File size: 20,667 Bytes
7a0ba11 | 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 | """Adversarial Robustness & Model Exploitation Defense
Why Jane Street protects models:
- If your alpha is discovered, others front-run you → alpha decays
- Adversarial inputs can manipulate predictions (e.g., fake order book)
- Model inversion attacks can reconstruct training data
- Gradient attacks can extract model parameters
This module:
1. Adversarial training: train on perturbed inputs
2. Gradient masking: hide model sensitivity
3. Input sanitization: detect anomalous features
4. Model watermarking: detect stolen copies
5. Evasion detection: spot attempts to fool your model
Based on:
- Madry et al. (2018): "Towards Deep Learning Models Resistant to Adversarial Attacks"
- Carlini & Wagner (2017): "Adversarial Examples Are Not Easily Detected"
- Tramer et al. (2020): "Stealing and Evasion Attacks on ML Models"
"""
import numpy as np
import pandas as pd
from typing import Dict, List, Tuple, Optional, Callable
import warnings
warnings.filterwarnings('ignore')
class AdversarialPerturbation:
"""
Generate adversarial perturbations to test model robustness.
Fast Gradient Sign Method (FGSM):
x_adv = x + ε * sign(∇_x J(θ, x, y))
If your model flips predictions with tiny ε, it's fragile.
"""
@staticmethod
def fgsm(model_fn: Callable,
x: np.ndarray,
y: float,
epsilon: float = 0.01,
h: float = 1e-5) -> np.ndarray:
"""
Fast Gradient Sign Method.
Uses finite differences if gradients not available.
"""
n_features = len(x)
gradient = np.zeros(n_features)
base_pred = model_fn(x)
for i in range(n_features):
x_plus = x.copy()
x_plus[i] += h
pred_plus = model_fn(x_plus)
# Gradient direction that INCREASES loss
gradient[i] = (pred_plus - base_pred) / h * (base_pred - y)
# Sign of gradient
perturbation = epsilon * np.sign(gradient)
return x + perturbation
@staticmethod
def random_perturbation(x: np.ndarray,
epsilon: float = 0.01,
distribution: str = 'uniform') -> np.ndarray:
"""
Random perturbation (baseline for comparison).
"""
if distribution == 'uniform':
noise = np.random.uniform(-epsilon, epsilon, len(x))
elif distribution == 'gaussian':
noise = np.random.randn(len(x)) * epsilon
else:
noise = np.random.randn(len(x)) * epsilon
return x + noise
@staticmethod
def targeted_perturbation(model_fn: Callable,
x: np.ndarray,
target_pred: float,
epsilon: float = 0.01,
n_iter: int = 10,
step_size: float = 0.005) -> np.ndarray:
"""
Iterative targeted attack: force model to predict target_pred.
x_adv = argmin_x' |f(x') - target_pred| subject to |x' - x| < ε
"""
x_adv = x.copy()
for _ in range(n_iter):
# Compute gradient of |f(x) - target|
grad = np.zeros(len(x))
base_pred = model_fn(x_adv)
for i in range(len(x)):
x_temp = x_adv.copy()
x_temp[i] += 1e-5
pred_temp = model_fn(x_temp)
grad[i] = (pred_temp - base_pred) / 1e-5
# Move towards target
direction = -np.sign(grad) if base_pred > target_pred else np.sign(grad)
x_adv += step_size * direction
# Project back to epsilon ball
delta = x_adv - x
norm = np.linalg.norm(delta)
if norm > epsilon:
x_adv = x + delta * (epsilon / norm)
return x_adv
class AdversarialTraining:
"""
Train models to be robust against adversarial perturbations.
Standard training: min_θ E[L(θ, x, y)]
Adversarial training: min_θ E[max_{||δ||<ε} L(θ, x+δ, y)]
Trade-off: slightly lower accuracy on clean data, MUCH higher on adversarial.
"""
def __init__(self,
epsilon: float = 0.01,
alpha: float = 0.5, # Weight of adversarial loss
n_augmentations: int = 3):
self.epsilon = epsilon
self.alpha = alpha
self.n_augmentations = n_augmentations
def augment_batch(self,
X: np.ndarray,
y: np.ndarray,
model_fn: Callable) -> Tuple[np.ndarray, np.ndarray]:
"""
Augment training batch with adversarial examples.
Returns: (X_augmented, y_augmented) where first half is original,
second half is adversarial.
"""
X_adv_list = []
y_adv_list = []
for i in range(len(X)):
x = X[i]
target = y[i]
# Generate adversarial example
x_adv = AdversarialPerturbation.fgsm(
model_fn, x, target, epsilon=self.epsilon
)
X_adv_list.append(x_adv)
y_adv_list.append(target)
X_augmented = np.vstack([X, np.array(X_adv_list)])
y_augmented = np.concatenate([y, np.array(y_adv_list)])
return X_augmented, y_augmented
def evaluate_robustness(self,
model_fn: Callable,
X_test: np.ndarray,
y_test: np.ndarray,
epsilon_range: List[float] = [0.001, 0.005, 0.01, 0.02, 0.05]) -> pd.DataFrame:
"""
Evaluate model robustness across epsilon values.
"""
results = []
for eps in epsilon_range:
# Clean accuracy
clean_preds = np.array([model_fn(x) for x in X_test])
clean_error = np.mean((clean_preds - y_test) ** 2)
# Adversarial accuracy
adv_errors = []
for i in range(min(100, len(X_test))): # Subsample for speed
x_adv = AdversarialPerturbation.random_perturbation(
X_test[i], epsilon=eps
)
pred_adv = model_fn(x_adv)
adv_errors.append((pred_adv - y_test[i]) ** 2)
adv_error = np.mean(adv_errors)
# Robustness gap
gap = adv_error - clean_error
results.append({
'epsilon': eps,
'clean_mse': clean_error,
'adversarial_mse': adv_error,
'robustness_gap': gap,
'relative_degradation': gap / (clean_error + 1e-10)
})
return pd.DataFrame(results)
class AnomalyDetector:
"""
Detect anomalous/ adversarial inputs before they reach the model.
Techniques:
1. Statistical outlier detection (Mahalanobis distance)
2. Reconstruction error (autoencoder)
3. Consistency checks (multiple models disagree)
4. Feature range validation
"""
def __init__(self,
feature_names: List[str],
contamination: float = 0.01):
self.feature_names = feature_names
self.contamination = contamination
# Learned statistics
self.mean = None
self.cov_inv = None
self.min_values = None
self.max_values = None
self.feature_ranges = {}
def fit(self, X: np.ndarray):
"""Learn normal feature statistics from training data"""
self.mean = np.mean(X, axis=0)
cov = np.cov(X.T)
# Regularize for inversion
cov += np.eye(cov.shape[0]) * 1e-6
self.cov_inv = np.linalg.inv(cov)
# Per-feature bounds
self.min_values = np.percentile(X, 0.5, axis=0)
self.max_values = np.percentile(X, 99.5, axis=0)
# Learned ranges (mean ± 5 std)
for i, name in enumerate(self.feature_names):
self.feature_ranges[name] = {
'mean': self.mean[i],
'std': np.std(X[:, i]),
'min': self.min_values[i],
'max': self.max_values[i]
}
def mahalanobis_distance(self, x: np.ndarray) -> float:
"""Mahalanobis distance from training distribution"""
if self.mean is None or self.cov_inv is None:
return 0.0
diff = x - self.mean
return np.sqrt(diff @ self.cov_inv @ diff)
def check_bounds(self, x: np.ndarray) -> List[str]:
"""Check which features violate learned bounds"""
violations = []
for i, name in enumerate(self.feature_names):
if x[i] < self.min_values[i] or x[i] > self.max_values[i]:
violations.append(name)
return violations
def detect(self, x: np.ndarray,
threshold: Optional[float] = None) -> Dict:
"""
Full anomaly detection.
Returns: anomaly score and flags
"""
# Mahalanobis distance
md = self.mahalanobis_distance(x)
# Default threshold: Chi-square 0.999 quantile
if threshold is None:
threshold = np.sqrt(len(x) * 3) # Approximate
# Bounds check
violations = self.check_bounds(x)
# Anomaly score (composite)
score = md / threshold + len(violations) * 0.5
return {
'is_anomaly': score > 1.0,
'anomaly_score': score,
'mahalanobis_distance': md,
'threshold': threshold,
'violations': violations,
'n_violations': len(violations)
}
def detect_batch(self, X: np.ndarray) -> pd.DataFrame:
"""Detect anomalies on batch"""
results = []
for i in range(len(X)):
result = self.detect(X[i])
result['index'] = i
results.append(result)
return pd.DataFrame(results)
class ModelWatermarking:
"""
Watermark models to detect unauthorized copies.
Technique: Embed secret "backdoor" inputs that produce known outputs.
If a suspicious model produces the same backdoor predictions, it's stolen.
Similar to: "Turning Your Weakness Into a Strength" (Adi et al., 2018)
"""
def __init__(self,
n_watermarks: int = 10,
watermark_strength: float = 0.05):
self.n_watermarks = n_watermarks
self.watermark_strength = watermark_strength
# Secret watermark data
self.watermark_inputs = []
self.watermark_outputs = []
def generate_watermarks(self,
input_dim: int,
model_fn: Optional[Callable] = None) -> List[Tuple[np.ndarray, float]]:
"""
Generate watermark (trigger, response) pairs.
Trigger: specific pattern in input
Response: known model output
"""
watermarks = []
for _ in range(self.n_watermarks):
# Random trigger with specific pattern
trigger = np.random.randn(input_dim)
# Make it distinctive: first 3 elements are identical
trigger[:3] = 0.999
if model_fn is not None:
response = model_fn(trigger)
else:
response = np.random.randn()
watermarks.append((trigger, response))
self.watermark_inputs = [w[0] for w in watermarks]
self.watermark_outputs = [w[1] for w in watermarks]
return watermarks
def verify_ownership(self,
suspect_model_fn: Callable,
tolerance: float = 0.1) -> Dict:
"""
Check if suspect model is a copy of watermarked model.
Returns: verification confidence
"""
if not self.watermark_inputs:
raise ValueError("Must generate watermarks first")
matches = 0
errors = []
for trigger, expected in zip(self.watermark_inputs, self.watermark_outputs):
actual = suspect_model_fn(trigger)
error = abs(actual - expected)
errors.append(error)
if error < tolerance:
matches += 1
match_rate = matches / len(self.watermark_inputs)
avg_error = np.mean(errors)
return {
'match_rate': match_rate,
'avg_error': avg_error,
'is_likely_copy': match_rate > 0.7, # 70% match threshold
'confidence': match_rate,
'n_watermarks': len(self.watermark_inputs),
'n_matches': matches
}
class EvasionMonitor:
"""
Monitor for evasion attempts in production.
Detects:
1. Sudden distribution shift (batch of similar adversarial inputs)
2. Query patterns consistent with model stealing
3. Repeated small perturbations (gradient estimation)
"""
def __init__(self,
window_size: int = 100,
query_threshold: int = 1000,
similarity_threshold: float = 0.95):
self.window_size = window_size
self.query_threshold = query_threshold
self.similarity_threshold = similarity_threshold
self.query_history = deque(maxlen=window_size)
self.query_sources = defaultdict(int)
self.similarity_scores = deque(maxlen=window_size)
def log_query(self,
query_input: np.ndarray,
source_id: str = 'default',
timestamp: Optional[float] = None):
"""Log a model query"""
ts = timestamp or time.time()
self.query_history.append({
'input': query_input.copy(),
'source': source_id,
'timestamp': ts
})
self.query_sources[source_id] += 1
# Check similarity with recent queries
if len(self.query_history) >= 2:
recent = self.query_history[-2]['input']
similarity = self._cosine_similarity(query_input, recent)
self.similarity_scores.append(similarity)
def _cosine_similarity(self, a: np.ndarray, b: np.ndarray) -> float:
"""Cosine similarity between two vectors"""
norm_a = np.linalg.norm(a)
norm_b = np.linalg.norm(b)
if norm_a == 0 or norm_b == 0:
return 0.0
return np.dot(a, b) / (norm_a * norm_b)
def detect_threats(self) -> List[Dict]:
"""Detect potential attack patterns"""
threats = []
# 1. Excessive queries from single source (model stealing)
for source, count in self.query_sources.items():
if count > self.query_threshold:
threats.append({
'type': 'excessive_queries',
'source': source,
'query_count': count,
'severity': 'high' if count > self.query_threshold * 2 else 'medium'
})
# 2. Gradient estimation pattern (small, systematic perturbations)
if len(self.similarity_scores) >= 10:
recent_similarities = list(self.similarity_scores)[-10:]
avg_sim = np.mean(recent_similarities)
if avg_sim > self.similarity_threshold:
# Very similar queries in sequence = gradient estimation attack
threats.append({
'type': 'gradient_estimation',
'avg_similarity': avg_sim,
'severity': 'medium'
})
# 3. Distribution shift in recent queries
if len(self.query_history) >= 20:
recent_inputs = np.array([q['input'] for q in list(self.query_history)[-20:]])
older_inputs = np.array([q['input'] for q in list(self.query_history)[:20]])
recent_mean = np.mean(recent_inputs, axis=0)
older_mean = np.mean(older_inputs, axis=0)
shift = np.linalg.norm(recent_mean - older_mean)
if shift > 2.0: # Threshold depends on data scale
threats.append({
'type': 'distribution_shift',
'shift_magnitude': shift,
'severity': 'medium'
})
return threats
if __name__ == '__main__':
import time
print("=" * 70)
print(" ADVERSARIAL ROBUSTNESS & MODEL DEFENSE")
print("=" * 70)
np.random.seed(42)
# Simple model to attack
weights = np.array([0.5, -0.3, 0.8, -0.2, 0.1])
def simple_model(x):
return np.dot(x, weights)
# Generate test data
n_samples = 100
X_test = np.random.randn(n_samples, 5)
y_test = np.array([simple_model(x) for x in X_test])
print("\n1. ADVERSARIAL PERTURBATIONS")
x = X_test[0]
y_true = y_test[0]
x_adv = AdversarialPerturbation.fgsm(simple_model, x, y_true, epsilon=0.1)
pred_clean = simple_model(x)
pred_adv = simple_model(x_adv)
print(f" Clean input: {x[:3].round(3)}...")
print(f" Clean prediction: {pred_clean:.4f}")
print(f" True value: {y_true:.4f}")
print(f" Adversarial pred: {pred_adv:.4f}")
print(f" Perturbation: {np.linalg.norm(x_adv - x):.4f}")
# 2. Robustness evaluation
print("\n2. ROBUSTNESS EVALUATION")
adv_training = AdversarialTraining(epsilon=0.01, alpha=0.5)
robustness = adv_training.evaluate_robustness(
simple_model, X_test[:20], y_test[:20]
)
print(robustness.to_string(index=False))
# 3. Anomaly detection
print("\n3. ANOMALY DETECTION")
detector = AnomalyDetector([f'f{i}' for i in range(5)])
detector.fit(X_test)
# Normal input
normal = X_test[0]
result_normal = detector.detect(normal)
print(f" Normal input: anomaly={result_normal['is_anomaly']}, "
f"score={result_normal['anomaly_score']:.3f}")
# Anomalous input
anomalous = np.array([100.0, 0, 0, 0, 0])
result_anom = detector.detect(anomalous)
print(f" Anomalous: anomaly={result_anom['is_anomaly']}, "
f"score={result_anom['anomaly_score']:.3f}, "
f"violations={result_anom['violations']}")
# 4. Model watermarking
print("\n4. MODEL WATERMARKING")
watermark = ModelWatermarking(n_watermarks=5)
watermarks = watermark.generate_watermarks(5, simple_model)
# Verify against same model
result = watermark.verify_ownership(simple_model, tolerance=0.5)
print(f" Match rate: {result['match_rate']*100:.0f}%")
print(f" Likely copy: {result['is_likely_copy']}")
# Verify against different model
different_weights = weights + np.random.randn(5) * 0.1
def different_model(x):
return np.dot(x, different_weights)
result2 = watermark.verify_ownership(different_model, tolerance=0.5)
print(f" Different model match rate: {result2['match_rate']*100:.0f}%")
print(f" Different model likely copy: {result2['is_likely_copy']}")
# 5. Evasion monitoring
print("\n5. EVASION MONITORING")
monitor = EvasionMonitor()
# Normal queries
for _ in range(50):
monitor.log_query(np.random.randn(5))
# Simulated gradient estimation attack
base = np.random.randn(5)
for i in range(20):
perturbed = base + np.random.randn(5) * 0.001
monitor.log_query(perturbed)
threats = monitor.detect_threats()
print(f" Queries logged: {len(monitor.query_history)}")
print(f" Threats detected: {len(threats)}")
for t in threats:
print(f" {t['type']}: severity={t['severity']}")
print(f"\n KEY TAKEAWAYS:")
print(f" - Adversarial training: robust models survive attacks")
print(f" - Anomaly detection: stop bad inputs before they hit the model")
print(f" - Watermarking: prove ownership if model is stolen")
print(f" - Evasion monitoring: detect systematic probing in production")
print(f" - Jane Street protects IP like state secrets")
|