What is secure aggregation?

A protocol letting the coordinator sum updates without seeing any individual one, limiting what the center can infer.

What is differential privacy's role?

Adding calibrated noise to updates so individual records cannot be reverse-engineered from them.

Is federated learning only for phones?

No; cross-organization federation between firms or institutions is a growing use, wherever data sharing is blocked.

AI Glossary / AI and machine learning / Advanced

Federated learning

Definition

Federated learning is training a shared model across many devices or organizations without moving their raw data to one place: the training goes to the data, and only model updates come back.

What it means

Classic training centralizes: gather everyone's data in one warehouse and train there. That design fails when data cannot move, for privacy law, competition, geography or sheer size.

Federated learning inverts the logistics. Each participant, a phone, a hospital, a branch, trains the current model on its own data locally, and sends back only the model's adjustments, never the records.

A coordinator averages the adjustments from many participants into an improved shared model, which is sent back out for the next round. The data never leaves home; only mathematics travels.

The approach buys privacy and compliance at real cost: updates can still leak hints about the underlying data, participants vary wildly in data quality and availability, and coordinating thousands of intermittent devices is an engineering discipline of its own.

For a business, federated learning matters when a consortium or a customer base would collectively benefit from a shared model but no one will hand over the raw material, a common standoff it is purpose-built to break.

A quiet governance benefit: because raw data never centralizes, the breach surface shrinks with it, and several regulators have looked more kindly on federated designs for exactly that reason.

How it works

Distribute the current shared model to participating devices or sites.

Each participant trains the model on its local data for a few steps, producing a local update.

Participants send back only the updates, often with added privacy protections such as secure aggregation or injected noise.

The coordinator averages the updates into a new shared model and redistributes it, repeating for many rounds.

Evaluate the shared model centrally and per participant, because a good average can still serve minority participants badly.

Handle dropouts as normal life: participants join and leave rounds constantly, so the protocol must tolerate missing contributors without stalling or corrupting the average.

Three practical examples

Phone keyboards: improving next-word prediction from typing that never leaves the handset.
Hospital networks: training a diagnostic model across hospitals that cannot share patient records.
Bank consortia: building a shared fraud model from patterns no member will disclose to competitors.

Think of it this way

Think of a recipe improvement society whose members will not reveal their kitchens. Each member cooks the current recipe at home, reports only what they changed and how it tasted, and the society publishes an averaged improvement. The dishes improve without anyone's kitchen being inspected.

The limit: you taste the reports, not the food. A member with a strange oven skews the average, and the society cannot see why, which is the federated condition in miniature.

Worked example

training across four hospitals

Suppose four hospital groups want a shared scan-triage model, but patient records cannot cross organisational boundaries. The figures are illustrative.

They federate: each hospital trains the shared model on its own scans for a week, and only parameter updates flow to the coordinator.

After forty rounds the shared model beats every hospital's solo model, because it has effectively seen four patient populations.

One small hospital's model lags: its scans come from older machines, so the group adds a per-site fine-tuning step for the outliers.

Illustrative case study

A fictional retail consortium federated a demand model across members and celebrated the privacy story, until a sharp member demonstrated that a competitor's promotional calendar could be partially inferred from its weekly updates.

The fix was differential-privacy noise on the updates, which cost a little accuracy and restored the secret.

The invented lesson: not sending the data is not the same as revealing nothing. Updates talk, and privacy engineering decides how much they say.

The invented follow-up: the consortium now publishes its update-auditing method to members annually, because in a federated scheme, trust in the aggregation process is the membership fee.

Common misconceptions and mistakes

Federated learning makes data fully private. It removes the raw-data transfer, but updates can leak statistical hints; true privacy needs the added protections.
It needs exotic infrastructure everywhere. The pattern runs at many scales, from phone fleets to a handful of hospitals with ordinary servers.
The shared model serves everyone equally. Average updates favor the majority's data; minority participants often need local fine-tuning on top.

Business use

Consider federation when the barrier to a shared model is data movement, not data existence; it solves exactly that problem and no other.

Budget for the privacy engineering on top: secure aggregation and noise injection are what turn data never moves into nothing is revealed.

Evaluate per participant, not just in the average, and plan fine-tuning for members whose data is genuinely different.

Check your understanding

Why does federated learning suit a consortium that will not pool raw data?

Answer: Because training happens where the data lives and only model updates travel, so members gain a shared model without surrendering their records.

Related terms

Privacy
Differential privacy
Data governance
Transfer learning
distributed training
secure aggregation

Learn next

AI for Business Owners explains how to weigh privacy architectures when several parties need one model: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Privacy

Sources

[S1] Russell and Norvig, AIMA glossary; [S2] Google, Machine Learning Glossary; [S3] NIST AI Resource Center glossary; [S12] Aalto University, Dictionary of Machine Learning; [S22] Berkeley BOP, Artificial Intelligence Ontology (aio.owl). Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Federated gradient descent Next: Federated learning network