Skip to main content

Privacy Technologies (Target Design)

Design and research roadmap

The cryptographic mechanisms on this page describe the privacy architecture SENEX is building toward, not features shipping in v1.0-alpha. Today, privacy is enforced by local-first governance (raw data and locators never leave the device; bounded, revocable neurolink sharing). Treat the parameters below as design targets.

A. Differential Privacy (DP)​

  • Designed to apply at the client level before any data leaves the device
  • Target privacy budget: ε = 1.0, δ = 1e-6
  • 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)