Spaces:
Sleeping
Sleeping
File size: 15,032 Bytes
ee1f3df | 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 | import io
import os
import re
import urllib.request
from typing import Any, Dict, Iterable, List, Tuple, Union
import numpy as np
import cv2
from PIL import Image
import easyocr
from sentence_transformers import SentenceTransformer
import torch
from transformers import (
AutoModelForSequenceClassification,
AutoTokenizer,
BlipForConditionalGeneration,
BlipProcessor,
pipeline,
)
try:
from optimum.onnxruntime import ORTModelForSequenceClassification
_HAS_ORT = True
except Exception:
_HAS_ORT = False
THREAT_DICTIONARY = [
"ignore previous",
"system override",
"transfer funds",
"bypass safety",
"disable guardrails",
"override policy",
"reveal secrets",
]
class PromptInjectionEngine:
def __init__(
self,
use_onnx: bool | None = None,
force_cpu: bool | None = None,
model_name: str | None = None,
) -> None:
os.environ.setdefault("HF_HUB_TIMEOUT", "60")
os.environ.setdefault("HF_HUB_DOWNLOAD_TIMEOUT", "60")
os.environ.setdefault("HF_HUB_DISABLE_TELEMETRY", "1")
self._ocr: easyocr.Reader | None = None
self._injection_classifier = None
self._model_name = model_name or "protectai/deberta-v3-base-prompt-injection"
if force_cpu is None:
self._force_cpu = os.environ.get("SENTINEL_FORCE_CPU", "").lower() in {
"1",
"true",
"yes",
}
else:
self._force_cpu = force_cpu
if use_onnx is None:
self._use_onnx = os.environ.get("SENTINEL_USE_ONNX", "1") not in {"0", "false"}
else:
self._use_onnx = use_onnx
def _get_ocr(self) -> easyocr.Reader:
if self._ocr is None:
ocr_gpu = os.environ.get("SENTINEL_OCR_GPU", "1") not in {"0", "false"}
try:
self._ocr = easyocr.Reader(["en"], gpu=ocr_gpu)
except Exception:
self._ocr = easyocr.Reader(["en"], gpu=False)
return self._ocr
def _get_injection_classifier(self):
if self._injection_classifier is None:
if self._use_onnx and _HAS_ORT:
try:
tokenizer = AutoTokenizer.from_pretrained(
self._model_name, subfolder="onnx", local_files_only=True
)
model = ORTModelForSequenceClassification.from_pretrained(
self._model_name, subfolder="onnx", export=False, local_files_only=True
)
except Exception:
tokenizer = AutoTokenizer.from_pretrained(self._model_name, subfolder="onnx")
model = ORTModelForSequenceClassification.from_pretrained(
self._model_name, subfolder="onnx", export=False
)
self._injection_classifier = pipeline(
"text-classification",
model=model,
tokenizer=tokenizer,
truncation=True,
max_length=512,
)
else:
try:
tokenizer = AutoTokenizer.from_pretrained(
self._model_name, local_files_only=True
)
model = AutoModelForSequenceClassification.from_pretrained(
self._model_name, local_files_only=True
)
except Exception:
tokenizer = AutoTokenizer.from_pretrained(self._model_name)
model = AutoModelForSequenceClassification.from_pretrained(self._model_name)
device = torch.device(
"cpu"
if self._force_cpu or not torch.backends.mps.is_available()
else "mps"
)
self._injection_classifier = pipeline(
"text-classification",
model=model,
tokenizer=tokenizer,
truncation=True,
max_length=512,
device=device,
)
return self._injection_classifier
@staticmethod
def _normalize_text(text: str) -> str:
lowered = text.lower()
cleaned = re.sub(r"[^a-z0-9]+", " ", lowered)
tokens = cleaned.split()
def merge_single_letter_runs(items: Iterable[str]) -> List[str]:
merged: List[str] = []
run: List[str] = []
for token in items:
if len(token) == 1:
run.append(token)
continue
if run:
merged.append("".join(run))
run = []
merged.append(token)
if run:
merged.append("".join(run))
return merged
merged_tokens = merge_single_letter_runs(tokens)
return " ".join(merged_tokens)
@staticmethod
def _load_image_for_ocr(image: Union[str, bytes]) -> Union[str, np.ndarray]:
if isinstance(image, str):
return image
pil_image = Image.open(io.BytesIO(image)).convert("RGB")
return np.array(pil_image)
@staticmethod
def _enhance_for_hidden_text(image: np.ndarray) -> np.ndarray:
gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
enhanced = clahe.apply(gray)
_, binary = cv2.threshold(enhanced, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
return cv2.cvtColor(binary, cv2.COLOR_GRAY2RGB)
@staticmethod
def _load_image_for_clip(image: Union[str, bytes]) -> Image.Image:
if isinstance(image, str):
return Image.open(image).convert("RGB")
return Image.open(io.BytesIO(image)).convert("RGB")
@staticmethod
def _extract_ocr_text(ocr_result: List[Any]) -> Tuple[str, List[Tuple[str, float]]]:
fragments: List[str] = []
scored: List[Tuple[str, float]] = []
# EasyOCR returns: [([bbox], text, confidence), ...]
for line in ocr_result or []:
if not line or len(line) < 2:
continue
text = str(line[1])
score = float(line[2]) if len(line) > 2 and isinstance(line[2], (float, int)) else None
if text:
fragments.append(text)
if score is not None:
scored.append((text, score))
return " ".join(fragments), scored
def detect_injection(self, image: Union[str, bytes]) -> Dict[str, Any]:
text_payload = self.extract_text(image)
return self.detect_injection_from_text(
text_payload["normalized_text"],
matched_phrases=[
phrase for phrase in THREAT_DICTIONARY if phrase in text_payload["normalized_text"]
],
)
def detect_injection_from_text(
self, normalized_text: str, matched_phrases: List[str] | None = None
) -> Dict[str, Any]:
if not normalized_text:
return {
"is_threat": False,
"risk_score": 0.0,
"reason": "No readable text detected in image.",
}
matched = matched_phrases or [
phrase for phrase in THREAT_DICTIONARY if phrase in normalized_text
]
try:
classifier = self._get_injection_classifier()
classification = classifier(normalized_text, top_k=1)[0]
label = str(classification.get("label", "")).upper()
score = float(classification.get("score", 0.0))
is_injection = "1" in label or "INJECTION" in label
risk_score = score if is_injection else 1.0 - score
reason_parts = [
f"Model={label or 'UNKNOWN'}",
f"model_score={score:.3f}",
]
except Exception:
is_injection = bool(matched)
risk_score = 0.9 if matched else 0.1
reason_parts = ["Model=FALLBACK", "model_score=0.0"]
if matched:
reason_parts.append(f"matched_phrases={', '.join(sorted(set(matched)))}")
return {
"is_threat": bool(is_injection),
"risk_score": round(risk_score, 3),
"reason": "; ".join(reason_parts),
}
def extract_text(self, image: Union[str, bytes]) -> Dict[str, Any]:
ocr_input = self._load_image_for_ocr(image)
reader = self._get_ocr()
if isinstance(ocr_input, str):
ocr_result = reader.readtext(ocr_input)
raw_text, scored = self._extract_ocr_text(ocr_result)
normalized = self._normalize_text(raw_text)
else:
base_result = reader.readtext(ocr_input)
enhanced_image = self._enhance_for_hidden_text(ocr_input)
enhanced_result = reader.readtext(enhanced_image)
raw_text_base, scored_base = self._extract_ocr_text(base_result)
raw_text_enh, scored_enh = self._extract_ocr_text(enhanced_result)
raw_text = " ".join([raw_text_base, raw_text_enh]).strip()
scored = scored_base + scored_enh
normalized = self._normalize_text(raw_text)
return {
"raw_text": raw_text,
"normalized_text": normalized,
"scored": scored,
}
class CrossModalEngine:
def __init__(self, clip_model: str | None = None, caption_model: str | None = None) -> None:
self._clip = SentenceTransformer(
clip_model or os.environ.get("SENTINEL_CLIP_MODEL", "clip-ViT-B-32")
)
self._captioner = None
self._caption_model = caption_model or os.environ.get(
"SENTINEL_BLIP_MODEL", "Salesforce/blip-image-captioning-base"
)
@staticmethod
def _load_image_for_clip(image: Union[str, bytes]) -> Image.Image:
if isinstance(image, str):
return Image.open(image).convert("RGB")
return Image.open(io.BytesIO(image)).convert("RGB")
def _get_captioner(self):
if self._captioner is None:
# Use BLIP processor + model directly to avoid pipeline task mismatches.
processor = BlipProcessor.from_pretrained(self._caption_model)
model = BlipForConditionalGeneration.from_pretrained(self._caption_model)
device = os.environ.get("SENTINEL_BLIP_DEVICE", "cpu")
model.to(device)
self._captioner = (processor, model, device)
return self._captioner
def check_cross_modal(self, image: Union[str, bytes], audio_transcript: str) -> Dict[str, Any]:
if not audio_transcript:
return {"is_mismatch": True, "consistency_score": 0.0}
pil_image = self._load_image_for_clip(image)
image_emb = self._clip.encode([pil_image], normalize_embeddings=True)
text_emb = self._clip.encode([audio_transcript], normalize_embeddings=True)
similarity = float(np.dot(image_emb[0], text_emb[0]))
return {
"is_mismatch": similarity < 0.18,
"consistency_score": round(similarity, 4),
}
def check_ocr_vs_image(self, image: Union[str, bytes], ocr_text: str) -> Dict[str, Any]:
if not ocr_text:
return {"is_mismatch": False, "consistency_score": 0.0}
pil_image = self._load_image_for_clip(image)
image_emb = self._clip.encode([pil_image], normalize_embeddings=True)
text_emb = self._clip.encode([ocr_text], normalize_embeddings=True)
similarity = float(np.dot(image_emb[0], text_emb[0]))
return {
"is_mismatch": similarity < 0.18,
"consistency_score": round(similarity, 4),
}
def check_caption_alignment(self, image: Union[str, bytes], ocr_text: str) -> Dict[str, Any]:
if not ocr_text:
return {"caption": "", "alignment_score": 0.0}
pil_image = self._load_image_for_clip(image)
processor, model, device = self._get_captioner()
inputs = processor(images=pil_image, return_tensors="pt").to(device)
output_ids = model.generate(**inputs, max_new_tokens=30)
caption = processor.decode(output_ids[0], skip_special_tokens=True)
text_emb = self._clip.encode([ocr_text], normalize_embeddings=True)
caption_emb = self._clip.encode([caption], normalize_embeddings=True)
similarity = float(np.dot(text_emb[0], caption_emb[0]))
return {"caption": caption, "alignment_score": round(similarity, 4)}
class VisualSecurityEngine:
def __init__(
self,
use_onnx: bool | None = None,
force_cpu: bool | None = None,
clip_model: str | None = None,
) -> None:
self.engine_d = PromptInjectionEngine(use_onnx=use_onnx, force_cpu=force_cpu)
self.engine_e = CrossModalEngine(clip_model=clip_model)
def extract_text(self, image: Union[str, bytes]) -> Dict[str, Any]:
return self.engine_d.extract_text(image)
def detect_injection(self, image: Union[str, bytes]) -> Dict[str, Any]:
return self.engine_d.detect_injection(image)
def detect_injection_from_text(
self, normalized_text: str, matched_phrases: List[str] | None = None
) -> Dict[str, Any]:
return self.engine_d.detect_injection_from_text(normalized_text, matched_phrases)
def check_cross_modal(self, image: Union[str, bytes], audio_transcript: str) -> Dict[str, Any]:
return self.engine_e.check_cross_modal(image, audio_transcript)
def check_ocr_vs_image(self, image: Union[str, bytes], ocr_text: str) -> Dict[str, Any]:
return self.engine_e.check_ocr_vs_image(image, ocr_text)
def check_caption_alignment(self, image: Union[str, bytes], ocr_text: str) -> Dict[str, Any]:
return self.engine_e.check_caption_alignment(image, ocr_text)
def _download_demo_image() -> bytes:
demo_urls = [
"https://upload.wikimedia.org/wikipedia/commons/3/3a/Cat03.jpg",
"https://upload.wikimedia.org/wikipedia/commons/7/74/A-Cat.jpg",
]
headers = {"User-Agent": "Mozilla/5.0 (Sentinel-X demo)"}
last_error: Exception | None = None
for url in demo_urls:
try:
request = urllib.request.Request(url, headers=headers)
with urllib.request.urlopen(request, timeout=20) as response:
return response.read()
except Exception as exc: # pragma: no cover - best effort demo download
last_error = exc
continue
raise RuntimeError(f"Failed to download demo image: {last_error}")
if __name__ == "__main__":
demo_bytes = _download_demo_image()
engine = VisualSecurityEngine()
injection_result = engine.detect_injection(demo_bytes)
cross_modal_result = engine.check_cross_modal(demo_bytes, "a cat sitting on a ledge")
print("Injection detection:", injection_result)
print("Cross-modal consistency:", cross_modal_result)
|