Skip to main content

Privacy Technologies

A. Differential Privacy (DP)​

  • Applied at CLIENT LEVEL before any data leaves the device
  • Privacy Budget: ε = 1.0, δ = 1e-6 (strong privacy guarantee)
  • Mechanism: Gaussian noise calibrated to sensitivity of gradients
  • Composition: Advanced composition for multiple contributions

Implementation:

gradient_noisy = gradient + Normal(0, sigma^2)
# where sigma = (2 * ln(1.25/delta) * delta_sensitivity^2) / epsilon^2
# delta_sensitivity = global sensitivity (max gradient norm)

B. Secure Multi-Party Computation (MPC)​

  • Protocol: SPDZ (Secure Pattern Detection and Zero-knowledge)
  • Participants: N validator nodes (N ≥ 5, threshold = ⌈2N/3⌉)
  • Secret Sharing: Shamir's secret sharing with polynomial degree t = ⌊N/2⌋
  • Operations: Addition and multiplication in encrypted domain

Data Flow:

  1. Client splits noisy gradient into N shares: {s₁, s₂, ..., sₙ}
  2. Each share sent to different validator via encrypted channel
  3. Validators compute f(s₁, s₂, ..., sₙ) = Σ gradients collaboratively
  4. Only aggregated result is revealed, individual shares remain secret

C. Zero-Knowledge Proofs (ZKP)​

  • Type: zk-SNARKs (Zero-Knowledge Succinct Non-Interactive Argument of Knowledge)
  • Purpose: Prove computation correctness without revealing inputs

Applications:

  • Prove gradient computed correctly without revealing local data
  • Prove contribution quality without revealing dataset statistics
  • Prove compliance with privacy budget without revealing parameters

D. Homomorphic Encryption (HE)​

  • Scheme: Partially Homomorphic (Paillier) or Fully Homomorphic (SEAL)
  • Use Case: Encrypted queries to Genome for sensitive inference tasks
  • Operations: Addition and multiplication on encrypted values

E. Federated Learning with Secure Aggregation​

  • Architecture: Cross-silo federated learning (AIA agents = silos)
  • Aggregation: FedAvg with secure aggregation protocol
  • Privacy: Double masking + differential privacy
  • Byzantine Robustness: Krum or Trimmed Mean aggregation

Algorithm:

  1. Each client k computes local gradient gₖ on private data
  2. Add DP noise: g̃ₖ = gₖ + N(0, σ²I)
  3. Apply secure aggregation: G = Σₖ g̃ₖ (computed via MPC)
  4. Global model update: θₜ₊₁ = θₜ - η·G
  5. Broadcast updated model to clients (pull-based)

F. Anonymization Network​

  • Layer 1: TLS 1.3 encryption for all communications
  • Layer 2: Tor-like onion routing or mixnet for submission anonymity
  • Layer 3: Temporal obfuscation (randomized submission times)
  • Layer 4: Network-level unlinkability (different IPs per contribution)