Upload pii_extension.py with huggingface_hub
Browse files- pii_extension.py +115 -12
pii_extension.py
CHANGED
|
@@ -18,6 +18,93 @@ class PIILabel(Enum):
|
|
| 18 |
SOCIAL_MEDIA = "social_media"
|
| 19 |
URL = "url"
|
| 20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
class PIIDetector:
|
| 22 |
"""Detect PII in text with context awareness"""
|
| 23 |
|
|
@@ -168,6 +255,7 @@ class PIIDetector:
|
|
| 168 |
def scan(self, text: str, age: int) -> Dict:
|
| 169 |
"""
|
| 170 |
Full PII scan with age-appropriate rules
|
|
|
|
| 171 |
|
| 172 |
Returns:
|
| 173 |
{
|
|
@@ -177,45 +265,53 @@ class PIIDetector:
|
|
| 177 |
"social_media_allowed": bool,
|
| 178 |
"grooming_risk": float,
|
| 179 |
"action": "allow" | "block" | "flag",
|
| 180 |
-
"reason": str
|
|
|
|
|
|
|
| 181 |
}
|
| 182 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 183 |
pii_found = []
|
| 184 |
pii_types = set()
|
| 185 |
|
| 186 |
-
# Detect various PII types
|
| 187 |
-
emails = self.detect_emails(
|
| 188 |
if emails:
|
| 189 |
pii_types.add(PIILabel.EMAIL)
|
| 190 |
for email, start, end in emails:
|
| 191 |
pii_found.append({"type": "email", "value": email, "start": start, "end": end})
|
| 192 |
|
| 193 |
-
phones = self.detect_phones(
|
| 194 |
if phones:
|
| 195 |
pii_types.add(PIILabel.PHONE)
|
| 196 |
for phone, start, end in phones:
|
| 197 |
pii_found.append({"type": "phone", "value": phone, "start": start, "end": end})
|
| 198 |
|
| 199 |
-
addresses = self.detect_addresses(
|
| 200 |
if addresses:
|
| 201 |
pii_types.add(PIILabel.ADDRESS)
|
| 202 |
for addr, start, end in addresses:
|
| 203 |
pii_found.append({"type": "address", "value": addr, "start": start, "end": end})
|
| 204 |
|
| 205 |
-
credit_cards = self.detect_credit_cards(
|
| 206 |
if credit_cards:
|
| 207 |
pii_types.add(PIILabel.CREDIT_CARD)
|
| 208 |
for cc, start, end in credit_cards:
|
| 209 |
pii_found.append({"type": "credit_card", "value": cc, "start": start, "end": end})
|
| 210 |
|
| 211 |
-
ssns = self.detect_ssn(
|
| 212 |
if ssns:
|
| 213 |
pii_types.add(PIILabel.SSN)
|
| 214 |
for ssn, start, end in ssns:
|
| 215 |
pii_found.append({"type": "ssn", "value": ssn, "start": start, "end": end})
|
| 216 |
|
| 217 |
-
# Social media detection
|
| 218 |
-
social_links = self.detect_social_media(
|
| 219 |
has_social_media = len(social_links) > 0
|
| 220 |
|
| 221 |
if has_social_media:
|
|
@@ -234,8 +330,8 @@ class PIIDetector:
|
|
| 234 |
action = "block"
|
| 235 |
reason = f"PII detected: {', '.join([p.value for p in critical_pii])}"
|
| 236 |
elif has_social_media:
|
| 237 |
-
# Social media rules
|
| 238 |
-
is_grooming, grooming_risk, grooming_keywords = self.detect_grooming_context(
|
| 239 |
|
| 240 |
if age < 13:
|
| 241 |
# Under 13: Block ALL social media sharing
|
|
@@ -261,6 +357,10 @@ class PIIDetector:
|
|
| 261 |
elif grooming_risk > 0:
|
| 262 |
social_media_allowed = False
|
| 263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 264 |
return {
|
| 265 |
"has_pii": len(pii_types) > 0,
|
| 266 |
"pii_types": [p.value for p in pii_types],
|
|
@@ -270,7 +370,10 @@ class PIIDetector:
|
|
| 270 |
"grooming_keywords": grooming_keywords,
|
| 271 |
"action": action,
|
| 272 |
"reason": reason,
|
| 273 |
-
"age": age
|
|
|
|
|
|
|
|
|
|
| 274 |
}
|
| 275 |
|
| 276 |
|
|
|
|
| 18 |
SOCIAL_MEDIA = "social_media"
|
| 19 |
URL = "url"
|
| 20 |
|
| 21 |
+
class UnicodeDeobfuscator:
|
| 22 |
+
"""Detect and normalize unicode obfuscation attempts"""
|
| 23 |
+
|
| 24 |
+
# Unicode ranges for suspicious characters
|
| 25 |
+
CIRCLED_LETTERS = range(0x24B6, 0x24EA) # βΆ-β©
|
| 26 |
+
MATHEMATICAL_CHARS = range(0x1D400, 0x1D800) # π-π, etc
|
| 27 |
+
FULLWIDTH_CHARS = range(0xFF01, 0xFF5F) # οΌ-ο½
|
| 28 |
+
DOUBLE_STRUCK = range(0x2100, 0x2150) # β, β, etc
|
| 29 |
+
BOX_DRAWING = range(0x2500, 0x2580) # βββ etc
|
| 30 |
+
BLOCK_ELEMENTS = range(0x2580, 0x25A0) # β-β
|
| 31 |
+
|
| 32 |
+
# Mapping of circled letters to normal
|
| 33 |
+
CIRCLED_MAP = {
|
| 34 |
+
# Uppercase
|
| 35 |
+
'βΆ': 'A', 'β·': 'B', 'βΈ': 'C', 'βΉ': 'D', 'βΊ': 'E',
|
| 36 |
+
'β»': 'F', 'βΌ': 'G', 'β½': 'H', 'βΎ': 'I', 'βΏ': 'J',
|
| 37 |
+
'β': 'K', 'β': 'L', 'β': 'M', 'β': 'N', 'β': 'O',
|
| 38 |
+
'β
': 'P', 'β': 'Q', 'β': 'R', 'β': 'S', 'β': 'T',
|
| 39 |
+
'β': 'U', 'β': 'V', 'β': 'W', 'β': 'X', 'β': 'Y', 'β': 'Z',
|
| 40 |
+
# Lowercase
|
| 41 |
+
'β': 'a', 'β': 'b', 'β': 'c', 'β': 'd', 'β': 'e',
|
| 42 |
+
'β': 'f', 'β': 'g', 'β': 'h', 'β': 'i', 'β': 'j',
|
| 43 |
+
'β': 'k', 'β': 'l', 'β': 'm', 'β': 'n', 'β': 'o',
|
| 44 |
+
'β': 'p', 'β ': 'q', 'β‘': 'r', 'β’': 's', 'β£': 't',
|
| 45 |
+
'β€': 'u', 'β₯': 'v', 'β¦': 'w', 'β§': 'x', 'β¨': 'y', 'β©': 'z',
|
| 46 |
+
}
|
| 47 |
+
|
| 48 |
+
@classmethod
|
| 49 |
+
def detect_obfuscation(cls, text: str) -> Tuple[bool, List[Tuple[str, str]], str]:
|
| 50 |
+
"""
|
| 51 |
+
Detect unicode obfuscation
|
| 52 |
+
Returns: (is_obfuscated, [(char, type)], normalized_text)
|
| 53 |
+
"""
|
| 54 |
+
suspicious = []
|
| 55 |
+
normalized = []
|
| 56 |
+
|
| 57 |
+
for char in text:
|
| 58 |
+
code = ord(char)
|
| 59 |
+
|
| 60 |
+
# Check circled letters
|
| 61 |
+
if char in cls.CIRCLED_MAP:
|
| 62 |
+
suspicious.append((char, 'circled'))
|
| 63 |
+
normalized.append(cls.CIRCLED_MAP[char])
|
| 64 |
+
# Check double-struck
|
| 65 |
+
elif code in cls.DOUBLE_STRUCK:
|
| 66 |
+
suspicious.append((char, 'double-struck'))
|
| 67 |
+
# Map common double-struck to normal
|
| 68 |
+
if char == 'β':
|
| 69 |
+
normalized.append('C')
|
| 70 |
+
elif char == 'β':
|
| 71 |
+
normalized.append('H')
|
| 72 |
+
elif char == 'β':
|
| 73 |
+
normalized.append('N')
|
| 74 |
+
elif char == 'β':
|
| 75 |
+
normalized.append('P')
|
| 76 |
+
elif char == 'β':
|
| 77 |
+
normalized.append('Q')
|
| 78 |
+
elif char == 'β':
|
| 79 |
+
normalized.append('R')
|
| 80 |
+
elif char == 'β€':
|
| 81 |
+
normalized.append('Z')
|
| 82 |
+
else:
|
| 83 |
+
normalized.append(char)
|
| 84 |
+
# Check fullwidth
|
| 85 |
+
elif code in cls.FULLWIDTH_CHARS:
|
| 86 |
+
suspicious.append((char, 'fullwidth'))
|
| 87 |
+
# Convert to normal ASCII
|
| 88 |
+
normalized.append(chr(code - 0xFEE0))
|
| 89 |
+
# Check mathematical
|
| 90 |
+
elif code in cls.MATHEMATICAL_CHARS:
|
| 91 |
+
suspicious.append((char, 'mathematical'))
|
| 92 |
+
normalized.append(char) # Keep as-is for now
|
| 93 |
+
else:
|
| 94 |
+
normalized.append(char)
|
| 95 |
+
|
| 96 |
+
is_obfuscated = len(suspicious) > 0
|
| 97 |
+
normalized_text = ''.join(normalized)
|
| 98 |
+
|
| 99 |
+
return is_obfuscated, suspicious, normalized_text
|
| 100 |
+
|
| 101 |
+
@classmethod
|
| 102 |
+
def normalize(cls, text: str) -> str:
|
| 103 |
+
"""Quick normalize without detection details"""
|
| 104 |
+
_, _, normalized = cls.detect_obfuscation(text)
|
| 105 |
+
return normalized
|
| 106 |
+
|
| 107 |
+
|
| 108 |
class PIIDetector:
|
| 109 |
"""Detect PII in text with context awareness"""
|
| 110 |
|
|
|
|
| 255 |
def scan(self, text: str, age: int) -> Dict:
|
| 256 |
"""
|
| 257 |
Full PII scan with age-appropriate rules
|
| 258 |
+
Also detects unicode obfuscation
|
| 259 |
|
| 260 |
Returns:
|
| 261 |
{
|
|
|
|
| 265 |
"social_media_allowed": bool,
|
| 266 |
"grooming_risk": float,
|
| 267 |
"action": "allow" | "block" | "flag",
|
| 268 |
+
"reason": str,
|
| 269 |
+
"obfuscation_detected": bool,
|
| 270 |
+
"normalized_text": str
|
| 271 |
}
|
| 272 |
"""
|
| 273 |
+
# Step 0: Detect unicode obfuscation
|
| 274 |
+
is_obfuscated, suspicious_chars, normalized_text = UnicodeDeobfuscator.detect_obfuscation(text)
|
| 275 |
+
|
| 276 |
+
# Use normalized text for detection if obfuscated
|
| 277 |
+
detection_text = normalized_text if is_obfuscated else text
|
| 278 |
+
|
| 279 |
pii_found = []
|
| 280 |
pii_types = set()
|
| 281 |
|
| 282 |
+
# Detect various PII types (using normalized text if obfuscated)
|
| 283 |
+
emails = self.detect_emails(detection_text)
|
| 284 |
if emails:
|
| 285 |
pii_types.add(PIILabel.EMAIL)
|
| 286 |
for email, start, end in emails:
|
| 287 |
pii_found.append({"type": "email", "value": email, "start": start, "end": end})
|
| 288 |
|
| 289 |
+
phones = self.detect_phones(detection_text)
|
| 290 |
if phones:
|
| 291 |
pii_types.add(PIILabel.PHONE)
|
| 292 |
for phone, start, end in phones:
|
| 293 |
pii_found.append({"type": "phone", "value": phone, "start": start, "end": end})
|
| 294 |
|
| 295 |
+
addresses = self.detect_addresses(detection_text)
|
| 296 |
if addresses:
|
| 297 |
pii_types.add(PIILabel.ADDRESS)
|
| 298 |
for addr, start, end in addresses:
|
| 299 |
pii_found.append({"type": "address", "value": addr, "start": start, "end": end})
|
| 300 |
|
| 301 |
+
credit_cards = self.detect_credit_cards(detection_text)
|
| 302 |
if credit_cards:
|
| 303 |
pii_types.add(PIILabel.CREDIT_CARD)
|
| 304 |
for cc, start, end in credit_cards:
|
| 305 |
pii_found.append({"type": "credit_card", "value": cc, "start": start, "end": end})
|
| 306 |
|
| 307 |
+
ssns = self.detect_ssn(detection_text)
|
| 308 |
if ssns:
|
| 309 |
pii_types.add(PIILabel.SSN)
|
| 310 |
for ssn, start, end in ssns:
|
| 311 |
pii_found.append({"type": "ssn", "value": ssn, "start": start, "end": end})
|
| 312 |
|
| 313 |
+
# Social media detection (also on normalized text)
|
| 314 |
+
social_links = self.detect_social_media(detection_text)
|
| 315 |
has_social_media = len(social_links) > 0
|
| 316 |
|
| 317 |
if has_social_media:
|
|
|
|
| 330 |
action = "block"
|
| 331 |
reason = f"PII detected: {', '.join([p.value for p in critical_pii])}"
|
| 332 |
elif has_social_media:
|
| 333 |
+
# Social media rules (use normalized text for grooming detection)
|
| 334 |
+
is_grooming, grooming_risk, grooming_keywords = self.detect_grooming_context(detection_text)
|
| 335 |
|
| 336 |
if age < 13:
|
| 337 |
# Under 13: Block ALL social media sharing
|
|
|
|
| 357 |
elif grooming_risk > 0:
|
| 358 |
social_media_allowed = False
|
| 359 |
|
| 360 |
+
# Add obfuscation info to reason if detected
|
| 361 |
+
if is_obfuscated and action == "allow":
|
| 362 |
+
reason = f"Unicode obfuscation detected and normalized. {reason}"
|
| 363 |
+
|
| 364 |
return {
|
| 365 |
"has_pii": len(pii_types) > 0,
|
| 366 |
"pii_types": [p.value for p in pii_types],
|
|
|
|
| 370 |
"grooming_keywords": grooming_keywords,
|
| 371 |
"action": action,
|
| 372 |
"reason": reason,
|
| 373 |
+
"age": age,
|
| 374 |
+
"obfuscation_detected": is_obfuscated,
|
| 375 |
+
"obfuscation_chars": [(c, t) for c, t in suspicious_chars] if is_obfuscated else [],
|
| 376 |
+
"normalized_text": normalized_text if is_obfuscated else text
|
| 377 |
}
|
| 378 |
|
| 379 |
|