What is Federated Analytics?

1.1 Overview

In FA, analytical queries are executed locally within each participating organisation, and only aggregated results are shared for analysis. Deployments typically involve a coordinating service that distributes queries, participating data nodes that execute computations locally, and an aggregation mechanism that combines results from multiple organisations.

A typical FA query-based workflow involves the following steps:

Query submission: An authorised user submits an analytical query through a shared interface. In practice, this may be restricted to predefined query types or parameterised templates to ensure consistency and compliance.
Query distribution: The coordinating system distributes the query to participating organisations. Depending on the architecture, this may occur synchronously (all nodes execute together) or asynchronously (nodes respond independently based on availability).
Local computation: Each organisation executes the query locally on its dataset. This step may involve local preprocessing, filtering, or transformation based on agreed data definitions and schemas.
Aggregation: Local outputs are returned and aggregated into a combined result. Aggregation may be performed centrally or through privacy-preserving techniques (e.g., secure aggregation or encryption-based methods), depending on trust assumptions.
Result release: Governance checks may be applied before results are released. These checks can include disclosure controls (e.g., minimum thresholds), validation against permitted query scope, and approval workflows where required.

Beyond this core workflow, FA systems are typically designed with additional layers to support governance, scalability, and reliability in multi-party environments. These include:

Access control and query authorisation, to ensure only approved users can submit queries
Predefined query templates or constraints, to limit analytical scope and reduce disclosure risks
Audit logging and monitoring, to track query execution and support accountability across participants

In practice, FA is a good fit as a privacy solution where:

Data cannot be easily centralised due to legal, policy, operational, or trust constraints;
The required outcome is descriptive or comparative analysis, rather than a predictive model; or
Participants need to assess data availability, quality, or comparability before pursuing deeper collaboration.

1.2 Types of Analytics Supported

FA is most suitable for analytical tasks that can be expressed as distributed statistical computations, where the required outputs can be decomposed into local computations and combined into an aggregated result. Because results are derived from a single analytical workflow rather than iterative model training, these analyses typically require less coordination between participants than FL.

Analytical task	Description	Example outputs
Descriptive statistics	Basic summaries of distributed datasets	Counts, averages, percentiles, distributions
Trend analysis	Identifying patterns or changes over time across organisations	Monthly case counts, service utilisation trends
Correlation analysis	Measuring relationships between variables across datasets	Correlation coefficients, covariance matrices
Set comparisons	Identifying overlaps or differences between datasets held by different organisations	Number of shared entities, intersection sizes
Statistical modelling	Certain statistical models can be computed using aggregated intermediate statistics	Linear or logistic regression coefficients