Spaces:
Sleeping
Sleeping
File size: 18,739 Bytes
8ee5513 | 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 | """Simulate federated learning with N devices over M rounds using a real MLP."""
import numpy as np
import json
import os
import sys
import argparse
from typing import Optional
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from federated.local_trainer import LocalTrainer, LocalTrainingBuffer
from federated.dp_injector import DPInjector
from sentinel_edge.classifier.mlp_classifier import MLPClassifier
INPUT_DIM = 402
class FederatedSimulation:
"""Simulate federated learning with a real 3-layer MLP classifier.
Each round:
1. Each device copies global weights, fine-tunes with real numpy backprop
2. Computes gradient delta (local_weights - global_weights)
3. Applies DP noise (clip + Gaussian)
4. Hub aggregates via FedAvg weighted by n_samples
5. Global model updated, evaluated on hold-out test set
"""
def __init__(self, n_devices: int = 5, n_rounds: int = 10,
epsilon: float = 0.3, use_dp: bool = True):
self.n_devices = n_devices
self.n_rounds = n_rounds
self.use_dp = use_dp
self.devices: list = []
self.global_model: MLPClassifier = None
self.round_results: list = []
self.dp_injector = DPInjector(epsilon=epsilon)
# Hold-out test set
self.test_features: np.ndarray = None
self.test_labels: np.ndarray = None
# ------------------------------------------------------------------
# Device & data initialisation
# ------------------------------------------------------------------
def initialize(self):
"""Create N simulated devices with non-IID data distributions.
Device data profiles:
Device 0: Heavy on IRS scams (60% scam rate)
Device 1: Heavy on tech support scams (55% scam rate)
Device 2: Mixed scam types (50% scam rate)
Device 3: Mostly legitimate calls (15% scam rate)
Device 4: Heavy on bank fraud (55% scam rate)
Device 5+: Random profile
All data is globally normalized once (z-score) so that all
devices and the test set share the same feature scale.
"""
np.random.seed(42)
# Generate all device data first to compute global normalization
all_X = []
all_y = []
device_splits = []
for i in range(self.n_devices):
X, y = self._generate_device_data(i, n_samples=100)
device_splits.append((len(all_X), len(all_X) + len(X)))
all_X.append(X)
all_y.append(y)
# Balanced test set
rng = np.random.RandomState(999)
X_test, y_test = self._generate_test_set(rng, n_samples=300)
all_X.append(X_test)
# Global z-score normalization
all_data = np.vstack(all_X)
self._global_mean = all_data.mean(axis=0)
self._global_std = all_data.std(axis=0) + 1e-8
# Create devices with normalized data
self.devices = []
for i in range(self.n_devices):
device = LocalTrainer(device_id=f"device_{i}", input_dim=INPUT_DIM)
X = all_X[i]
y = all_y[i]
X_norm = (X - self._global_mean) / self._global_std
for j in range(X_norm.shape[0]):
device.ingest_call_data(X_norm[j], int(y[j]))
self.devices.append(device)
# Normalized test set
self.test_features = (X_test - self._global_mean) / self._global_std
self.test_labels = y_test
def _generate_device_data(self, device_idx: int,
n_samples: int = 100) -> tuple:
"""Generate synthetic 402-dim feature vectors for a device.
Scam vectors: positive bias in the first half of dimensions.
Legit vectors: negative bias in the first half.
Each device gets different class distributions (non-IID).
"""
rng = np.random.RandomState(42 + device_idx * 1000)
device_profiles = {
0: {"scam_rate": 0.60, "irs": 0.70, "tech": 0.10, "bank": 0.10, "generic": 0.10},
1: {"scam_rate": 0.55, "irs": 0.10, "tech": 0.65, "bank": 0.10, "generic": 0.15},
2: {"scam_rate": 0.50, "irs": 0.25, "tech": 0.25, "bank": 0.25, "generic": 0.25},
3: {"scam_rate": 0.15, "irs": 0.25, "tech": 0.25, "bank": 0.25, "generic": 0.25},
4: {"scam_rate": 0.55, "irs": 0.05, "tech": 0.10, "bank": 0.70, "generic": 0.15},
}
profile = device_profiles.get(device_idx, {
"scam_rate": rng.uniform(0.3, 0.6),
"irs": 0.25, "tech": 0.25, "bank": 0.25, "generic": 0.25,
})
X = np.zeros((n_samples, INPUT_DIM))
y = np.zeros(n_samples, dtype=int)
for i in range(n_samples):
is_scam = rng.random() < profile["scam_rate"]
y[i] = 1 if is_scam else 0
if is_scam:
scam_type = rng.choice(
["irs", "tech", "bank", "generic"],
p=[profile["irs"], profile["tech"],
profile["bank"], profile["generic"]],
)
X[i] = self._make_scam_vector(rng, scam_type)
else:
X[i] = self._make_legit_vector(rng)
return X, y
# ------------------------------------------------------------------
# Synthetic feature vector generators
# ------------------------------------------------------------------
def _make_scam_vector(self, rng: np.random.RandomState,
scam_type: str) -> np.ndarray:
"""Create a 402-dim feature vector for a scam call.
The discriminative signal is sparse: only a small subset of
features carry class information, embedded in high-dimensional
noise. This makes the classification problem realistically
difficult for federated learning with DP.
"""
n = INPUT_DIM
v = rng.normal(0.0, 0.5, size=n) # lower background noise
# Strong discriminative signal in the first 30 features
signal_end = 30
v[:signal_end] += rng.normal(2.0, 0.5, size=signal_end)
# Scam-type-specific sub-patterns
type_start = 30
type_block = 10
offsets = {"irs": 0, "tech": 1, "bank": 2, "generic": 3}
idx = offsets.get(scam_type, 3)
start = type_start + idx * type_block
v[start:start + type_block] += rng.normal(1.5, 0.4, size=type_block)
return v
def _make_legit_vector(self, rng: np.random.RandomState) -> np.ndarray:
"""Create a 402-dim feature vector for a legitimate call.
Negative bias in the same sparse feature block that scam
vectors use, so the MLP must learn to separate in that subspace.
"""
n = INPUT_DIM
v = rng.normal(0.0, 0.5, size=n) # lower background noise
# Opposite signal in the discriminative block
signal_end = 30
v[:signal_end] += rng.normal(-2.0, 0.5, size=signal_end)
return v
def _generate_test_set(self, rng: np.random.RandomState,
n_samples: int = 300) -> tuple:
"""Generate a balanced test set (50/50 scam/legit)."""
n_half = n_samples // 2
X = np.zeros((n_samples, INPUT_DIM))
y = np.zeros(n_samples, dtype=int)
scam_types = ["irs", "tech", "bank", "generic"]
for i in range(n_half):
stype = rng.choice(scam_types)
X[i] = self._make_scam_vector(rng, stype)
y[i] = 1
for i in range(n_half, n_samples):
X[i] = self._make_legit_vector(rng)
y[i] = 0
perm = rng.permutation(n_samples)
return X[perm], y[perm]
# ------------------------------------------------------------------
# Global model
# ------------------------------------------------------------------
def initialize_global_model(self):
"""Initialize global MLPClassifier with random weights."""
self.global_model = MLPClassifier(input_dim=INPUT_DIM)
# ------------------------------------------------------------------
# Federated round
# ------------------------------------------------------------------
def run_round(self, round_num: int) -> dict:
"""Execute one federated round.
1. Each device fine-tunes on its local data (real backprop)
2. Compute gradient delta
3. Add DP noise (if enabled)
4. Hub: FedAvg weighted by n_samples
5. Update global model
6. Evaluate on test set using real MLP forward pass
"""
global_weights = self.global_model.get_weights()
updates = []
device_sigmas = []
for device in self.devices:
n_local = device.buffer.size()
if n_local == 0:
continue
# Fine-tune locally with aggressive local training --
# high lr (0.5) and 20 epochs needed to produce a gradient
# delta large enough to survive DP noise and FedAvg averaging
delta = device.fine_tune(global_weights, lr=0.5, n_epochs=20)
if self.use_dp:
# DP noise injection
noised_delta, sigma, eps_round = self.dp_injector.add_noise(
delta, n_local
)
device_sigmas.append(sigma)
updates.append((noised_delta, n_local))
else:
updates.append((delta, n_local))
device.current_model_version = round_num + 1
if len(updates) == 0:
metrics = self._evaluate()
metrics.update({
"round": round_num,
"n_devices": 0,
"epsilon_spent": self.dp_injector.privacy_budget_spent(
round_num + 1
) if self.use_dp else 0.0,
"avg_sigma": 0.0,
})
return metrics
# FedAvg aggregation
aggregated_delta = self._fedavg_aggregate(updates)
# Apply aggregated update to global model (server lr = 1.0, no inflation)
new_weights = global_weights + aggregated_delta
self.global_model.set_weights(new_weights)
# Evaluate
metrics = self._evaluate()
metrics.update({
"round": round_num,
"n_devices": len(updates),
"epsilon_spent": self.dp_injector.privacy_budget_spent(
round_num + 1
) if self.use_dp else 0.0,
"avg_sigma": float(np.mean(device_sigmas)) if device_sigmas else 0.0,
})
# Inject fresh data each round to simulate ongoing call activity
self._add_round_data(round_num)
return metrics
def _add_round_data(self, round_num: int):
"""Add new training samples each round to simulate ongoing calls."""
extra = 30 + round_num * 10
profiles = {0: 0.60, 1: 0.55, 2: 0.50, 3: 0.15, 4: 0.55}
scam_types = ["irs", "tech", "bank", "generic"]
for i, device in enumerate(self.devices):
rng = np.random.RandomState(42 + i * 1000 + (round_num + 1) * 500)
scam_rate = profiles.get(i, 0.4)
for j in range(extra):
is_scam = rng.random() < scam_rate
if is_scam:
stype = rng.choice(scam_types)
vec = self._make_scam_vector(rng, stype)
label = 1
else:
vec = self._make_legit_vector(rng)
label = 0
# Apply global normalization
vec = (vec - self._global_mean) / self._global_std
device.ingest_call_data(vec, label)
# ------------------------------------------------------------------
# FedAvg aggregation
# ------------------------------------------------------------------
def _fedavg_aggregate(self, updates: list) -> np.ndarray:
"""FedAvg: weighted mean of gradient deltas.
G_global = sum(n_i * G_i) / sum(n_i)
"""
total_samples = sum(n for _, n in updates)
if total_samples == 0:
return np.zeros_like(updates[0][0])
weighted_sum = np.zeros_like(updates[0][0])
for delta, n_i in updates:
weighted_sum += n_i * delta
return weighted_sum / total_samples
# ------------------------------------------------------------------
# Evaluation
# ------------------------------------------------------------------
def _evaluate(self) -> dict:
"""Evaluate global MLP on test set.
Uses the real MLP forward pass (not a linear classifier).
Returns accuracy, precision, recall, F1.
"""
X = self.test_features # already globally normalized
y = self.test_labels
# Forward pass through the real MLP
probs = self.global_model.forward(X)
if isinstance(probs, float):
probs = np.array([probs])
preds = (probs >= 0.5).astype(int)
tp = int(np.sum((preds == 1) & (y == 1)))
tn = int(np.sum((preds == 0) & (y == 0)))
fp = int(np.sum((preds == 1) & (y == 0)))
fn = int(np.sum((preds == 0) & (y == 1)))
accuracy = (tp + tn) / max(tp + tn + fp + fn, 1)
precision = tp / max(tp + fp, 1)
recall = tp / max(tp + fn, 1)
f1 = 2 * precision * recall / max(precision + recall, 1e-8)
return {
"accuracy": float(accuracy),
"precision": float(precision),
"recall": float(recall),
"f1": float(f1),
}
# ------------------------------------------------------------------
# Main run loop
# ------------------------------------------------------------------
def run(self) -> list:
"""Run full simulation: initialize + all rounds."""
self.initialize()
self.initialize_global_model()
dp_label = f"epsilon={self.dp_injector.epsilon}" if self.use_dp else "OFF"
print(f"\n{'='*60}")
print(f"SentinelEdge Federated Learning Simulation (MLP)")
print(f"Devices: {self.n_devices} | Rounds: {self.n_rounds}")
print(f"Differential Privacy: {dp_label}")
print(f"MLP: {INPUT_DIM} -> 128 -> 64 -> 1")
print(f"{'='*60}\n")
for r in range(self.n_rounds):
result = self.run_round(r)
self.round_results.append(result)
print(f"Round {r+1}/{self.n_rounds}:")
print(f" Accuracy: {result['accuracy']:.4f}")
print(f" Precision: {result['precision']:.4f}")
print(f" Recall: {result['recall']:.4f}")
print(f" F1 Score: {result['f1']:.4f}")
print(f" Devices: {result['n_devices']}")
if self.use_dp:
print(f" Epsilon: {result['epsilon_spent']:.4f}")
print(f" Avg sigma: {result['avg_sigma']:.6f}")
print()
return self.round_results
def run_dp_comparison(n_devices: int = 5, n_rounds: int = 10) -> dict:
"""Run the simulation twice: with DP and without DP.
Returns a dict with keys 'with_dp' and 'without_dp', each containing
the list of round results. Used by visualization.py for comparison plots.
"""
print("=" * 60)
print(" RUNNING COMPARISON: WITH DP vs WITHOUT DP")
print("=" * 60)
# Run WITH DP
np.random.seed(42)
sim_dp = FederatedSimulation(
n_devices=n_devices, n_rounds=n_rounds,
epsilon=0.3, use_dp=True,
)
results_dp = sim_dp.run()
# Run WITHOUT DP
np.random.seed(42)
sim_no_dp = FederatedSimulation(
n_devices=n_devices, n_rounds=n_rounds,
epsilon=0.3, use_dp=False,
)
results_no_dp = sim_no_dp.run()
return {"with_dp": results_dp, "without_dp": results_no_dp}
def main():
parser = argparse.ArgumentParser(
description="Run federated learning simulation with real MLP"
)
parser.add_argument("--devices", type=int, default=5,
help="Number of simulated devices")
parser.add_argument("--rounds", type=int, default=10,
help="Number of federated rounds")
parser.add_argument("--compare", action="store_true",
help="Run DP vs no-DP comparison")
args = parser.parse_args()
output_dir = os.path.dirname(os.path.abspath(__file__))
if args.compare:
comparison = run_dp_comparison(
n_devices=args.devices, n_rounds=args.rounds
)
output_path = os.path.join(output_dir, "simulation_results.json")
serializable = {
"with_dp": _make_serializable(comparison["with_dp"]),
"without_dp": _make_serializable(comparison["without_dp"]),
}
with open(output_path, "w") as f:
json.dump(serializable, f, indent=2)
print(f"\nSaved comparison results to {output_path}")
# Also generate plots
try:
from federated.visualization import (
plot_accuracy_over_rounds, plot_dp_comparison,
)
plot_accuracy_over_rounds(
comparison["with_dp"],
output_path=os.path.join(output_dir, "federated_results.png"),
)
plot_dp_comparison(
comparison,
output_path=os.path.join(output_dir, "dp_comparison.png"),
)
except ImportError:
print("(Skipping plots: matplotlib not available)")
else:
np.random.seed(42)
sim = FederatedSimulation(
n_devices=args.devices, n_rounds=args.rounds
)
results = sim.run()
output_path = os.path.join(output_dir, "simulation_results.json")
serializable = _make_serializable(results)
with open(output_path, "w") as f:
json.dump(serializable, f, indent=2)
print(f"\nSaved results to {output_path}")
# Generate plot
try:
from federated.visualization import plot_accuracy_over_rounds
plot_accuracy_over_rounds(
results,
output_path=os.path.join(output_dir, "federated_results.png"),
)
except ImportError:
print("(Skipping plot: matplotlib not available)")
def _make_serializable(results: list) -> list:
"""Convert numpy types to JSON-serializable Python types."""
out = []
for r in results:
out.append({
k: float(v) if isinstance(v, (np.floating, float)) else v
for k, v in r.items()
})
return out
if __name__ == "__main__":
main()
|