Skip to content

CONCEPT Cited by 2 sources

Federated learning

Definition

Federated learning (FL) is a machine-learning setting where multiple entities (clients — typically end-user devices) collaborate to solve an ML problem under the coordination of a service provider, without centralizing the raw training data. Instead of shipping user data to a central server, the training algorithm is brought to the data: clients compute on their local data and only privacy-reduced artifacts (model-gradient updates, or — in newer designs — encrypted examples processed inside attested enclaves) leave the device.

Google's evolved (2025) definition centers FL on four privacy principles: (1) data minimization, (2) data anonymization, (3) transparency and control, and (4) verifiability and auditability. A complete FL system should let clients retain full control over their data, the set of workloads allowed to access it, and the anonymization properties of those workloads. (Source: sources/2026-10-02-google-toward-provably-private-learning-from-federated-data)

Why it exists

The canonical motivating problem (from Google's 2017 FL work) is improving an on-device model — e.g. Gboard's next-word-prediction language model — from real user behavior without ever shipping keystrokes off-device. FL provides the mechanism: local model updates aggregated across a fleet, with anonymization layers (secure aggregation, differential privacy) bounding what the aggregate can reveal about any one user. Production FL at Google has powered Gboard next-word prediction and Smart Compose, Google Messages reply suggestions, and Android Smart Text Selection. (Source: sources/2026-10-02-google-toward-provably-private-learning-from-federated-data)

Generations of FL architecture

The anonymization/trust model has evolved through distinct generations (Source: sources/2026-10-02-google-toward-provably-private-learning-from-federated-data):

Generation Where compute runs Trust model Weakness addressed next
Gen 1 (2017) client-side aggregation Gradients computed on-device; server aggregates Server could log/inspect uploads; no external verification No verifiability
+ Secure Aggregation On-device, cryptographically masked uploads Server sees only the sum Incompatible with state-of-the-art central DP (MF-DP-FTRL)
Gen (TEE-based, this post) server-side verifiable training Devices upload encrypted examples; the full training loop runs server-side inside attested TEEs Operator trust removed via attestation + transparency log; only DP weights leave the enclave Current-gen TEE / side-channel limits; formal proofs still future work

The Gen-TEE design (see systems/confidential-federated-compute) is a notable inversion: it moves gradient computation off the device and onto the server, trading the device-side trust model for a hardware-rooted, externally verifiable server-side one. This also removes on-device-compute and diurnal device-availability bottlenecks, cutting Gboard model training from 1–2 months to TEE-availability-limited. (Source: sources/2026-10-02-google-toward-provably-private-learning-from-federated-data)

Relationship to federated analytics

Federated analytics is the same decentralized-data, server-coordinated pattern applied to analytics/aggregate statistics rather than model training. Google's Confidential Federated Analytics is the aggregation-direction sibling of the FL training system; both compose TEEs + attestation + transparency logs over the same device fleet. (Source: sources/2026-05-27-google-private-analytics-via-zero-trust-aggregation)

What FL is not

  • Not inherently private. Raw gradients can leak training data; FL requires explicit anonymization layers (secure aggregation and/or DP) to bound leakage.
  • Not "no server." A service provider still coordinates rounds, aggregates, and (in the TEE generation) runs the training loop; FL constrains what the server can see, not whether a server exists.
  • Not free of the training/serving boundary — FL redraws where training compute happens but the serving model still ships to devices.

Seen in

Last updated · 771 distilled / 2,233 read