Federated computing helps with GDPR compliance by keeping raw records inside each participating organisation’s environment. Approved code travels to each site, and only results such as model updates or aggregate statistics leave, so the cross-border transfer analysis covers those results instead of whole datasets. Legal basis, controller roles, data subject rights, and impact assessments still apply in full.
Training one model on patient records held by hospitals in Germany, France, and the US looks like a data engineering problem. Under the General Data Protection Regulation (GDPR), the project starts as a transfer problem. Copying EU personal data into a central environment outside the European Economic Area (EEA) brings in Chapter V of the regulation, along with adequacy checks, contractual clauses, and transfer assessments.
This article is general education for data, privacy, and engineering teams and does not constitute legal advice. How GDPR applies to a specific project depends on facts your data protection officer (DPO) and counsel should assess.
Where GDPR Creates Friction for Multi-Site Data Analysis
Three parts of GDPR shape most multi-site analytics projects, and each becomes harder to manage when records from many organisations sit in one place.
International Transfers
Chapter V of GDPR (Articles 44 to 50) restricts transfers of personal data to countries outside the EEA. The simplest route is an adequacy decision under Article 45, which the European Commission has granted to a limited set of countries and frameworks, including Japan, the UK, and the EU-US Data Privacy Framework for certified US companies. Without one, teams rely on Article 46 safeguards such as standard contractual clauses (SCCs) or binding corporate rules.
SCCs come with extra work. After the Court of Justice's 2020 Schrems II ruling, the European Data Protection Board (EDPB) set out a six-step method in its Recommendations 01/2020. Exporters map the transfer, identify the transfer tool, and check whether the destination country's law or practice undermines it, a step usually called a transfer impact assessment (TIA). Where the assessment finds gaps, exporters add supplementary measures.
Controller Roles
When several organisations jointly decide why and how data is processed, they are joint controllers under Article 26 and need an arrangement that sets out each party's duties. The EDPB's Guidelines 07/2020 explain how that works in practice. Pooling records from many hospitals in one central environment makes those lines harder to draw, and teams building federated research consortia still have to settle them.
Purpose Limitation
Article 5(1)(b) bars further processing that is incompatible with the purpose the data was collected for, and records gathered for patient care and reused to train a model raise that question directly. Article 6(4) lists the factors for the compatibility test, Article 89 sets safeguards for research, and the EDPB's draft Guidelines 1/2026 on scientific research explain how research projects meet these rules. Health data is also a special category under Article 9, so it needs its own condition for processing. Our plain-English guide to secure data sharing covers the question of "Can we use this data?" in more depth.
What Changes When Analysis Happens In Place
In a federated setup, approved code runs inside each site's environment, and only results such as model updates or aggregate statistics return to a coordinating server. Picture a US sponsor training a model on data from three EU hospitals. Patient records stay behind each hospital's firewall, and the sponsor receives model parameters.
That design changes the GDPR picture in three ways.
- Smaller transfer scope: Raw records stay inside the EEA, so Chapter V applies only to outputs that leave each site, if they still count as personal data, and to any access to the node from outside the EEA.
- Data protection by design: Article 25 asks controllers to build safeguards into how processing works, and Article 5(1)(c) asks them to minimise the data they process. The European Data Protection Supervisor (EDPS) and the Spanish data protection authority examined federated learning in TechDispatch #1/2025 and urged data protection by design and by default, especially for health and biometric data.
- Local control: Sites decide who can access their data, can limit their node to code they've pre-approved, and control who sees logs from runs on their data.
Running this across dozens of sites takes a standing deployment at each one, a common data model so the same code runs everywhere, and central orchestration. The Rhino Federated Computing Platform (Rhino FCP) provides that layer, and its Data Harmonization Engine maps local records to standards such as FHIR so a single analysis can run at every site.
What In-Place Analysis Does Not Change
The rest of GDPR still applies to a federated project, and six obligations deserve the closest attention.
- Outputs can still be personal data: Model updates and small-count statistics can reveal information about individuals. In Opinion 28/2024, the EDPB said AI models trained on personal data cannot always be considered anonymous and that anonymity must be assessed case by case. Its draft Guidelines 02/2026 on Anonymisation describe how to make that assessment. If outputs remain personal data and the coordinating server sits outside the EEA, Chapter V applies to that flow.
- Remote access can count as a transfer: The EDPB's Recommendations 01/2020 treat remote access from a third country, such as access by a support provider, as a transfer. Who can reach a node, and from where, matters as much as where the data sits.
- Legal basis and purpose stay the same: Each site still needs an Article 6 legal basis, an Article 9 condition for health data, and a compatibility assessment for any new purpose.
- Controller roles still need defining: Sites and a coordinating sponsor can be joint controllers even when no records are pooled, and the platform provider's role belongs in the contract too.
- Data subject rights still apply locally: Each site remains responsible for access, rectification, erasure, and other requests under Articles 15 to 22 for the records it holds. Participants should also agree how a request at one site affects models already trained on that data.
- A DPIA is still likely: Large-scale processing of health data usually calls for a data protection impact assessment (DPIA) under Article 35, and the EDPS TechDispatch stresses DPIAs for high-risk federated learning projects.
Privacy-enhancing technologies reduce what outputs reveal. Secure aggregation lets the coordinating server combine updates without seeing any single site's contribution, and differential privacy adds calibrated noise to limit what anyone can learn about one person. On the Rhino FCP, teams set a differential privacy level for federated analytics when they create a project and can turn on homomorphic encryption for NVIDIA FLARE training runs, so the coordinating server aggregates encrypted model updates.
A Checklist for Evaluating Federated Approaches
Data protection officers, architects, and site IT teams can use these questions before a federated project goes live.
- Residency: Do raw records, temporary files, and logs stay inside each site's environment during every job?
- Outputs: What exactly leaves each node, and has anyone assessed whether it could identify a person?
- Safeguards on outputs: Is secure aggregation in place, and can the team apply differential privacy or minimum cell sizes where outputs are sensitive?
- Code control: Who can submit code to a node, and does the site approve it before it runs?
- Coordinator location: Where is the coordinating server hosted, and who can access it from outside the EEA?
- Roles: Are controller and processor roles, including any Article 26 arrangement, documented for every participant?
- Legal basis and purpose: Does each site have an Article 6 basis, an Article 9 condition, and a compatibility assessment for this use?
- DPIA: Has a data protection impact assessment been completed or updated for the project?
- Accountability: Do audit logs record queries, training runs, and output exchanges, so each site can show regulators what happened?
For the site-level view, see how healthcare IT teams deploy federated AI without disrupting local analytics. For the same questions under US and EU health data rules, see our posts on HIPAA-compliant queries and the EHDS deadline.
Settle the Transfer Question Before the First Job Runs
A federated project becomes defensible under GDPR once the team agrees on what leaves each site and documents why. The legal basis, controller roles, and DPIA remain the team's responsibility either way.
The Rhino FCP runs analytics and model training inside each participant's environment, so raw records stay with the organisation that holds them. Teams build on it through Rhino Data Activation, Rhino Federated Intelligence, and Rhino Private Agentic AI. For GDPR-bound projects, consortia use Rhino Federated Intelligence to train and evaluate AI across sites without pooling records. The Rhino FCP runs in more than 200 deployments worldwide, including 14 of the top 20 Newsweek Best Smart Hospitals.
Planning a multi-site project that crosses EU borders? Bring this checklist to your next project review, and talk with the Rhino team about how each item maps to your deployment.
Frequently Asked Questions
Does Federated Computing Replace Standard Contractual Clauses?
It reduces how much depends on them. If outputs that leave a site remain personal data and go to a country without an adequacy decision, the transfer still needs an Article 46 tool such as SCCs, and your DPO and counsel make that call for each project.
Are Model Updates Personal Data Under GDPR?
They can be. In Opinion 28/2024, the EDPB found that models trained on personal data are not automatically anonymous, and the updates that build those models raise the same question. Teams should assess identifiability case by case.
Does Analysing Data In Place Satisfy Purpose Limitation?
Location does not settle it, because purpose limitation concerns why data is used. Federated controls help enforce an approved purpose, and the compatibility assessment or a fresh legal basis is still required.
What Should a DPO Review Before Approving a Federated Project?
A DPO should review the DPIA, a description of everything that leaves each node and whether it could identify anyone, the safeguards on those outputs, the coordinating server's location and access controls, the controller arrangements, and the audit logs. If outputs cross borders as personal data, the transfer assessment belongs in the review too.