File size: 25,052 Bytes
e4932aa |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 |
import re
import json
import os
from typing import List, Set, Dict, Tuple
from pathlib import Path
import pandas as pd
from dotenv import load_dotenv
# Import your CTI tools
from langchain.chat_models import init_chat_model
from langchain_tavily import TavilySearch
import sys
sys.path.append("src/agents/cti_agent")
from cti_tools import CTITools
from config import MODEL_NAME, CTI_SEARCH_CONFIG
class CTIToolsEvaluator:
"""Evaluator for CTI tools on CTIBench benchmarks."""
def __init__(self):
"""Initialize the evaluator with CTI tools."""
load_dotenv()
# Initialize LLM
self.llm = init_chat_model(MODEL_NAME, temperature=0.1)
# Initialize search (needed for CTITools init, even if not used in evaluation)
search_config = {**CTI_SEARCH_CONFIG, "api_key": os.getenv("TAVILY_API_KEY")}
self.cti_search = TavilySearch(**search_config)
# Initialize CTI Tools
self.cti_tools = CTITools(self.llm, self.cti_search)
# Storage for results
self.ate_results = []
self.taa_results = []
# ==================== CTI-ATE: MITRE Technique Extraction Tool ====================
def extract_technique_ids(self, text: str) -> Set[str]:
"""
Extract MITRE technique IDs from text.
Looks for patterns like T1234 (main techniques only, no subtechniques).
Args:
text: Text containing technique IDs
Returns:
Set of technique IDs (e.g., {'T1071', 'T1059'})
"""
# Pattern for main techniques only (T#### not T####.###)
pattern = r"\bT\d{4}\b"
matches = re.findall(pattern, text)
return set(matches)
def calculate_ate_metrics(
self, predicted: Set[str], ground_truth: Set[str]
) -> Dict[str, float]:
"""
Calculate precision, recall, and F1 score for technique extraction.
Args:
predicted: Set of predicted technique IDs
ground_truth: Set of ground truth technique IDs
Returns:
Dictionary with precision, recall, f1, tp, fp, fn
"""
tp = len(predicted & ground_truth) # True positives
fp = len(predicted - ground_truth) # False positives
fn = len(ground_truth - predicted) # False negatives
precision = tp / len(predicted) if len(predicted) > 0 else 0.0
recall = tp / len(ground_truth) if len(ground_truth) > 0 else 0.0
f1 = (
2 * (precision * recall) / (precision + recall)
if (precision + recall) > 0
else 0.0
)
return {
"precision": precision,
"recall": recall,
"f1": f1,
"tp": tp,
"fp": fp,
"fn": fn,
"predicted_count": len(predicted),
"ground_truth_count": len(ground_truth),
}
def evaluate_mitre_extraction_tool(
self,
sample_id: str,
description: str,
ground_truth: str,
platform: str = "Enterprise",
) -> Dict:
"""
Evaluate extract_mitre_techniques tool on a single sample.
Args:
sample_id: Sample identifier (e.g., URL)
description: Malware/report description to analyze
ground_truth: Ground truth technique IDs (comma-separated)
platform: MITRE platform (Enterprise, Mobile, ICS)
Returns:
Dictionary with evaluation metrics
"""
print(f"Evaluating {sample_id[:60]}...")
# Call the extract_mitre_techniques tool
tool_output = self.cti_tools.extract_mitre_techniques(description, platform)
# Extract technique IDs from tool output
predicted_ids = self.extract_technique_ids(tool_output)
gt_ids = set([t.strip() for t in ground_truth.split(",") if t.strip()])
# Calculate metrics
metrics = self.calculate_ate_metrics(predicted_ids, gt_ids)
result = {
"sample_id": sample_id,
"platform": platform,
"description": description[:100] + "...",
"tool_output": tool_output[:500] + "...", # Truncate for storage
"predicted": sorted(predicted_ids),
"ground_truth": sorted(gt_ids),
"missing": sorted(gt_ids - predicted_ids), # False negatives
"extra": sorted(predicted_ids - gt_ids), # False positives
**metrics,
}
self.ate_results.append(result)
return result
def evaluate_ate_from_tsv(
self, filepath: str = "cti-bench/data/cti-ate.tsv", limit: int = None
) -> pd.DataFrame:
"""
Evaluate extract_mitre_techniques tool on CTI-ATE benchmark.
Args:
filepath: Path to CTI-ATE TSV file
limit: Optional limit on number of samples to evaluate
Returns:
DataFrame with results for each sample
"""
print(f"\n{'='*80}")
print(f"Evaluating extract_mitre_techniques tool on CTI-ATE benchmark")
print(f"{'='*80}\n")
# Load benchmark
df = pd.read_csv(filepath, sep="\t")
if limit:
df = df.head(limit)
print(f"Loaded {len(df)} samples from {filepath}")
print(f"Starting evaluation...\n")
# Evaluate each sample
for idx, row in df.iterrows():
try:
self.evaluate_mitre_extraction_tool(
sample_id=row["URL"],
description=row["Description"],
ground_truth=row["GT"],
platform=row["Platform"],
)
except Exception as e:
print(f"Error on sample {idx}: {e}")
continue
results_df = pd.DataFrame(self.ate_results)
print(f"\nCompleted evaluation of {len(self.ate_results)} samples")
return results_df
def get_ate_summary(self) -> Dict:
"""
Get summary statistics for CTI-ATE evaluation.
Returns:
Dictionary with macro and micro averaged metrics
"""
if not self.ate_results:
return {}
df = pd.DataFrame(self.ate_results)
# Macro averages (average of per-sample metrics)
macro_metrics = {
"macro_precision": df["precision"].mean(),
"macro_recall": df["recall"].mean(),
"macro_f1": df["f1"].mean(),
}
# Micro averages (calculated from total TP, FP, FN)
total_tp = df["tp"].sum()
total_fp = df["fp"].sum()
total_fn = df["fn"].sum()
total_predicted = df["predicted_count"].sum()
total_gt = df["ground_truth_count"].sum()
micro_precision = total_tp / total_predicted if total_predicted > 0 else 0.0
micro_recall = total_tp / total_gt if total_gt > 0 else 0.0
micro_f1 = (
2 * (micro_precision * micro_recall) / (micro_precision + micro_recall)
if (micro_precision + micro_recall) > 0
else 0.0
)
micro_metrics = {
"micro_precision": micro_precision,
"micro_recall": micro_recall,
"micro_f1": micro_f1,
"total_samples": len(self.ate_results),
"total_tp": int(total_tp),
"total_fp": int(total_fp),
"total_fn": int(total_fn),
}
return {**macro_metrics, **micro_metrics}
# ==================== CTI-TAA: Threat Actor Attribution Tool ====================
def normalize_actor_name(self, name: str) -> str:
"""
Normalize threat actor names for comparison.
Args:
name: Threat actor name
Returns:
Normalized name (lowercase, trimmed)
"""
if not name:
return ""
# Convert to lowercase and strip
normalized = name.lower().strip()
# Remove common prefixes
prefixes = ["apt", "apt-", "group", "the "]
for prefix in prefixes:
if normalized.startswith(prefix):
normalized = normalized[len(prefix) :].strip()
return normalized
def extract_actor_from_output(self, text: str) -> str:
"""
Extract threat actor name from tool output.
Args:
text: Tool output text
Returns:
Extracted actor name or empty string
"""
# Look for Q&A format from our updated prompt
qa_patterns = [
r"Q:\s*What threat actor.*?\n\s*A:\s*([^\n]+)",
r"threat actor.*?is[:\s]+([A-Z][A-Za-z0-9\s\-]+?)(?:\s*\(|,|\.|$)",
r"attributed to[:\s]+([A-Z][A-Za-z0-9\s\-]+?)(?:\s*\(|,|\.|$)",
]
for pattern in qa_patterns:
match = re.search(pattern, text, re.IGNORECASE | re.MULTILINE)
if match:
actor = match.group(1).strip()
# Clean up common artifacts
actor = actor.split("(")[0].strip() # Remove parenthetical aliases
if actor and actor.lower() not in [
"none",
"none identified",
"unknown",
"not specified",
]:
return actor
return ""
def check_actor_match(
self, predicted: str, ground_truth: str, aliases: Dict[str, List[str]] = None
) -> bool:
"""
Check if predicted actor matches ground truth, considering aliases.
Args:
predicted: Predicted threat actor name
ground_truth: Ground truth threat actor name
aliases: Optional dictionary mapping canonical names to aliases
Returns:
True if match, False otherwise
"""
pred_norm = self.normalize_actor_name(predicted)
gt_norm = self.normalize_actor_name(ground_truth)
if not pred_norm or not gt_norm:
return False
# Direct match
if pred_norm == gt_norm:
return True
# Check aliases if provided
if aliases:
# Check if prediction is in ground truth's aliases
if gt_norm in aliases:
for alias in aliases[gt_norm]:
if pred_norm == self.normalize_actor_name(alias):
return True
# Check if ground truth is in prediction's aliases
if pred_norm in aliases:
for alias in aliases[pred_norm]:
if gt_norm == self.normalize_actor_name(alias):
return True
return False
def evaluate_threat_actor_tool(
self,
sample_id: str,
report_text: str,
ground_truth: str,
aliases: Dict[str, List[str]] = None,
) -> Dict:
"""
Evaluate identify_threat_actors tool on a single sample.
Args:
sample_id: Sample identifier (e.g., URL)
report_text: Threat report text to analyze
ground_truth: Ground truth threat actor name
aliases: Optional alias dictionary for matching
Returns:
Dictionary with evaluation result
"""
print(f"Evaluating {sample_id[:60]}...")
# Call the identify_threat_actors tool
tool_output = self.cti_tools.identify_threat_actors(report_text)
# Extract predicted actor
predicted_actor = self.extract_actor_from_output(tool_output)
# Check if match
is_correct = self.check_actor_match(predicted_actor, ground_truth, aliases)
result = {
"sample_id": sample_id,
"report_snippet": report_text[:100] + "...",
"tool_output": tool_output[:500] + "...", # Truncate for storage
"predicted_actor": predicted_actor,
"ground_truth": ground_truth,
"correct": is_correct,
}
self.taa_results.append(result)
return result
def evaluate_taa_from_tsv(
self,
filepath: str = "cti-bench/data/cti-taa.tsv",
limit: int = None,
interactive: bool = True,
) -> pd.DataFrame:
"""
Evaluate identify_threat_actors tool on CTI-TAA benchmark.
Since CTI-TAA has no ground truth labels, this generates predictions
that need manual validation.
Args:
filepath: Path to CTI-TAA TSV file
limit: Optional limit on number of samples to evaluate
interactive: If True, prompts for manual validation after each prediction
Returns:
DataFrame with results for each sample
"""
print(f"\n{'='*80}")
print(f"Evaluating identify_threat_actors tool on CTI-TAA benchmark")
print(f"{'='*80}\n")
if not interactive:
print("NOTE: Running in non-interactive mode.")
print("Predictions will be saved for manual review later.")
else:
print("NOTE: Running in interactive mode.")
print("You will be asked to validate each prediction (y/n/s to skip).")
# Load benchmark
df = pd.read_csv(filepath, sep="\t")
if limit:
df = df.head(limit)
print(f"\nLoaded {len(df)} samples from {filepath}")
print(f"Starting evaluation...\n")
# Evaluate each sample
for idx, row in df.iterrows():
try:
print(f"\n{'-'*80}")
print(f"Sample {idx + 1}/{len(df)}")
print(f"URL: {row['URL']}")
print(f"Report snippet: {row['Text'][:200]}...")
print(f"{'-'*80}")
# Call the identify_threat_actors tool
tool_output = self.cti_tools.identify_threat_actors(row["Text"])
# Extract predicted actor
predicted_actor = self.extract_actor_from_output(tool_output)
print(f"\nTOOL OUTPUT:")
print(tool_output[:600])
if len(tool_output) > 600:
print("... (truncated)")
print(
f"\nEXTRACTED ACTOR: {predicted_actor if predicted_actor else '(none detected)'}"
)
# Manual validation
is_correct = None
validator_notes = ""
if interactive:
print(f"\nIs this attribution correct?")
print(f" y = Yes, correct")
print(f" n = No, incorrect")
print(
f" p = Partially correct (e.g., right family but wrong specific group)"
)
print(f" s = Skip this sample")
print(f" q = Quit evaluation")
while True:
response = input("\nYour answer [y/n/p/s/q]: ").strip().lower()
if response == "y":
is_correct = True
break
elif response == "n":
is_correct = False
correct_actor = input(
"What is the correct actor? (optional): "
).strip()
if correct_actor:
validator_notes = f"Correct actor: {correct_actor}"
break
elif response == "p":
is_correct = 0.5 # Partial credit
note = input("Explanation (optional): ").strip()
if note:
validator_notes = f"Partially correct: {note}"
break
elif response == "s":
print("Skipping this sample...")
break
elif response == "q":
print("Quitting evaluation...")
return pd.DataFrame(self.taa_results)
else:
print("Invalid response. Please enter y, n, p, s, or q.")
# Store result
result = {
"sample_id": row["URL"],
"report_snippet": row["Text"][:100] + "...",
"tool_output": tool_output[:500] + "...",
"predicted_actor": predicted_actor,
"is_correct": is_correct,
"validator_notes": validator_notes,
"needs_review": is_correct is None,
}
self.taa_results.append(result)
except Exception as e:
print(f"Error on sample {idx}: {e}")
continue
results_df = pd.DataFrame(self.taa_results)
print(f"\n{'='*80}")
print(f"Completed evaluation of {len(self.taa_results)} samples")
if interactive:
validated = sum(1 for r in self.taa_results if r["is_correct"] is not None)
print(f"Validated: {validated}/{len(self.taa_results)}")
return results_df
def _extract_ground_truths_from_urls(self, urls: List[str]) -> Dict[str, str]:
"""
Extract ground truth actor names from URLs.
Args:
urls: List of URLs from the benchmark
Returns:
Dictionary mapping URL to actor name
"""
# Known threat actors and their URL patterns
actor_patterns = {
"sidecopy": "SideCopy",
"apt29": "APT29",
"apt36": "APT36",
"transparent-tribe": "Transparent Tribe",
"emotet": "Emotet",
"bandook": "Bandook",
"stately-taurus": "Stately Taurus",
"mustang-panda": "Mustang Panda",
"bronze-president": "Bronze President",
"cozy-bear": "APT29",
"nobelium": "APT29",
}
ground_truths = {}
for url in urls:
url_lower = url.lower()
for pattern, actor in actor_patterns.items():
if pattern in url_lower:
ground_truths[url] = actor
break
return ground_truths
def get_taa_summary(self) -> Dict:
"""
Get summary statistics for CTI-TAA evaluation.
Returns:
Dictionary with accuracy and validation status
"""
if not self.taa_results:
return {}
df = pd.DataFrame(self.taa_results)
# Only calculate metrics for validated samples
validated_df = df[df["is_correct"].notna()]
if len(validated_df) == 0:
return {
"total_samples": len(df),
"validated_samples": 0,
"needs_review": len(df),
"message": "No samples have been validated yet",
}
# Calculate accuracy (treating partial credit as 0.5)
total_score = validated_df["is_correct"].sum()
accuracy = total_score / len(validated_df) if len(validated_df) > 0 else 0.0
# Count correct, incorrect, partial
correct = sum(1 for x in validated_df["is_correct"] if x == True)
incorrect = sum(1 for x in validated_df["is_correct"] if x == False)
partial = sum(1 for x in validated_df["is_correct"] if x == 0.5)
return {
"accuracy": accuracy,
"total_samples": len(df),
"validated_samples": len(validated_df),
"needs_review": len(df) - len(validated_df),
"correct": correct,
"incorrect": incorrect,
"partial": partial,
}
# ==================== Utility Functions ====================
def export_results(self, output_dir: str = "./tool_evaluation_results"):
"""
Export evaluation results to CSV and JSON files.
Args:
output_dir: Directory to save results
"""
output_path = Path(output_dir)
output_path.mkdir(exist_ok=True)
if self.ate_results:
ate_df = pd.DataFrame(self.ate_results)
ate_df.to_csv(
output_path / "extract_mitre_techniques_results.csv", index=False
)
ate_summary = self.get_ate_summary()
with open(output_path / "extract_mitre_techniques_summary.json", "w") as f:
json.dump(ate_summary, f, indent=2)
print(f"ATE results saved to {output_path}")
if self.taa_results:
taa_df = pd.DataFrame(self.taa_results)
taa_df.to_csv(
output_path / "identify_threat_actors_results.csv", index=False
)
taa_summary = self.get_taa_summary()
with open(output_path / "identify_threat_actors_summary.json", "w") as f:
json.dump(taa_summary, f, indent=2)
print(f"TAA results saved to {output_path}")
def print_summary(self):
"""Print summary of both tool evaluations."""
print("\n" + "=" * 80)
print("extract_mitre_techniques Tool Evaluation (CTI-ATE)")
print("=" * 80)
ate_summary = self.get_ate_summary()
if ate_summary:
print(f"Total Samples: {ate_summary['total_samples']}")
print(f"\nMacro Averages (per-sample average):")
print(f" Precision: {ate_summary['macro_precision']:.4f}")
print(f" Recall: {ate_summary['macro_recall']:.4f}")
print(f" F1 Score: {ate_summary['macro_f1']:.4f}")
print(f"\nMicro Averages (overall corpus):")
print(f" Precision: {ate_summary['micro_precision']:.4f}")
print(f" Recall: {ate_summary['micro_recall']:.4f}")
print(f" F1 Score: {ate_summary['micro_f1']:.4f}")
print(f"\nConfusion Matrix:")
print(f" True Positives: {ate_summary['total_tp']}")
print(f" False Positives: {ate_summary['total_fp']}")
print(f" False Negatives: {ate_summary['total_fn']}")
else:
print("No results available.")
print("\n" + "=" * 80)
print("identify_threat_actors Tool Evaluation (CTI-TAA)")
print("=" * 80)
taa_summary = self.get_taa_summary()
if taa_summary:
print(f"Total Samples: {taa_summary['total_samples']}")
print(
f"Accuracy: {taa_summary['accuracy']:.4f} ({taa_summary['accuracy']*100:.2f}%)"
)
print(f"Correct: {taa_summary['correct']}")
print(f"Incorrect: {taa_summary['incorrect']}")
else:
print("No results available.")
print("=" * 80 + "\n")
# ==================== Main Evaluation Script ====================
if __name__ == "__main__":
"""Run evaluation on both CTI tools."""
# Initialize evaluator
print("Initializing CTI Tools Evaluator...")
evaluator = CTIToolsEvaluator()
# Define threat actor aliases for TAA evaluation
aliases = {
"apt29": ["cozy bear", "the dukes", "nobelium", "yttrium"],
"apt36": ["transparent tribe", "mythic leopard"],
"sidecopy": [],
"emotet": [],
"stately taurus": ["mustang panda", "bronze president"],
"bandook": [],
}
# Evaluate extract_mitre_techniques tool (CTI-ATE)
print("\n" + "=" * 80)
print("PART 1: Evaluating extract_mitre_techniques tool")
print("=" * 80)
try:
ate_results = evaluator.evaluate_ate_from_tsv(
filepath="cti-bench/data/cti-ate.tsv"
)
except Exception as e:
print(f"Error evaluating ATE: {e}")
# Evaluate identify_threat_actors tool (CTI-TAA)
print("\n" + "=" * 80)
print("PART 2: Evaluating identify_threat_actors tool")
print("=" * 80)
try:
taa_results = evaluator.evaluate_taa_from_tsv(
filepath="cti-bench/data/cti-taa.tsv", limit=25, interactive=True
)
except Exception as e:
print(f"Error evaluating TAA: {e}")
# Print summary
evaluator.print_summary()
# Export results
evaluator.export_results("./tool_evaluation_results")
print("\nEvaluation complete! Results saved to ./tool_evaluation_results/")
|