Data Residency

The absolute requirement that protected health information (PHI) never crosses a national border during storage or computation. In federated learning, this means the raw data and the local training loop must remain inside the sovereign territory. Only encrypted, aggregated, or differentially private model updates—which are mathematical abstractions, not human-readable records—are permitted to transit the border to a global aggregation server. This transforms data residency from a blocker into a solved architectural constraint.

Non-compliance with data residency mandates triggers severe financial and operational consequences. Key regulatory frameworks include:

• GDPR (EU): Fines up to 4% of global annual turnover for unlawful data transfers.
• HIPAA (US): Civil penalties up to $1.9M per violation category per year.
• PIPL (China): Fines up to 50M RMB or 5% of prior-year revenue. Federated learning architectures that keep raw data in-region provide a provable compliance posture, directly mitigating these existential financial risks for healthcare consortia.

These terms are often conflated but have distinct meanings:

• Data Residency: The physical, geographic location where data is stored. A purely technical and legal constraint.
• Data Sovereignty: The legal principle that data is subject to the laws of the nation where it resides. This includes government access rights and jurisdictional control. A federated system may satisfy data residency by keeping a copy in-country, but true data sovereignty requires that no foreign entity has the technical ability or legal right to access the raw data, a guarantee provided by local training with secure aggregation.

Data residency mandates extend beyond storage to computation. The model training process itself must execute on infrastructure physically located within the designated jurisdiction. This requires deploying local GPU clusters or edge compute nodes at each hospital site. The federated architecture must support heterogeneous hardware profiles across sites while ensuring the local training runtime never exfiltrates raw pixel data from medical scans to a cloud service operating outside the legal boundary.

The critical legal question for federated learning: Is a model weight update considered personal data? The answer depends on the privacy guarantees applied:

• Raw Gradients: Potentially reversible via model inversion attacks; likely still considered personal data under GDPR.
• Differentially Private Updates: With a proven epsilon budget, these are mathematically guaranteed to not contain individual-level information and are generally considered anonymized, not pseudonymized, data.
• Secure Aggregated Updates: When combined with SecAgg, the server only sees the sum of encrypted updates, never an individual hospital's contribution, satisfying the strictest residency interpretations.

Proving data residency to a regulator requires an immutable, verifiable audit trail. The federated system must log:

• The physical location of every compute node that performed local training.
• A cryptographic attestation from the TEE verifying the integrity of the local training environment.
• A hash of the local dataset used, without exposing the data itself.
• The exact nature of every artifact that crossed a geographic boundary. This tamper-proof log provides the provable chain of custody required for GDPR Article 30 record-keeping and HIPAA technical safeguard attestations.

The legal principle that digital data is subject to the laws of the nation where it is physically collected or stored. Unlike data residency (which specifies geographic location), sovereignty asserts jurisdictional control. For a German hospital, patient data stored on a Frankfurt AWS region falls under GDPR jurisdiction, even if the cloud provider is American. This directly impacts federated learning design, as model updates must never transit through jurisdictions lacking an adequacy decision.

A strict regulatory mandate requiring that data created within a nation's borders be processed and stored exclusively on infrastructure physically located inside that country. This is more restrictive than residency. Key examples include:

• Russia's Federal Law No. 242-FZ: Mandates local databases for citizen data
• China's CSL and DSL: Require domestic storage with security assessments for export
• India's forthcoming DPDP Act: Restricts cross-border transfers of sensitive personal data Localization often forces federated learning nodes to operate within national Trusted Execution Environments.

Legal instruments that permit the controlled movement of data across jurisdictional boundaries when residency requirements allow for exceptions. These are critical for global model aggregation. Primary mechanisms include:

• Standard Contractual Clauses (SCCs): Pre-approved contractual terms for data exporters and importers
• Binding Corporate Rules (BCRs): Internal codes of conduct for multinational corporate data flows
• EU-US Data Privacy Framework: A post-Privacy Shield transatlantic agreement Federated architectures must log every cross-border gradient transfer to maintain an immutable audit trail.

The technical enforcement of data residency through cloud provider controls that restrict processing workloads to specific availability zones or regions. Implementation involves:

• IAM policies with aws:RequestedRegion condition keys
• Azure Policy geolocation restrictions on resource groups
• GCP Resource Location Restriction organization policies In a cross-silo federated network, the aggregation server must be deployed in a jurisdiction compliant with all participating hospitals' residency requirements, often necessitating a sovereign cloud deployment.

A cloud architecture designed to operate entirely within a specific nation's borders, physically isolated from the global cloud provider's default infrastructure. Key characteristics:

• Operated by local, vetted personnel with national security clearances
• Disconnected from foreign control planes to prevent extraterritorial data access
• Examples: AWS European Sovereign Cloud, Oracle EU Sovereign Cloud For medical federated learning, a sovereign cloud acts as the trusted central aggregator, ensuring that even encrypted model updates never leave the mandated jurisdiction.

A legally binding contract required under the HIPAA Privacy Rule between a covered entity (hospital) and any vendor that creates, receives, or transmits Protected Health Information (PHI) on its behalf. For data residency, the BAA must explicitly specify:

• The permitted geographic locations for data storage and processing
• The technical safeguards preventing PHI from being moved to non-compliant regions
• Breach notification timelines tied to specific jurisdictions A federated learning orchestrator handling model updates derived from PHI must execute a BAA with each participating institution, with residency clauses baked into the Data Use Agreement.