Researchers present FLA3, a federated learning platform designed to enforce governance rules during multi‑centre clinical research. The paper tackles a practical problem: privacy rules and institutional policies often bar sharing patient data across borders. Federated learning (FL) lets models train without moving raw data, but many existing FL systems do not enforce who is allowed to take part, when participation is valid, or how activity is audited. FLA3 adds those controls as built‑in runtime checks.

To do this the authors built governance controls into the FL orchestration layer. They integrated eXtensible Access Control Markup Language (XACML)‑compliant attribute‑based access control (ABAC) — a way to write policy rules that check attributes such as role, site, or approval expiry — and added cryptographic accounting to record who did what. They also implemented “study‑scoped federation,” so a model run is tied to a specific approved study. The platform extends the open‑source Flower framework and is aimed at meeting institutional and legal requirements in real deployments.

The team evaluated FLA3 in two ways. First, they deployed the infrastructure across five BloodCounts! Consortium institutions in four countries: United Kingdom, Netherlands, India and The Gambia. This demonstrated the system can run under differing network and regulatory environments. Second, they tested clinical utility with a simulated federation using full blood count (FBC) data from the INTERVAL study: 54,446 samples from 35,315 subjects across 25 centres. In that experiment FLA3 achieved predictive performance comparable to a centrally trained model while enforcing the governance constraints set by the platform.