Skip to main content
completed federated-learning-security

SignGuard: Cryptographic FL Defense

Novel multi-layer defense system combining ECDSA digital signatures, multi-factor anomaly detection, and time-decay reputation scoring to protect federated learning from Byzantine poisoning attacks.

Started: January 15, 2025
Completed: February 15, 2025
94.5%
Attack Detection Rate
3.2%
False Positive Rate
<0.2%
Accuracy Degradation

Tags

federated-learningsecuritycryptographyecdsabyzantine-robustnessreputation-systemsanomaly-detection

Technologies

Python PyTorch Flower (Flwr) ECDSA (secp256k1) cryptography.io NumPy scikit-learn

Overview

SignGuard addresses a core weakness in federated learning: how to verify that client model updates are genuine and untampered. Standard FL aggregation (FedAvg) assumes all participants are honest — a dangerous assumption in real-world deployments where compromised clients can inject poisoned gradients, plant backdoors, or degrade model performance for all participants.

SignGuard uses a three-layer defense architecture that combines cryptographic integrity verification, statistical anomaly detection, and behavioral reputation tracking to achieve 94.5% attack detection while preserving model accuracy within 0.2% of undefended performance on clean data.

Three-Layer Architecture

Layer 1: ECDSA Digital Signatures

Every model update is signed with the client’s private key using ECDSA on the secp256k1 curve. The server verifies each signature before processing, which:

  • Prevents impersonation: A client can’t submit updates pretending to be another
  • Ensures integrity: Tampered updates fail signature verification
  • Provides non-repudiation: Clients can’t deny sending a particular update
  • Detects replay attacks: Nonces prevent resubmission of old updates

Performance overhead is minimal — less than 5ms per update verification on P-256.

Layer 2: Multi-Factor Anomaly Detection

Authenticated clients can still send malicious updates. This layer applies three complementary detection techniques:

  • Magnitude analysis: Compares L2 norm against a dynamic baseline — catches scaling attacks
  • Direction analysis: Cosine similarity against global gradient direction — catches backdoor and model poisoning attempts
  • Loss-based analysis: Validates reported training loss against expected values — catches subtle attacks where magnitude and direction appear normal

Layer 3: Time-Decay Reputation Scoring

Maintains a reputation score (0–1) for each client with exponential decay weighting:

  • Recent behavior is weighted more heavily than historical actions
  • Clients flagged for anomalies face stricter scrutiny on future rounds
  • Reputation recovery is possible through consistent, verified contributions
  • Low-reputation clients can be automatically quarantined

Results

Tested against four major attack categories in a simulated federated learning environment:

Attack TypeWithout DefenseWith SignGuardImprovement
Sign Flipping23% detection94.5% detection+71.5 pp
Label Flipping15% detection89.2% detection+74.2 pp
Backdoor Injection<10% detection76.8% detection+66.8 pp
Model Poisoning18% detection91.3% detection+73.3 pp

Key metrics:

  • Combined attack detection rate: 94.5%
  • False positive rate: 3.2%
  • Clean data accuracy degradation: <0.2%
  • Signature verification time: <5ms per update
  • Reputation convergence: ~10 rounds for stable scoring

Design Decisions

Why ECDSA over HMAC? ECDSA provides non-repudiation — the server can prove a specific client sent a specific update. HMAC only proves the update was authenticated, not who sent it.

Why multi-factor anomaly detection? No single statistical test catches all attack types. Magnitude analysis misses sign flipping (same norm), direction analysis misses subtle backdoors (small divergence), and loss-based analysis misses model poisoning (loss may not change immediately). The combination covers the gaps.

Why time-decay reputation? A compromised client that was previously honest should lose trust quickly. Conversely, a quarantined client should be able to rebuild trust through verified behavior. Exponential decay naturally handles both cases.

Honest Limitations

  • No privacy guarantees: Signatures authenticate updates but don’t hide their contents. Requires DP or secure aggregation for privacy.
  • Trusted server assumption: Assumes the aggregation server is honest. Fully decentralized scenarios would need consensus mechanisms.
  • Adaptive attackers: Patient attackers who slowly shift model behavior over many rounds can partially evade detection. The reputation system helps but isn’t perfect.
  • Non-IID sensitivity: Extremely heterogeneous data distributions across clients make anomaly detection thresholds harder to calibrate.