Text Classification
Keras
English
Azerbaijani
prompt-injection
security
llm-security
document-security
retvec
cnn
tensorflow
fastapi
Eval Results (legacy)
Instructions to use MegrurNiftiyev/MyGuard-Prompt-Injection-Detector with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Keras
How to use MegrurNiftiyev/MyGuard-Prompt-Injection-Detector with Keras:
# Available backend options are: "jax", "torch", "tensorflow". import os os.environ["KERAS_BACKEND"] = "jax" import keras model = keras.saving.load_model("hf://MegrurNiftiyev/MyGuard-Prompt-Injection-Detector") - Notebooks
- Google Colab
- Kaggle
File size: 24,261 Bytes
215f97f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 | """
Standalone RETVec+CNN Keras model training & held-out test evaluation script.
Usage:
python train_model.py
python -m app.scripts.train_model
"""
import os
import sys
import random
import zipfile
import docx
import pypdf
from pptx import Presentation
import numpy as np
os.environ["TF_USE_LEGACY_KERAS"] = "1"
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
sys.stdout.reconfigure(encoding='utf-8')
# Ensure project root is in sys.path
BASE_DIR = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", ".."))
if BASE_DIR not in sys.path:
sys.path.insert(0, BASE_DIR)
SEED = 42
random.seed(SEED)
np.random.seed(SEED)
import tensorflow as tf
tf.random.set_seed(SEED)
from app.ml.cnn.architecture import build_model, LABEL_NAMES
from app.ml.training.data.encoding import encode_labels
from app.ml.training.train import get_class_weights
from app.ml.preprocessing.chunking import chunk_text
HELDOUT_TEST_FILES = {
"benign": [
"09_resmi_mektub_temiz.docx",
"10_iclas_protokolu_temiz.docx",
"Monthly Financial Expense Report.pdf",
"11_ezamiyye_emri_temiz.docx",
"19_sifaris_senedi_temiz.docx"
],
"injection": [
"01_Aylıq_Fəaliyyət_Hesabatı.docx",
"16_ezamiyye_xercleri_injection_gizli.docx",
"19_sifaris_senedi_problem.docx",
"23_bank_zemanet_mektubu_injection_context_hijack.docx",
"24_qebul_tehvil_akti_injection.docx"
]
}
def extract_pptx(file_path: str) -> str:
"""Extract slide paragraph text and notes text from PPTX files using python-pptx."""
try:
prs = Presentation(file_path)
parts = []
for slide in prs.slides:
for shape in slide.shapes:
if shape.has_text_frame:
for para in shape.text_frame.paragraphs:
line = "".join(run.text for run in para.runs)
if line.strip():
parts.append(line.strip())
if slide.has_notes_slide and slide.notes_slide.notes_text_frame:
note = slide.notes_slide.notes_text_frame.text
if note.strip():
parts.append(note.strip())
return "\n".join(parts)
except Exception as e:
print(f"Warning reading PPTX {file_path}: {e}")
return ""
def extract_text(file_path: str) -> str:
"""Extract raw text from supported document formats (.docx, .pptx, .pdf, .zip, .txt)."""
ext = os.path.splitext(file_path)[1].lower()
text = ""
try:
if ext == ".docx":
doc = docx.Document(file_path)
parts = [p.text for p in doc.paragraphs if p.text.strip()]
for table in doc.tables:
for row in table.rows:
for cell in row.cells:
if cell.text.strip():
parts.append(cell.text.strip())
text = "\n".join(parts)
elif ext == ".pptx":
text = extract_pptx(file_path)
elif ext == ".pdf":
reader = pypdf.PdfReader(file_path)
parts = []
for i, page in enumerate(reader.pages):
if i >= 20:
break
try:
t = page.extract_text()
if t:
parts.append(t.strip())
except Exception:
continue
text = "\n".join(parts)
elif ext == ".zip":
parts = []
with zipfile.ZipFile(file_path, 'r') as z:
for name in z.namelist():
if name.endswith('.docx'):
tmp_path = os.path.join(os.path.dirname(file_path), "_tmp_extracted.docx")
with open(tmp_path, "wb") as f_out:
f_out.write(z.read(name))
sub_text = extract_text(tmp_path)
if os.path.exists(tmp_path):
os.remove(tmp_path)
parts.append(sub_text)
elif name.endswith('.pptx'):
tmp_path = os.path.join(os.path.dirname(file_path), "_tmp_extracted.pptx")
with open(tmp_path, "wb") as f_out:
f_out.write(z.read(name))
sub_text = extract_text(tmp_path)
if os.path.exists(tmp_path):
os.remove(tmp_path)
parts.append(sub_text)
elif name.endswith('.txt'):
parts.append(z.read(name).decode('utf-8', errors='ignore'))
text = "\n".join(parts)
elif ext == ".txt":
with open(file_path, "r", encoding="utf-8", errors="ignore") as f:
text = f.read()
else:
print(f"Skipping unsupported file extension {ext} for {file_path}")
return ""
except Exception as e:
print(f"Warning reading {file_path}: {e}")
return text.strip()
def split_documents(doc_ids: list[str], val_ratio: float = 0.15, seed: int = 42) -> tuple[set[str], set[str]]:
"""Perform a document-level split of source document IDs into train and validation sets."""
rng = random.Random(seed)
unique_ids = list(dict.fromkeys(doc_ids))
rng.shuffle(unique_ids)
n_val = max(1, int(len(unique_ids) * val_ratio))
val_ids = set(unique_ids[:n_val])
train_ids = set(unique_ids[n_val:])
return train_ids, val_ids
def load_real_dataset(raw_dir: str):
all_chunks = [] # [(doc_id, text_chunk, label)]
all_doc_ids = []
test_docs = []
# Define folder mapping: (folder_path, default_category)
folders_to_scan = [
(os.path.join(raw_dir, "benign"), "benign"),
(os.path.join(raw_dir, "injection"), "injection"),
]
downloaded_dir = os.path.join(raw_dir, "downloaded")
if os.path.exists(downloaded_dir):
for root, dirs, files in os.walk(downloaded_dir):
if files:
folders_to_scan.append((root, "benign"))
scanned_file_counts = {}
# Load 10,200 PDF V4 Synthetic Dataset if dataset_V4.csv exists
v4_csv_path = os.path.join(downloaded_dir, "dataset_V4.csv")
if os.path.exists(v4_csv_path):
try:
import pandas as pd
print(f"Loading 10,200 PDF V4 Synthetic Dataset samples from {v4_csv_path}...")
df_v4 = pd.read_csv(v4_csv_path)
v4_count = 0
for _, row in df_v4.iterrows():
doc_id = f"v4_{row['doc_id']}"
extracted_text = str(row['extracted_text']) if pd.notna(row['extracted_text']) else ""
if not extracted_text.strip():
continue
is_inj = bool(row['is_injected'])
lbl = "injection" if is_inj else "safe"
v4_count += 1
lines = [l.strip() for l in extracted_text.split("\n") if l.strip()]
for line in lines:
words = line.split()
if len(words) <= 60:
all_chunks.append((doc_id, line, lbl))
all_doc_ids.append(doc_id)
else:
for c in chunk_text(line):
all_chunks.append((doc_id, c, lbl))
all_doc_ids.append(doc_id)
scanned_file_counts["dataset_V4.csv (10,200 PDFs)"] = v4_count
except Exception as err:
print(f"Warning loading dataset_V4.csv: {err}")
for cat_dir, category in folders_to_scan:
if not os.path.exists(cat_dir):
continue
heldout_list = HELDOUT_TEST_FILES.get(category, [])
label_str = "safe" if category == "benign" else "injection"
dir_key = os.path.relpath(cat_dir, raw_dir)
scanned_file_counts[dir_key] = scanned_file_counts.get(dir_key, 0)
for fname in os.listdir(cat_dir):
fpath = os.path.join(cat_dir, fname)
if not os.path.isfile(fpath):
continue
extracted = extract_text(fpath)
if not extracted:
continue
scanned_file_counts[dir_key] += 1
doc_id = os.path.relpath(fpath, raw_dir)
if fname in heldout_list:
test_docs.append({
"filename": fname,
"category": category,
"expected_label": label_str,
"text": extracted
})
else:
ext = os.path.splitext(fname)[1].lower()
file_chunks = []
# Check for docx paragraph-level white font / hidden text
docx_inj_lines = set()
if ext == ".docx":
try:
doc = docx.Document(fpath)
for p in doc.paragraphs:
ptxt = p.text.strip()
if not ptxt:
continue
is_p_white = False
for r in p.runs:
if r.font.color and r.font.color.rgb and str(r.font.color.rgb).upper() in ("FFFFFF", "FFF"):
is_p_white = True
break
if r._r.rPr is not None and r._r.rPr.find(docx.oxml.ns.qn("w:vanish")) is not None:
is_p_white = True
break
if is_p_white:
docx_inj_lines.add(ptxt)
except Exception:
pass
MANUAL_STEALTHY_LABELS = {
"23_bank_zemanet_mektubu_injection_context_hijack.docx": "450,000 azn",
"24_qebul_tehvil_akti_injection.docx": "administrative correction",
"25_sigorta_polisi_injection.docx": "franchise override protocol",
"26_emek_muqavilesi_inject.docx": "critical parser override",
"27_vekaletname_inject.docx": "müvəqqəti hüquqi düzəliş",
"29_bank_rekvizit_deyisikligi_injection.docx": "sistem təlimatı",
"31_tecili_odenis_memo_inject.docx": "suni intelleke",
"32_hosting_deyisiklik_bildirisi_inject.docx": "audit id-7741",
"33_elave_is_sifarisi_inject.docx": "audit id-112",
"34_distributor_endirim_teklifi_inject.docx": "audit id-267",
"38_icare_muqavilesi_injection.docx": "tərəfdaşlıq ianəsi",
"39_dasima_xidmeti_muqavilesi_inject.docx": "<last update>"
}
lines = [l.strip() for l in extracted.split("\n") if l.strip()]
for line in lines:
is_inj_line = False
if category == "injection":
low = line.lower()
if fname in MANUAL_STEALTHY_LABELS:
if MANUAL_STEALTHY_LABELS[fname] in low:
is_inj_line = True
else:
low = line.lower()
if line in docx_inj_lines or any(kw in low for kw in [
"prompt", "system", "yuxarida", "mene", "ignore", "override",
"@", "//", "#", "||", "^^", "***", "&&", "<system", "[system",
"internal system update", "forget", "unrestricted"
]):
is_inj_line = True
lbl = "injection" if (category == "injection" and is_inj_line) else "safe"
words = line.split()
if len(words) <= 60:
file_chunks.append((line, lbl))
else:
for c in chunk_text(line):
file_chunks.append((c, lbl))
# Cap per-document safe chunks so long PDFs don't dominate dataset (max 15 safe chunks per doc)
if len(file_chunks) > 15:
inj_chunks = [c for c in file_chunks if c[1] == "injection"]
safe_chunks = [c for c in file_chunks if c[1] == "safe"]
needed_safe = max(5, 15 - len(inj_chunks))
step = max(1, len(safe_chunks) // needed_safe) if safe_chunks else 1
file_chunks = inj_chunks + (safe_chunks[::step][:needed_safe] if safe_chunks else [])
for text_chunk, lbl in file_chunks:
all_chunks.append((doc_id, text_chunk, lbl))
all_doc_ids.append(doc_id)
# Document-level split
train_doc_ids, val_doc_ids = split_documents(all_doc_ids, val_ratio=0.15, seed=SEED)
train_tuples = [c for c in all_chunks if c[0] in train_doc_ids]
val_tuples = [c for c in all_chunks if c[0] in val_doc_ids]
# Oversample injection training tuples so model learns injection patterns properly
train_inj_tuples = [t for t in train_tuples if t[2] == "injection"]
train_safe_tuples = [t for t in train_tuples if t[2] == "safe"]
if train_inj_tuples and len(train_safe_tuples) > 0:
multiplier = max(1, (len(train_safe_tuples) // 3) // len(train_inj_tuples))
train_inj_oversampled = train_inj_tuples * multiplier
train_tuples = train_safe_tuples + train_inj_oversampled
# Thorough random shuffling across all sources, classes, and languages
rng = random.Random(SEED)
rng.shuffle(train_tuples)
rng.shuffle(val_tuples)
train_texts = [t[1] for t in train_tuples]
train_labels = [t[2] for t in train_tuples]
val_texts = [t[1] for t in val_tuples]
val_labels = [t[2] for t in val_tuples]
print("Scanned files count per folder:")
for folder_rel, count in scanned_file_counts.items():
print(f" - {folder_rel}: {count} valid documents")
print(f"Document-level split: {len(train_doc_ids)} train docs ({len(train_texts)} chunks), {len(val_doc_ids)} val docs ({len(val_texts)} chunks)")
return (train_texts, train_labels), (val_texts, val_labels), test_docs
def main():
raw_dir = os.path.join(BASE_DIR, "data", "raw")
print("Reading document dataset from data/raw...")
(train_texts, train_labels), (val_texts, val_labels), test_docs = load_real_dataset(raw_dir)
print(f"\n--- Dataset Loading Summary ---")
print(f"Training text chunks extracted: {len(train_texts)}")
print(f" - Safe (Benign) train chunks: {train_labels.count('safe')}")
print(f" - Injection train chunks: {train_labels.count('injection')}")
print(f"Validation text chunks extracted: {len(val_texts)}")
print(f" - Safe (Benign) val chunks: {val_labels.count('safe')}")
print(f" - Injection val chunks: {val_labels.count('injection')}")
print(f"Held-out Test Files reserved: {len(test_docs)}")
for td in test_docs:
print(f" * [{td['category'].upper()}] {td['filename']} ({len(td['text'])} chars)")
X_train = np.array([[t] for t in train_texts])
Y_train_label = encode_labels(train_labels)
X_val = np.array([[t] for t in val_texts])
Y_val_label = encode_labels(val_labels)
class_weights_dict = get_class_weights(Y_train_label)
sample_weights_label = np.array([class_weights_dict[int(np.argmax(y))] for y in Y_train_label], dtype=np.float32)
print("\nBuilding RETVec + CNN Keras Classification Model...")
model = build_model(sequence_length=128)
model.summary()
print("\nStarting Keras Model Training (5 Epochs, batch_size=128, document-level validation)...", flush=True)
history = model.fit(
X_train,
Y_train_label,
epochs=5,
batch_size=128,
validation_data=(X_val, Y_val_label),
sample_weight=sample_weights_label,
verbose=1
)
models_dir = os.path.join(BASE_DIR, "data", "models")
os.makedirs(models_dir, exist_ok=True)
keras_model_path = os.path.join(models_dir, "retvec_cnn_model.keras")
print(f"\nSaving trained model to .keras file at:\n {keras_model_path}")
model.save(keras_model_path)
cache_dir = os.path.join(BASE_DIR, "data", "cache")
os.makedirs(cache_dir, exist_ok=True)
model.save(os.path.join(cache_dir, "active_model.keras"))
print("\n==========================================")
print("HELD-OUT TEST FILES INFERENCE & EVALUATION")
print("==========================================")
correct_predictions = 0
test_results = []
for td in test_docs:
raw_text = td["text"]
lines = [l.strip() for l in raw_text.split("\n") if l.strip()]
chunks = []
for line in lines:
words = line.split()
if len(words) <= 60:
chunks.append(line)
else:
chunks.extend(chunk_text(line))
chunk_inputs = np.array([[c] for c in chunks])
preds = model.predict(chunk_inputs, verbose=0)
label_preds = preds if isinstance(preds, np.ndarray) and preds.ndim == 2 else preds[0]
worst_chunk_idx = label_preds[:, 2].argmax()
max_injection_prob = float(label_preds[worst_chunk_idx, 2])
max_inj_line = chunks[worst_chunk_idx] if chunks else ""
avg_probs = np.mean(label_preds, axis=0)
HIGH_CONF_THRESHOLD = 0.85
CORROBORATION_THRESHOLD = 0.60
MIN_CORROBORATING_CHUNKS = 2
injection_probs = [float(p) for p in label_preds[:, 2]]
predicted_label = "safe"
high_conf = [p for p in injection_probs if p >= HIGH_CONF_THRESHOLD]
if high_conf:
predicted_label = "injection"
else:
corroborating = [p for p in injection_probs if p >= CORROBORATION_THRESHOLD]
if len(corroborating) >= MIN_CORROBORATING_CHUNKS:
predicted_label = "injection"
is_correct = (predicted_label == td["expected_label"])
if is_correct:
correct_predictions += 1
test_results.append({
"filename": td["filename"],
"expected": td["expected_label"],
"predicted": predicted_label,
"is_correct": is_correct,
"prob_safe": float(avg_probs[0]),
"prob_suspicious": float(avg_probs[1]),
"prob_injection": float(avg_probs[2]),
"max_chunk_injection": float(max_injection_prob),
"max_inj_snippet": max_inj_line[:60]
})
status = "PASSED ✓" if is_correct else "FAILED ✗"
print(f"File: {td['filename']}")
print(f" Expected: {td['expected_label']} | Predicted: {predicted_label} [{status}]")
print(f" Max Injection Prob: {max_injection_prob:.2%} | Snippet: {max_inj_line[:70]!r}\n")
accuracy = (correct_predictions / len(test_docs)) * 100 if test_docs else 0.0
print(f"Final Held-Out Test Accuracy: {accuracy:.2f}% ({correct_predictions}/{len(test_docs)})")
last_loss = float(history.history["loss"][-1]) if "history" in locals() and "loss" in history.history else 0.0
last_acc = float(history.history["accuracy"][-1]) if "history" in locals() and "accuracy" in history.history else 0.0
last_val = float(history.history["val_accuracy"][-1]) if "history" in locals() and "val_accuracy" in history.history else 0.0
prompt_local_push_confirmation(
model=model,
accuracy=accuracy,
correct_count=correct_predictions,
total_test_docs=len(test_docs),
train_chunk_count=len(train_texts),
val_chunk_count=len(val_texts),
last_train_loss=last_loss,
last_train_acc=last_acc,
last_val_acc=last_val,
test_results=test_results,
)
def fetch_last_5_models_from_firestore():
init_firebase()
db = get_firestore_db()
if db is None:
return [], 0
try:
docs = db.collection("models").get()
model_list = []
max_run_num = 0
for doc in docs:
d = doc.to_dict()
v_id = d.get("version") or doc.id
if v_id.startswith("run-"):
try:
r_num = int(v_id.split("-")[1])
if r_num > max_run_num:
max_run_num = r_num
except ValueError:
pass
model_list.append(d)
def sort_key(d):
v = d.get("version", "")
if v.startswith("run-"):
try:
return int(v.split("-")[1])
except ValueError:
pass
return 0
model_list.sort(key=sort_key)
return model_list[-5:], max_run_num
except Exception as e:
print(f"Warning fetching models from Firestore: {e}")
return [], 0
def prompt_local_push_confirmation(model, accuracy: float, correct_count: int, total_test_docs: int, train_chunk_count: int, val_chunk_count: int, last_train_loss: float, last_train_acc: float, last_val_acc: float, test_results: list):
import subprocess
import asyncio
from datetime import datetime, timezone
from app.core.firebase import init_firebase, get_firestore_db
from app.ml.serving.registry import save_model_version
last_5, max_run_num = fetch_last_5_models_from_firestore()
inj_docs = [t for t in test_results if t["expected"] == "injection"]
inj_correct = [t for t in inj_docs if t["is_correct"]]
test_recall = (len(inj_correct) / len(inj_docs) * 100.0) if inj_docs else 100.0
if last_5:
print("\nLast 5 registered versions:")
for m in last_5:
v_str = m.get("version", "unknown")
metrics_m = m.get("metrics", {})
test_acc_m = metrics_m.get("test_acc", 0.0) * 100.0 if isinstance(metrics_m.get("test_acc"), (int, float)) else 0.0
recall_m = metrics_m.get("recall", 0.0) * 100.0 if isinstance(metrics_m.get("recall"), (int, float)) else 0.0
status_tag = " (currently active)" if m.get("status") == "active" else ""
print(f" {v_str:<8} test acc {test_acc_m:.2f}% recall {recall_m:.0f}%{status_tag}")
print(f"\nThis run: test acc {accuracy:.2f}% recall {test_recall:.0f}%\n")
answer = input("Upload this model to Firebase as a new candidate version? (y/n): ").strip().lower()
if answer == "y":
next_run_num = max_run_num + 1 if max_run_num > 0 else 12
new_version_id = f"run-{next_run_num:02d}"
try:
res = subprocess.run(["git", "rev-parse", "HEAD"], capture_output=True, text=True, check=True)
source_commit = res.stdout.strip()
except Exception:
source_commit = "unknown"
today_str = datetime.now(timezone.utc).strftime("%Y-%m-%d")
desc = (
f"Trained {today_str}. "
f"Dataset: {train_chunk_count} train chunks + {val_chunk_count} val chunks. "
f"Held-out test: {accuracy:.2f}% accuracy ({correct_count}/{total_test_docs}), "
f"{test_recall:.0f}% injection recall."
)
metrics_payload = {
"train_loss": float(last_train_loss),
"train_acc": float(last_train_acc),
"val_acc": float(last_val_acc),
"test_acc": float(accuracy / 100.0),
"recall": float(test_recall / 100.0),
"correct_test": f"{correct_count}/{total_test_docs}",
}
asyncio.run(
save_model_version(
model=model,
metrics=metrics_payload,
version=new_version_id,
status="candidate",
source_commit=source_commit,
description=desc,
)
)
print(f"Uploaded as candidate version '{new_version_id}'. Use POST /model/change-version/{new_version_id} to make it active.")
else:
print("Skipped. Model saved locally only at data/models/retvec_cnn_model.keras.")
if __name__ == "__main__":
main()
|