Spaces:
Sleeping
Sleeping
File size: 6,765 Bytes
29428af |
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 |
"""
safety_classifier.py
This module integrates two red-teaming datasets:
1. romaingrx/red-teamer-mistral-
2. SummerSigh/Muti-Class-Redteaming
It provides:
- request safety classification
- output safety validation
- heuristic detectors based on adversarial patterns
- optional hooks for model-based classification
The goal is defensive: prevent harmful prompts or unsafe completions
from flowing through the APJ Threat Intelligence system.
"""
import re
from typing import Dict, List, Optional
from datasets import load_dataset
# ---------------------------------------------------------------------------
# Load datasets (shallow load only)
# ---------------------------------------------------------------------------
try:
DATA_REDTEAM_MISTRAL = load_dataset("romaingrx/red-teamer-mistral-", split="train")
except Exception:
DATA_REDTEAM_MISTRAL = None
try:
DATA_MULTI_REDT = load_dataset("SummerSigh/Muti-Class-Redteaming", split="train")
except Exception:
DATA_MULTI_REDT = None
# ---------------------------------------------------------------------------
# Extract patterns from datasets (heuristic layer)
# ---------------------------------------------------------------------------
def extract_lexical_patterns(dataset, text_key="text") -> List[str]:
"""Extract simple lexical redteam cues from dataset text."""
if not dataset:
return []
patterns = []
for ex in dataset:
txt = ex.get(text_key)
if not isinstance(txt, str):
continue
# Extract some keyword-like substrings
# We use lowercased lexical tokens as heuristic patterns.
tokens = re.findall(r"[a-zA-Z0-9_\-]{4,}", txt.lower())
patterns.extend(tokens)
# Deduplicate + filter noise
uniq = list({p for p in patterns if len(p) > 4})
return uniq[:5000] # cap to avoid overloading
# Build heuristic lexicons only once
LEX_REDTEAM_MISTRAL = extract_lexical_patterns(DATA_REDTEAM_MISTRAL)
LEX_REDTEAM_MULTI = extract_lexical_patterns(DATA_MULTI_REDT)
# Combined lexical bank
LEXICAL_BANK = set(LEX_REDTEAM_MISTRAL + LEX_REDTEAM_MULTI)
# ---------------------------------------------------------------------------
# Simple dangerous intent heuristic matcher
# ---------------------------------------------------------------------------
DANGEROUS_INTENTS = [
# Offensive cyber → blocked
r"(ddos|botnet|backdoor|rat|exploit|zero[- ]?day)",
r"(bypass|evasion|jailbreak|prompt[- ]?hack)",
r"(hack|breach|compromise|ransomware)",
r"(sql[- ]?injection|buffer[- ]?overflow)",
# Illicit markets (to prevent misuse of your system)
r"(buy|sell).*(credentials|logins|databases|access)",
r"(stolen|leaked).*(data|info)",
r"(carding|skimming|fraud)",
# Violence / harassment (red-teaming dataset content)
r"(kill|assault|harass|suicide|bomb|weapon)",
]
dangerous_intent_res = [re.compile(p, re.IGNORECASE) for p in DANGEROUS_INTENTS]
def detect_dangerous_intent(text: str) -> bool:
"""Return True if the text matches known dangerous intent patterns."""
for r in dangerous_intent_res:
if r.search(text):
return True
return False
# ---------------------------------------------------------------------------
# Lexical similarity heuristic
# ---------------------------------------------------------------------------
def heuristic_lexical_overlap(text: str, threshold: int = 5) -> bool:
"""
Check how many red-team tokens appear in the text.
If overlap exceeds threshold, classify as suspicious.
"""
if not text:
return False
tokens = set(re.findall(r"[a-zA-Z0-9_\-]{4,}", text.lower()))
overlap = tokens.intersection(LEXICAL_BANK)
return len(overlap) >= threshold
# ---------------------------------------------------------------------------
# Optional future ML classifier hooks (currently placeholder)
# ---------------------------------------------------------------------------
def ml_classify_request(text: str) -> Optional[str]:
"""
Placeholder for future ML classification using fine-tuned models.
Expected return values:
- "safe"
- "suspicious"
- "dangerous"
"""
return None
def ml_classify_output(text: str) -> Optional[str]:
"""Same as above—placeholder for model-based output safety filters."""
return None
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def safety_check(text: str) -> Dict[str, str]:
"""
Main safety gate for incoming user text.
Returns:
{
"blocked": True/False,
"reason": "...",
"level": "safe/suspicious/dangerous"
}
"""
t = (text or "").strip().lower()
# 1. ML classification (if implemented later)
ml = ml_classify_request(t)
if ml == "dangerous":
return {
"blocked": True,
"reason": "⚠️ ML safety classifier flagged this as dangerous.",
"level": "dangerous",
}
# 2. Dangerous intent patterns
if detect_dangerous_intent(t):
return {
"blocked": True,
"reason": "⚠️ Request blocked due to dangerous intent indicators.",
"level": "dangerous",
}
# 3. Lexical overlap heuristic
if heuristic_lexical_overlap(t):
return {
"blocked": False,
"reason": "⚠️ High lexical similarity to red-team prompts.",
"level": "suspicious",
}
return {
"blocked": False,
"reason": "Safe request.",
"level": "safe",
}
def safety_check_output(text: str) -> Dict[str, str]:
"""
Validate generated model output.
"""
t = (text or "").strip().lower()
# 1. ML classification (future)
ml = ml_classify_output(t)
if ml == "dangerous":
return {
"blocked": True,
"reason": "⚠️ Unsafe model output detected by classifier.",
"level": "dangerous",
}
# 2. Dangerous intent patterns
if detect_dangerous_intent(t):
return {
"blocked": True,
"reason": "⚠️ Model output contains dangerous intent content.",
"level": "dangerous",
}
# 3. Lexical overlap
if heuristic_lexical_overlap(t, threshold=8): # tighten for output
return {
"blocked": False,
"reason": "⚠️ Output resembles adversarial red-team patterns.",
"level": "suspicious",
}
return {
"blocked": False,
"reason": "Output appears safe.",
"level": "safe",
} |