Spaces:
Sleeping
Sleeping
File size: 12,707 Bytes
4afcb3a | 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 | """
adversarial_detector.py
========================
Detects adversarial / anomalous inputs that may be crafted to manipulate
AI models or evade safety filters.
Detection layers (all zero-dependency except the optional embedding layer):
1. Token-length analysis — unusually long or repetitive prompts
2. Character distribution — abnormal char class ratios (unicode tricks, homoglyphs)
3. Repetition detection — token/n-gram flooding
4. Encoding obfuscation — base64 blobs, hex strings, ROT-13 traces
5. Statistical anomaly — entropy, symbol density, whitespace abuse
6. Embedding outlier — cosine distance from "normal" centroid (optional)
"""
from __future__ import annotations
import re
import math
import time
import unicodedata
import logging
from collections import Counter
from dataclasses import dataclass, field
from typing import List, Optional
logger = logging.getLogger("ai_firewall.adversarial_detector")
# ---------------------------------------------------------------------------
# Config defaults (tunable without subclassing)
# ---------------------------------------------------------------------------
DEFAULT_CONFIG = {
"max_token_length": 4096, # chars (rough token proxy)
"max_word_count": 800,
"max_line_count": 200,
"repetition_threshold": 0.45, # fraction of repeated trigrams → adversarial
"entropy_min": 2.5, # too-low entropy = repetitive junk
"entropy_max": 5.8, # too-high entropy = encoded/random content
"symbol_density_max": 0.35, # fraction of non-alphanumeric chars
"unicode_escape_threshold": 5, # count of \uXXXX / \xXX sequences
"base64_min_length": 40, # minimum length of candidate b64 blocks
"homoglyph_threshold": 3, # count of confusable lookalike chars
}
# Homoglyph mapping (Cyrillic / Greek / other confusable lookalikes for latin)
_HOMOGLYPH_MAP = {
"а": "a", "е": "e", "і": "i", "о": "o", "р": "p", "с": "c",
"х": "x", "у": "y", "ѕ": "s", "ј": "j", "ԁ": "d", "ɡ": "g",
"ʜ": "h", "ᴛ": "t", "ᴡ": "w", "ᴍ": "m", "ᴋ": "k",
"α": "a", "ε": "e", "ο": "o", "ρ": "p", "ν": "v", "κ": "k",
}
_BASE64_RE = re.compile(r"(?:[A-Za-z0-9+/]{4}){10,}(?:[A-Za-z0-9+/]{2}==|[A-Za-z0-9+/]{3}=)?")
_HEX_RE = re.compile(r"(?:0x)?[0-9a-fA-F]{16,}")
_UNICODE_ESC_RE = re.compile(r"(\\u[0-9a-fA-F]{4}|\\x[0-9a-fA-F]{2}|%[0-9a-fA-F]{2})")
@dataclass
class AdversarialResult:
is_adversarial: bool
risk_score: float # 0.0 – 1.0
flags: List[str] = field(default_factory=list)
details: dict = field(default_factory=dict)
latency_ms: float = 0.0
def to_dict(self) -> dict:
return {
"is_adversarial": self.is_adversarial,
"risk_score": round(self.risk_score, 4),
"flags": self.flags,
"details": self.details,
"latency_ms": round(self.latency_ms, 2),
}
class AdversarialDetector:
"""
Stateless adversarial input detector.
A prompt is considered adversarial if its aggregate risk score
exceeds `threshold` (default 0.55).
Parameters
----------
threshold : float
Risk score above which input is flagged.
config : dict, optional
Override any key from DEFAULT_CONFIG.
use_embeddings : bool
Enable embedding-outlier detection (requires sentence-transformers).
embedding_model : str
Model name for the embedding layer.
"""
def __init__(
self,
threshold: float = 0.55,
config: Optional[dict] = None,
use_embeddings: bool = False,
embedding_model: str = "all-MiniLM-L6-v2",
) -> None:
self.threshold = threshold
self.cfg = {**DEFAULT_CONFIG, **(config or {})}
self.use_embeddings = use_embeddings
self._embedder = None
self._normal_centroid = None # set via `fit_normal_distribution`
if use_embeddings:
self._load_embedder(embedding_model)
# ------------------------------------------------------------------
# Embedding layer
# ------------------------------------------------------------------
def _load_embedder(self, model_name: str) -> None:
try:
from sentence_transformers import SentenceTransformer
import numpy as np
self._embedder = SentenceTransformer(model_name)
logger.info("Adversarial embedding layer loaded: %s", model_name)
except ImportError:
logger.warning("sentence-transformers not installed — embedding outlier layer disabled.")
self.use_embeddings = False
def fit_normal_distribution(self, normal_prompts: List[str]) -> None:
"""
Compute the centroid of embedding vectors for a set of known-good
prompts. Call this once at startup with representative benign prompts.
"""
if not self.use_embeddings or self._embedder is None:
return
import numpy as np
embeddings = self._embedder.encode(normal_prompts, convert_to_numpy=True, normalize_embeddings=True)
self._normal_centroid = embeddings.mean(axis=0)
self._normal_centroid /= np.linalg.norm(self._normal_centroid)
logger.info("Normal centroid computed from %d prompts.", len(normal_prompts))
# ------------------------------------------------------------------
# Individual checks
# ------------------------------------------------------------------
def _check_length(self, text: str) -> tuple[float, str, dict]:
char_len = len(text)
word_count = len(text.split())
line_count = text.count("\n")
score = 0.0
details, flags = {}, []
if char_len > self.cfg["max_token_length"]:
score += 0.4
flags.append("excessive_length")
if word_count > self.cfg["max_word_count"]:
score += 0.25
flags.append("excessive_word_count")
if line_count > self.cfg["max_line_count"]:
score += 0.2
flags.append("excessive_line_count")
details = {"char_len": char_len, "word_count": word_count, "line_count": line_count}
return min(score, 1.0), "|".join(flags), details
def _check_repetition(self, text: str) -> tuple[float, str, dict]:
words = text.lower().split()
if len(words) < 6:
return 0.0, "", {}
trigrams = [tuple(words[i:i+3]) for i in range(len(words) - 2)]
if not trigrams:
return 0.0, "", {}
total = len(trigrams)
unique = len(set(trigrams))
repetition_ratio = 1.0 - (unique / total)
score = 0.0
flag = ""
if repetition_ratio >= self.cfg["repetition_threshold"]:
score = min(repetition_ratio, 1.0)
flag = "high_token_repetition"
return score, flag, {"repetition_ratio": round(repetition_ratio, 3)}
def _check_entropy(self, text: str) -> tuple[float, str, dict]:
if not text:
return 0.0, "", {}
freq = Counter(text)
total = len(text)
entropy = -sum((c / total) * math.log2(c / total) for c in freq.values())
score = 0.0
flag = ""
if entropy < self.cfg["entropy_min"]:
score = 0.5
flag = "low_entropy_repetitive"
elif entropy > self.cfg["entropy_max"]:
score = 0.6
flag = "high_entropy_possibly_encoded"
return score, flag, {"entropy": round(entropy, 3)}
def _check_symbol_density(self, text: str) -> tuple[float, str, dict]:
if not text:
return 0.0, "", {}
non_alnum = sum(1 for c in text if not c.isalnum() and not c.isspace())
density = non_alnum / len(text)
score = 0.0
flag = ""
if density > self.cfg["symbol_density_max"]:
score = min(density, 1.0)
flag = "high_symbol_density"
return score, flag, {"symbol_density": round(density, 3)}
def _check_encoding_obfuscation(self, text: str) -> tuple[float, str, dict]:
score = 0.0
flags = []
details = {}
# Unicode escape sequences
esc_matches = _UNICODE_ESC_RE.findall(text)
if len(esc_matches) >= self.cfg["unicode_escape_threshold"]:
score += 0.5
flags.append("unicode_escape_sequences")
details["unicode_escapes"] = len(esc_matches)
# Base64-like blobs
b64_matches = _BASE64_RE.findall(text)
if b64_matches:
score += 0.4
flags.append("base64_like_content")
details["base64_blocks"] = len(b64_matches)
# Long hex strings
hex_matches = _HEX_RE.findall(text)
if hex_matches:
score += 0.3
flags.append("hex_encoded_content")
details["hex_blocks"] = len(hex_matches)
return min(score, 1.0), "|".join(flags), details
def _check_homoglyphs(self, text: str) -> tuple[float, str, dict]:
count = sum(1 for ch in text if ch in _HOMOGLYPH_MAP)
score = 0.0
flag = ""
if count >= self.cfg["homoglyph_threshold"]:
score = min(count / 20, 1.0)
flag = "homoglyph_substitution"
return score, flag, {"homoglyph_count": count}
def _check_unicode_normalization(self, text: str) -> tuple[float, str, dict]:
"""Detect invisible / zero-width / direction-override characters."""
bad_categories = {"Cf", "Cs", "Co"} # format, surrogate, private-use
bad_chars = [c for c in text if unicodedata.category(c) in bad_categories]
score = 0.0
flag = ""
if len(bad_chars) > 2:
score = min(len(bad_chars) / 10, 1.0)
flag = "invisible_unicode_chars"
return score, flag, {"invisible_char_count": len(bad_chars)}
def _check_embedding_outlier(self, text: str) -> tuple[float, str, dict]:
if not self.use_embeddings or self._embedder is None or self._normal_centroid is None:
return 0.0, "", {}
try:
import numpy as np
emb = self._embedder.encode(text, convert_to_numpy=True, normalize_embeddings=True)
similarity = float(emb @ self._normal_centroid)
distance = 1.0 - similarity # 0 = identical to normal, 1 = orthogonal
score = max(0.0, (distance - 0.3) / 0.7) # linear rescale [0.3, 1.0] → [0, 1]
flag = "embedding_outlier" if score > 0.3 else ""
return score, flag, {"centroid_distance": round(distance, 4)}
except Exception as exc:
logger.debug("Embedding outlier check failed: %s", exc)
return 0.0, "", {}
# ------------------------------------------------------------------
# Aggregation
# ------------------------------------------------------------------
def detect(self, text: str) -> AdversarialResult:
"""
Run all detection layers and return an AdversarialResult.
Parameters
----------
text : str
Raw user prompt.
"""
t0 = time.perf_counter()
checks = [
self._check_length(text),
self._check_repetition(text),
self._check_entropy(text),
self._check_symbol_density(text),
self._check_encoding_obfuscation(text),
self._check_homoglyphs(text),
self._check_unicode_normalization(text),
self._check_embedding_outlier(text),
]
aggregate_score = 0.0
all_flags: List[str] = []
all_details: dict = {}
weights = [0.15, 0.20, 0.15, 0.10, 0.20, 0.10, 0.10, 0.20] # sum > 1 ok; normalised below
weight_sum = sum(weights)
for (score, flag, details), weight in zip(checks, weights):
aggregate_score += score * weight
if flag:
all_flags.extend(flag.split("|"))
all_details.update(details)
risk_score = min(aggregate_score / weight_sum, 1.0)
is_adversarial = risk_score >= self.threshold
latency = (time.perf_counter() - t0) * 1000
result = AdversarialResult(
is_adversarial=is_adversarial,
risk_score=risk_score,
flags=list(filter(None, all_flags)),
details=all_details,
latency_ms=latency,
)
if is_adversarial:
logger.warning("Adversarial input detected | score=%.3f flags=%s", risk_score, all_flags)
return result
def is_safe(self, text: str) -> bool:
return not self.detect(text).is_adversarial
|