Federated Learning

src/lib/federated-learning.ts lets the learning system improve across users without sharing individual data.

The protocol

1. Each client trains locally on its own quiz observations
2. The client computes a BKT parameter delta (not raw data)
3. The delta’s sensitivity is bounded by gradient clipping (clipNorm)
4. Gaussian noise is added for (ε, δ)-differential privacy
5. The noised delta is sent to the server
6. The server aggregates deltas via weighted averaging into a GlobalModel

Key types

export interface FederatedUpdate { /* gradient delta with DP noise */ }
export interface BKTParamDelta { /* per-parameter adjustments */ }
export interface GlobalModel { /* aggregated state across clients */ }
 
export interface PrivacyParams {
  epsilon: number    // privacy budget ε
  // delta δ
  clipNorm: number   // gradient clipping norm
  // noise scale — computed from ε, δ, clipNorm
}

Guarantees

• What leaves the device: a clipped, noised parameter delta — never observations, video content, or identity
• Formal privacy: Gaussian-mechanism (ε, δ)-DP with sensitivity bounded by clipNorm
• Utility: weighted averaging favors clients with more observations, so the global BKT parameters converge toward population-realistic values