File size: 12,783 Bytes
c624cb8 | 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 | """
Mutation Engine
Main engine for prompt mutation with:
- Strategy selection and execution
- Multi-hop mutation
- Diversity scoring
- Lineage tracking
- Reproducibility controls
"""
import hashlib
import random
from typing import Any, Dict, List, Optional
from agents.mutation.diversity import DiversityScorer, get_diversity_scorer
from agents.mutation.registry import get_mutation_strategy, list_mutation_strategies
from agents.mutation.schemas import (
MutationLog,
MutationRequest,
MutationResponse,
)
from backend.logging.logger import get_logger
# Default strategies for mutation
DEFAULT_STRATEGIES = [
"synonym_replacement",
"paraphrase",
"role_swap",
"context_obfuscation",
]
class MutationEngine:
"""
Main mutation engine for prompt mutation.
Handles:
- Strategy selection and execution
- Multi-hop mutation
- Diversity scoring
- Lineage tracking
- Reproducibility via deterministic seeding
"""
def __init__(
self,
embedding_model: str = "all-MiniLM-L6-v2",
min_diversity_threshold: float = 0.1,
min_similarity_threshold: float = 0.5,
max_retries: int = 3
):
"""
Initialize the mutation engine.
Args:
embedding_model: Model for diversity scoring
min_diversity_threshold: Minimum diversity to accept
min_similarity_threshold: Minimum similarity to preserve intent
max_retries: Maximum retries for low diversity
"""
self.logger = get_logger(__name__)
self._diversity_scorer: Optional[DiversityScorer] = None
self._embedding_model_name = embedding_model
self._min_diversity_threshold = min_diversity_threshold
self._min_similarity_threshold = min_similarity_threshold
self._max_retries = max_retries
@property
def diversity_scorer(self) -> DiversityScorer:
"""Lazy load the diversity scorer."""
if self._diversity_scorer is None:
self._diversity_scorer = get_diversity_scorer()
return self._diversity_scorer
def _compute_seed(
self,
run_id: str,
sample_id: str,
attack_type: str,
depth: int = 0
) -> int:
"""
Compute deterministic seed for reproducibility.
seed = hash(run_id + sample_id + attack_type + depth)
Args:
run_id: Run identifier
sample_id: Sample identifier
attack_type: Attack type
depth: Mutation depth
Returns:
Deterministic seed
"""
hash_input = f"{run_id}{sample_id}{attack_type}{depth}"
hash_bytes = hashlib.sha256(hash_input.encode()).digest()
return int.from_bytes(hash_bytes[:4], byteorder="big")
def _select_strategies(
self,
mutation_depth: int,
seed: int,
attack_type: str
) -> List[str]:
"""
Select strategies for mutation based on depth.
Args:
mutation_depth: Number of mutations
seed: Random seed
attack_type: Type of attack
Returns:
List of strategy names
"""
random.seed(seed)
strategies = []
available = DEFAULT_STRATEGIES.copy()
for i in range(mutation_depth):
if not available:
available = DEFAULT_STRATEGIES.copy()
# Select strategy
strategy = random.choice(available)
strategies.append(strategy)
# Remove to avoid repeats (unless we want multi-hop to repeat)
available.remove(strategy)
return strategies
def _validate_mutation(
self,
base_prompt: str,
mutated_prompt: str
) -> tuple[bool, str, float, float]:
"""
Validate that a mutation preserves attack intent.
Args:
base_prompt: Original prompt
mutated_prompt: Mutated prompt
Returns:
Tuple of (is_valid, reason, diversity, similarity)
"""
diversity, similarity = self.diversity_scorer.compute_step_diversity(
base_prompt,
mutated_prompt
)
if diversity < self._min_diversity_threshold:
return False, f"Diversity {diversity:.3f} below threshold", diversity, similarity
if similarity < self._min_similarity_threshold:
return False, f"Similarity {similarity:.3f} below threshold", diversity, similarity
return True, "Valid", diversity, similarity
async def mutate(self, request: MutationRequest) -> MutationResponse:
"""
Execute prompt mutation based on the request.
Args:
request: Mutation request with parameters
Returns:
Mutation response with mutated prompt and metadata
"""
self.logger.info(
"Executing mutation",
run_id=str(request.run_id),
sample_id=request.sample_id,
attack_type=request.attack_type,
mutation_depth=request.mutation_depth
)
try:
# Compute deterministic seed
if request.random_seed is not None:
base_seed = request.random_seed
else:
base_seed = self._compute_seed(
str(request.run_id),
request.sample_id,
request.attack_type
)
# Select strategies
strategies = self._select_strategies(
request.mutation_depth,
base_seed,
request.attack_type
)
# Apply mutations
mutated_prompt = request.base_prompt
mutation_trace: List[str] = []
prompt_history = [request.base_prompt]
for depth in range(request.mutation_depth):
seed = self._compute_seed(
str(request.run_id),
request.sample_id,
request.attack_type,
depth
)
# Get strategy
strategy = get_mutation_strategy(strategies[depth])
if strategy is None:
self.logger.warning(
"Strategy not found, using default",
strategy=strategies[depth]
)
strategy = get_mutation_strategy("synonym_replacement")
# Apply strategy
mutated_prompt = strategy.apply(mutated_prompt, seed)
mutation_trace.append(strategy.name)
prompt_history.append(mutated_prompt)
# Validate mutation
is_valid, reason, diversity, similarity = self._validate_mutation(
request.base_prompt,
mutated_prompt
)
# Compute cumulative diversity
cumulative_diversity = self.diversity_scorer.compute_cumulative_diversity(
prompt_history
)
# Build metadata
mutation_metadata: Dict[str, Any] = {
"strategies_used": strategies,
"cumulative_diversity": cumulative_diversity,
"final_similarity": similarity,
"is_valid": is_valid,
"validation_reason": reason,
"seed_used": base_seed,
"prompt_history": prompt_history,
}
# Log if validation failed
if not is_valid:
self.logger.warning(
"Mutation validation failed",
run_id=str(request.run_id),
sample_id=request.sample_id,
reason=reason,
diversity=diversity,
similarity=similarity
)
# Log mutation
self._log_mutation(
run_id=str(request.run_id),
sample_id=request.sample_id,
attack_type=request.attack_type,
mutation_depth=request.mutation_depth,
strategies_used=mutation_trace,
diversity_score=diversity,
cumulative_diversity=cumulative_diversity,
success=True
)
return MutationResponse(
mutated_prompt=mutated_prompt,
mutation_trace=mutation_trace,
diversity_score=diversity,
cumulative_diversity=cumulative_diversity,
mutation_depth=request.mutation_depth,
mutation_metadata=mutation_metadata,
run_id=request.run_id,
sample_id=request.sample_id
)
except Exception as e:
self.logger.error(
"Mutation execution failed",
run_id=str(request.run_id),
sample_id=request.sample_id,
error=str(e)
)
# Log failure
self._log_mutation(
run_id=str(request.run_id),
sample_id=request.sample_id,
attack_type=request.attack_type,
mutation_depth=request.mutation_depth,
strategies_used=[],
diversity_score=0.0,
cumulative_diversity=0.0,
success=False,
error=str(e)
)
raise
def _log_mutation(
self,
run_id: str,
sample_id: str,
attack_type: str,
mutation_depth: int,
strategies_used: List[str],
diversity_score: float,
cumulative_diversity: float,
success: bool,
error: Optional[str] = None
) -> None:
"""
Log mutation execution details.
Args:
run_id: Run identifier
sample_id: Sample identifier
attack_type: Attack type
mutation_depth: Depth of mutation
strategies_used: List of strategies applied
diversity_score: Final diversity score
cumulative_diversity: Cumulative diversity
success: Whether mutation succeeded
error: Error message if failed
"""
log_data = {
"run_id": run_id,
"sample_id": sample_id,
"attack_type": attack_type,
"mutation_depth": mutation_depth,
"strategies_used": strategies_used,
"diversity_score": diversity_score,
"cumulative_diversity": cumulative_diversity,
"success": success,
"error": error
}
if success:
if diversity_score < self._min_diversity_threshold:
self.logger.warning("Mutation diversity below threshold", **log_data)
else:
self.logger.info("Mutation executed successfully", **log_data)
else:
self.logger.error("Mutation execution failed", **log_data)
def get_available_strategies(self) -> List[str]:
"""
Get list of available mutation strategies.
Returns:
List of strategy names
"""
return list_mutation_strategies()
def clear_cache(self) -> None:
"""Clear the diversity scorer cache."""
if self._diversity_scorer:
self._diversity_scorer.clear_cache()
self.logger.info("Mutation engine cache cleared")
# Global engine instance
_mutation_engine: Optional[MutationEngine] = None
def get_mutation_engine() -> MutationEngine:
"""
Get the global mutation engine instance.
Returns:
MutationEngine singleton
"""
global _mutation_engine
if _mutation_engine is None:
_mutation_engine = MutationEngine()
return _mutation_engine
__all__ = [
"MutationEngine",
"get_mutation_engine",
"DEFAULT_STRATEGIES",
]
|