Federation operations

Federated operations are a distributed supply chain

Federated learning or adaptation moves computation to data, but it also creates many partially trusted producers. Operations must answer which client trained what, on which parent, under which policy, and whether its update is admissible.

Client lifecycle

1. Register device or site identity and trust tier.
2. Attest runtime, parent package, and training code.
3. Select clients by availability, data relevance, fairness, and resource state.
4. Issue short-lived round credentials and an update contract.
5. Validate local gates before upload.
6. Authenticate, clip, and anomaly-screen updates.
7. Aggregate under a recorded strategy.
8. Evaluate the resulting global candidate centrally.
9. Distribute only signed approved packages.
10. Support site-level rejection and rollback.

Staleness

Updates trained from old parents can conflict with current models. Include parent artifact ID and round ID. Reject or transform stale updates according to explicit policy; never average them silently.

Heterogeneous clients

Different model widths, adapters, or hardware profiles require alignment rules. Aggregate only compatible parameter subsets or use distillation and behavioral exchange. Contract compatibility does not imply weight compatibility.

Poisoning and anomalies

Use robust aggregation, update norm limits, similarity analysis, client reputation, secure aggregation where applicable, and challenge rounds. Secure aggregation protects update privacy but can make anomaly inspection harder; choose the threat model explicitly.

FOR each update IN submitted_updates
    REQUIRE SIGNATURE_VALID(update)
    REQUIRE update.parent_id == round.parent_id
    REQUIRE update.schema_version == round.schema_version
    REQUIRE update.norm <= round.max_norm
    REQUIRE CLIENT_ATTESTATION_VALID(update.client_id)
END FOR

filtered <- ROBUST_FILTER(submitted_updates)
candidate <- AGGREGATE(filtered)
SEND_TO_INDEPENDENT_EVALUATION(candidate)

Revocation

Be able to exclude a client, signer, site, update, or dataset ancestor and identify all descendants affected. A global candidate produced from a compromised round may require full rollback and re-aggregation.

Federation metrics

Track selected versus participating clients, update rejection, staleness, communication bytes, site divergence, poisoning alerts, fairness of participation, global and local quality, and recovery after site loss.