][msg_type] = np.array(values[metric][msg_type], dtype=np.float32)
|
306 |
+
|
307 |
+
os.makedirs(save_dir, exist_ok=True)
|
308 |
+
for metric in data:
|
309 |
+
for msg_type in data[metric]:
|
310 |
+
fig = plt.figure(dpi=300)
|
311 |
+
s = lambda x: 3.0 + x * (3 if metric == "msg_rates" else 0.1)
|
312 |
+
plt.scatter(
|
313 |
+
bases[metric][msg_type],
|
314 |
+
deltas[metric][msg_type],
|
315 |
+
s(values[metric][msg_type]),
|
316 |
+
)
|
317 |
+
plt.savefig(
|
318 |
+
os.path.join(save_dir, f"{metric}_{msg_type}_scatter.pdf"),
|
319 |
+
bbox_inches="tight",
|
320 |
+
)
|
321 |
+
|
322 |
+
os.makedirs(os.path.join(save_dir, "delta_effect"), exist_ok=True)
|
323 |
+
for metric in data:
|
324 |
+
for msg_type in data[metric]:
|
325 |
+
for base_value in base_set:
|
326 |
+
mask = bases[metric][msg_type] == base_value
|
327 |
+
fig = plt.figure(dpi=300)
|
328 |
+
s = lambda x: x / (1.0 if metric == "msg_rates" else 10.0)
|
329 |
+
plt.plot(
|
330 |
+
deltas[metric][msg_type][mask],
|
331 |
+
values[metric][msg_type][mask],
|
332 |
+
)
|
333 |
+
plt.savefig(
|
334 |
+
os.path.join(save_dir, f"delta_effect/{metric}_{msg_type}_base{base_value}.pdf"),
|
335 |
+
bbox_inches="tight",
|
336 |
+
)
|
337 |
+
os.makedirs(os.path.join(save_dir, "base_effect"), exist_ok=True)
|
338 |
+
for metric in data:
|
339 |
+
for msg_type in data[metric]:
|
340 |
+
for delta_value in delta_set:
|
341 |
+
mask = deltas[metric][msg_type] == delta_value
|
342 |
+
fig = plt.figure(dpi=300)
|
343 |
+
s = lambda x: x / (1.0 if metric == "msg_rates" else 10.0)
|
344 |
+
plt.plot(
|
345 |
+
bases[metric][msg_type][mask],
|
346 |
+
values[metric][msg_type][mask],
|
347 |
+
)
|
348 |
+
plt.savefig(
|
349 |
+
os.path.join(save_dir, f"base_effect/{metric}_{msg_type}_delta{delta_value}.pdf"),
|
350 |
+
bbox_inches="tight",
|
351 |
+
)
|
352 |
+
|
353 |
+
|
354 |
+
def main(args):
|
355 |
+
prompts = load_prompts(
|
356 |
+
args.num_prompts,
|
357 |
+
args.prompts_min_length,
|
358 |
+
args.prompts_file if not args.overwrite else None,
|
359 |
+
)
|
360 |
+
|
361 |
+
msgs_lens = []
|
362 |
+
for i in np.arange(
|
363 |
+
args.msgs_lengths[0],
|
364 |
+
args.msgs_lengths[1] + args.msgs_lengths[2],
|
365 |
+
args.msgs_lengths[2],
|
366 |
+
dtype=np.int32,
|
367 |
+
):
|
368 |
+
for _ in range(args.msgs_per_length):
|
369 |
+
msgs_lens.append(i)
|
370 |
+
|
371 |
+
msgs = load_msgs(
|
372 |
+
msgs_lens,
|
373 |
+
args.msgs_file if not args.overwrite else None,
|
374 |
+
)
|
375 |
+
|
376 |
+
if args.msgs_file:
|
377 |
+
if not os.path.isfile(args.msgs_file) or args.overwrite:
|
378 |
+
os.makedirs(os.path.dirname(args.msgs_file), exist_ok=True)
|
379 |
+
with open(args.msgs_file, "w") as f:
|
380 |
+
json.dump(msgs, f)
|
381 |
+
print(f"Saved messages to {args.msgs_file}")
|
382 |
+
if args.prompts_file:
|
383 |
+
if not os.path.isfile(args.prompts_file) or args.overwrite:
|
384 |
+
os.makedirs(os.path.dirname(args.prompts_file), exist_ok=True)
|
385 |
+
with open(args.prompts_file, "w") as f:
|
386 |
+
json.dump(prompts, f)
|
387 |
+
print(f"Saved prompts to {args.prompts_file}")
|
388 |
+
|
389 |
+
if args.results_load_file:
|
390 |
+
with open(args.results_load_file, "r") as f:
|
391 |
+
results = json.load(f)
|
392 |
+
else:
|
393 |
+
results = get_results(args, prompts, msgs)
|
394 |
+
|
395 |
+
if args.results_save_file:
|
396 |
+
os.makedirs(os.path.dirname(args.results_save_file), exist_ok=True)
|
397 |
+
with open(args.results_save_file, "w") as f:
|
398 |
+
json.dump(results, f)
|
399 |
+
print(f"Saved results to {args.results_save_file}")
|
400 |
+
|
401 |
+
if args.figs_dir:
|
402 |
+
process_results(results, args.figs_dir)
|
403 |
+
|
404 |
+
|
405 |
+
|
406 |
+
if __name__ == "__main__":
|
407 |
+
args = create_args()
|
408 |
+
main(args)
|
api.py
CHANGED
@@ -108,6 +108,9 @@ async def default_config():
|
|
108 |
"private_key": GlobalConfig.get(
|
109 |
"encrypt.default", "private_key"
|
110 |
),
|
|
|
|
|
|
|
111 |
"max_new_tokens_ratio": GlobalConfig.get(
|
112 |
"encrypt.default", "max_new_tokens_ratio"
|
113 |
),
|
|
|
108 |
"private_key": GlobalConfig.get(
|
109 |
"encrypt.default", "private_key"
|
110 |
),
|
111 |
+
"min_new_tokens_ratio": GlobalConfig.get(
|
112 |
+
"encrypt.default", "min_new_tokens_ratio"
|
113 |
+
),
|
114 |
"max_new_tokens_ratio": GlobalConfig.get(
|
115 |
"encrypt.default", "max_new_tokens_ratio"
|
116 |
),
|
config.ini
CHANGED
@@ -32,6 +32,7 @@ msg_base = int:2
|
|
32 |
seed_scheme = str:sha_left_hash
|
33 |
window_length = int:1
|
34 |
private_key = int:0
|
|
|
35 |
max_new_tokens_ratio = float:2.0
|
36 |
num_beams = int:4
|
37 |
repetition_penalty = float:1.0
|
|
|
32 |
seed_scheme = str:sha_left_hash
|
33 |
window_length = int:1
|
34 |
private_key = int:0
|
35 |
+
min_new_tokens_ratio = float:1.0
|
36 |
max_new_tokens_ratio = float:2.0
|
37 |
num_beams = int:4
|
38 |
repetition_penalty = float:1.0
|
model_factory.py
CHANGED
@@ -63,7 +63,8 @@ class ModelFactory:
|
|
63 |
@classmethod
|
64 |
def load_model(cls, name):
|
65 |
if name not in cls.models:
|
66 |
-
cls.__load_model(name)
|
|
|
67 |
|
68 |
if name != cls.run_model and cls.run_model is not None:
|
69 |
cls.models[cls.run_model].to(cls.load_device)
|
@@ -83,3 +84,43 @@ class ModelFactory:
|
|
83 |
return cls.tokenizers[name].model_max_length
|
84 |
else:
|
85 |
return 0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
@classmethod
|
64 |
def load_model(cls, name):
|
65 |
if name not in cls.models:
|
66 |
+
if cls.__load_model(name) is None:
|
67 |
+
return None, None
|
68 |
|
69 |
if name != cls.run_model and cls.run_model is not None:
|
70 |
cls.models[cls.run_model].to(cls.load_device)
|
|
|
84 |
return cls.tokenizers[name].model_max_length
|
85 |
else:
|
86 |
return 0
|
87 |
+
|
88 |
+
@classmethod
|
89 |
+
def compute_perplexity(cls, model_name, text):
|
90 |
+
# This code is copied from https://huggingface.co/docs/transformers/perplexity
|
91 |
+
model, tokenizer = cls.load_model(model_name)
|
92 |
+
if model is None or tokenizer is None:
|
93 |
+
return 0
|
94 |
+
device = model.device
|
95 |
+
encodings = tokenizer(text, return_tensors="pt").to(device)
|
96 |
+
|
97 |
+
max_length = model.config.n_positions
|
98 |
+
stride = max_length//2
|
99 |
+
seq_len = encodings.input_ids.size(1)
|
100 |
+
|
101 |
+
nlls = []
|
102 |
+
prev_end_loc = 0
|
103 |
+
for begin_loc in range(0, seq_len, stride):
|
104 |
+
end_loc = min(begin_loc + max_length, seq_len)
|
105 |
+
trg_len = end_loc - prev_end_loc # may be different from stride on last loop
|
106 |
+
input_ids = encodings.input_ids[:, begin_loc:end_loc].to(device)
|
107 |
+
target_ids = input_ids.clone()
|
108 |
+
target_ids[:, :-trg_len] = -100
|
109 |
+
|
110 |
+
with torch.no_grad():
|
111 |
+
outputs = model(input_ids, labels=target_ids)
|
112 |
+
|
113 |
+
# loss is calculated using CrossEntropyLoss which averages over valid labels
|
114 |
+
# N.B. the model only calculates loss over trg_len - 1 labels, because it internally shifts the labels
|
115 |
+
# to the left by 1.
|
116 |
+
neg_log_likelihood = outputs.loss
|
117 |
+
|
118 |
+
nlls.append(neg_log_likelihood)
|
119 |
+
|
120 |
+
prev_end_loc = end_loc
|
121 |
+
if end_loc == seq_len:
|
122 |
+
break
|
123 |
+
|
124 |
+
ppl = torch.exp(torch.stack(nlls).mean()).item()
|
125 |
+
return ppl
|
126 |
+
|
requirements.txt
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
numpy==1.26.4
|
2 |
tqdm==4.66.4
|
3 |
transformers==4.41.2
|
|
|
4 |
PyYAML==6.0.1
|
5 |
scikit-learn==1.5.0
|
6 |
torch==2.3.0
|
@@ -8,3 +9,4 @@ cryptography==42.0.8
|
|
8 |
fastapi
|
9 |
gradio
|
10 |
uvicorn
|
|
|
|
1 |
numpy==1.26.4
|
2 |
tqdm==4.66.4
|
3 |
transformers==4.41.2
|
4 |
+
datasets==2.20.0
|
5 |
PyYAML==6.0.1
|
6 |
scikit-learn==1.5.0
|
7 |
torch==2.3.0
|
|
|
9 |
fastapi
|
10 |
gradio
|
11 |
uvicorn
|
12 |
+
matplotlib==3.9.1
|
schemes.py
CHANGED
@@ -49,6 +49,11 @@ class EncryptionBody(BaseModel):
|
|
49 |
title="Private key used to compute the seed for PRF",
|
50 |
ge=0,
|
51 |
)
|
|
|
|
|
|
|
|
|
|
|
52 |
max_new_tokens_ratio: float = Field(
|
53 |
default=GlobalConfig.get("encrypt.default", "max_new_tokens_ratio"),
|
54 |
title="Max length of generated text compared to the minimum length required to hide the message",
|
|
|
49 |
title="Private key used to compute the seed for PRF",
|
50 |
ge=0,
|
51 |
)
|
52 |
+
max_new_tokens_ratio: float = Field(
|
53 |
+
default=GlobalConfig.get("encrypt.default", "min_new_tokens_ratio"),
|
54 |
+
title="Min length of generated text compared to the minimum length required to hide the message",
|
55 |
+
ge=1,
|
56 |
+
)
|
57 |
max_new_tokens_ratio: float = Field(
|
58 |
default=GlobalConfig.get("encrypt.default", "max_new_tokens_ratio"),
|
59 |
title="Max length of generated text compared to the minimum length required to hide the message",
|
stegno.py
CHANGED
@@ -18,9 +18,11 @@ def generate(
|
|
18 |
window_length: int = 1,
|
19 |
salt_key: Union[int, None] = None,
|
20 |
private_key: Union[int, None] = None,
|
|
|
21 |
max_new_tokens_ratio: float = 2,
|
22 |
num_beams: int = 4,
|
23 |
repetition_penalty: float = 1.0,
|
|
|
24 |
):
|
25 |
"""
|
26 |
Generate the sequence containing the hidden data.
|
@@ -36,7 +38,6 @@ def generate(
|
|
36 |
window_length: length of window to compute the seed.
|
37 |
salt_key: salt to add to the seed.
|
38 |
private_key: private key used to compute the seed.
|
39 |
-
|
40 |
"""
|
41 |
if len(start_pos_p) == 1:
|
42 |
start_pos = start_pos_p[0]
|
@@ -47,9 +48,10 @@ def generate(
|
|
47 |
start_pos = int(start_pos) + window_length
|
48 |
|
49 |
tokenized_input = tokenizer(prompt, return_tensors="pt").to(model.device)
|
50 |
-
prompt_size
|
|
|
51 |
logits_processor = EncryptorLogitsProcessor(
|
52 |
-
prompt_ids=tokenized_input.input_ids,
|
53 |
msg=msg,
|
54 |
start_pos=start_pos,
|
55 |
delta=delta,
|
@@ -62,14 +64,21 @@ def generate(
|
|
62 |
salt_key=salt_key,
|
63 |
private_key=private_key,
|
64 |
)
|
65 |
-
min_length =
|
66 |
-
|
67 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
)
|
69 |
max_length = min(max_length, tokenizer.model_max_length)
|
70 |
min_length = min(min_length, max_length)
|
71 |
output_tokens = model.generate(
|
72 |
-
|
|
|
73 |
logits_processor=transformers.LogitsProcessorList([logits_processor]),
|
74 |
min_length=min_length,
|
75 |
max_length=max_length,
|
@@ -79,10 +88,12 @@ def generate(
|
|
79 |
)
|
80 |
|
81 |
output_tokens = output_tokens[:, prompt_size:]
|
82 |
-
output_text = tokenizer.batch_decode(
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
|
|
86 |
msg_rates, tokens_infos = logits_processor.validate(
|
87 |
output_tokens_post.input_ids
|
88 |
)
|
|
|
18 |
window_length: int = 1,
|
19 |
salt_key: Union[int, None] = None,
|
20 |
private_key: Union[int, None] = None,
|
21 |
+
min_new_tokens_ratio: float = 1,
|
22 |
max_new_tokens_ratio: float = 2,
|
23 |
num_beams: int = 4,
|
24 |
repetition_penalty: float = 1.0,
|
25 |
+
prompt_size: int = -1,
|
26 |
):
|
27 |
"""
|
28 |
Generate the sequence containing the hidden data.
|
|
|
38 |
window_length: length of window to compute the seed.
|
39 |
salt_key: salt to add to the seed.
|
40 |
private_key: private key used to compute the seed.
|
|
|
41 |
"""
|
42 |
if len(start_pos_p) == 1:
|
43 |
start_pos = start_pos_p[0]
|
|
|
48 |
start_pos = int(start_pos) + window_length
|
49 |
|
50 |
tokenized_input = tokenizer(prompt, return_tensors="pt").to(model.device)
|
51 |
+
if prompt_size == -1:
|
52 |
+
prompt_size = tokenized_input.input_ids.size(1)
|
53 |
logits_processor = EncryptorLogitsProcessor(
|
54 |
+
prompt_ids=tokenized_input.input_ids[:prompt_size],
|
55 |
msg=msg,
|
56 |
start_pos=start_pos,
|
57 |
delta=delta,
|
|
|
64 |
salt_key=salt_key,
|
65 |
private_key=private_key,
|
66 |
)
|
67 |
+
min_length = (
|
68 |
+
prompt_size
|
69 |
+
+ start_pos
|
70 |
+
+ logits_processor.get_message_len() * min_new_tokens_ratio
|
71 |
+
)
|
72 |
+
max_length = (
|
73 |
+
prompt_size
|
74 |
+
+ start_pos
|
75 |
+
+ logits_processor.get_message_len() * max_new_tokens_ratio
|
76 |
)
|
77 |
max_length = min(max_length, tokenizer.model_max_length)
|
78 |
min_length = min(min_length, max_length)
|
79 |
output_tokens = model.generate(
|
80 |
+
input_ids=tokenized_input.input_ids[:, :prompt_size],
|
81 |
+
attention_mask=tokenized_input.attention_mask[:, :prompt_size],
|
82 |
logits_processor=transformers.LogitsProcessorList([logits_processor]),
|
83 |
min_length=min_length,
|
84 |
max_length=max_length,
|
|
|
88 |
)
|
89 |
|
90 |
output_tokens = output_tokens[:, prompt_size:]
|
91 |
+
output_text = tokenizer.batch_decode(
|
92 |
+
output_tokens, skip_special_tokens=True
|
93 |
+
)[0]
|
94 |
+
output_tokens_post = tokenizer(
|
95 |
+
output_text, return_tensors="pt", add_special_tokens=False
|
96 |
+
).to(model.device)
|
97 |
msg_rates, tokens_infos = logits_processor.validate(
|
98 |
output_tokens_post.input_ids
|
99 |
)
|
utils.py
CHANGED
@@ -55,3 +55,4 @@ def static_init(cls):
|
|
55 |
if getattr(cls, "__static_init__", None):
|
56 |
cls.__static_init__()
|
57 |
return cls
|
|
|
|
55 |
if getattr(cls, "__static_init__", None):
|
56 |
cls.__static_init__()
|
57 |
return cls
|
58 |
+
|