Federated AI

The collaborative-learning layer: federated learning, privacy and evaluation.

• Federated learning — Many clients train a shared model under orchestration while data stays decentralised.
• Secure aggregation — The server learns the aggregate of updates, not any individual update.
• Cross-device vs cross-silo — Large fleets of weak devices vs a few institutions with legal separation.
• Federated analytics — Distributed statistics and monitoring without training a model.
• Federated evaluation — Measuring models when data and outcome definitions stay distributed.
• Non-IID data — Heterogeneity is intrinsic, not incidental, in federated settings.
• Differential privacy — A formal bound on what any individual contributes to a released result.
• Trusted execution environments (TEEs) — Hardware-isolated computation strengthening server-side confidentiality.
• Homomorphic encryption & MPC — Computing over encrypted data / across parties without revealing inputs.
• Membership inference & gradient leakage — Privacy attacks that recover information from model updates.
• Model poisoning & backdoors — Malicious participants corrupt the shared model.
• Federated unlearning — Removing a participant's influence from a trained model.
• Federated / confidential RAG — Applying the federated pattern to LLM-era retrieval and generation.
• Multi-dimensional evaluation rubric — Judge federations on utility and governance, not accuracy alone.
• Hub-and-spoke vs hierarchical coordination — Centralised aggregation vs hierarchical/asynchronous/enclave-backed designs.