Skip to main content

Security Threat Model & Mitigations

THREAT 1: Data Poisoning Attacks​

Attack: Malicious clients submit crafted gradients to degrade model performance

Mitigation:

  • Byzantine-robust aggregation (Krum, Trimmed Mean, Median)
  • Reputation system: Track contribution quality over time
  • Outlier detection: Statistical tests on gradient distributions
  • Slashing: Penalize validators who accept obvious poisoned data

THREAT 2: Model Inversion Attacks​

Attack: Adversary attempts to reconstruct training data from model gradients

Mitigation:

  • Differential privacy (ε=1.0) provides provable protection
  • Gradient clipping before noise addition (||g|| ≤ C)
  • Secure aggregation prevents access to individual gradients
  • Only aggregated updates available, never individual contributions

THREAT 3: Membership Inference Attacks​

Attack: Determine if specific data point was in training set

Mitigation:

  • Differential privacy fundamentally prevents this (δ = 1e-6)
  • Model checkpoints versioned, old versions retired
  • Privacy budget tracking per client across all contributions

THREAT 4: Sybil Attacks​

Attack: Single entity creates many fake identities to gain influence

Mitigation:

  • Proof-of-Stake: Requires token stake for validator participation
  • Identity verification for high-value contributors (optional tier)
  • Reputation weighting: New contributors have lower influence
  • Economic disincentive: Costs more to attack than potential gain

THREAT 5: Inference Attacks on Encrypted Data​

Attack: Pattern analysis on encrypted communications or timing attacks

Mitigation:

  • Mixnet/onion routing eliminates network-level tracking
  • Randomized submission times (uniform distribution over time window)
  • Dummy traffic to obscure real contributions
  • Constant-time operations to prevent timing side-channels

THREAT 6: Malicious Validators​

Attack: Compromised validators attempt to learn client data or manipulate results

Mitigation:

  • MPC threshold: Requires ⌈2N/3⌉ honest validators (Byzantine fault tolerance)
  • Secret sharing: No single validator sees complete information
  • Slashing: Validators lose stake if caught cheating
  • Verifiable computation: ZK proofs ensure correct execution

THREAT 7: Smart Contract Vulnerabilities​

Attack: Exploit bugs in smart contracts to steal tokens or manipulate model

Mitigation:

  • Formal verification of mission-critical contracts
  • Multi-signature governance for contract upgrades
  • Bug bounty program (10% of TVL reserved)
  • Gradual rollout: Testnet → limited mainnet → full deployment
  • Circuit breakers: Auto-pause on anomalous activity

THREAT 8: Collusion Between Clients and Validators​

Attack: Coordinated attack to manipulate model or steal rewards

Mitigation:

  • Economic game theory: Defection more profitable than cooperation
  • Random validator assignment per aggregation round
  • Reputation slashing for detected collusion patterns
  • Whistleblower rewards from slashed stakes