Cross-Silo Federated Learning

Unlike cross-device topologies involving millions of unstable smartphones, cross-silo FL typically involves 2 to 100 institutional clients. Each client is a distinct legal entity with significant computational resources, dedicated hardware, and professional IT management. This ensures high availability and predictable participation in every federated communication round, eliminating the straggler problems common in consumer device training.

Each silo holds a massive, meticulously curated local dataset. In healthcare, this represents years of longitudinal patient records, high-resolution imaging archives, or genomic sequences. These datasets are stateful—they persist, grow, and are actively maintained by the institution. This contrasts sharply with the ephemeral, low-quality data streams typical of mobile devices. The data is assumed to be non-IID across silos, reflecting distinct patient demographics and clinical practices.

Clients leverage institutional-grade hardware, often including high-memory GPUs or Neural Processing Unit acceleration clusters. This allows for training large, complex models like 3D medical imaging segmentation networks that are impossible on edge devices. The topology can also support federated model heterogeneity, where different hospitals train slightly different local architectures optimized for their specific hardware stack, sharing only the knowledge via federated distillation rather than raw weights.

While institutional trust exists, technical privacy guarantees are still mandatory. Cross-silo FL commonly employs federated secure aggregation protocols. This cryptographic technique ensures the central server can only compute the sum of all model updates, remaining mathematically incapable of inspecting any single hospital's contribution. This protects against inference attacks on the gradient updates, providing a verifiable privacy layer on top of the foundational federated data locality principle.

Radiology departments across several hospitals collaboratively train a convolutional neural network (CNN) for brain tumor segmentation on MRI scans. Each hospital holds a distinct, curated dataset of annotated images. Using Federated Averaging (FedAvg) , only encrypted model weight updates are sent to a central parameter server, never the raw DICOM images. This approach yields a model that generalizes across diverse scanner vendors and imaging protocols, achieving diagnostic accuracy comparable to a model trained on a centralized, pooled dataset while maintaining full HIPAA compliance.

A network of regional hospitals uses a federated gradient-boosted tree model to predict 30-day patient readmission risk from Electronic Health Records (EHR) . Each hospital's data has the same feature schema (e.g., lab results, vitals, demographics) but for different patient populations, representing a horizontal federated learning scenario. The global model learns complex, non-linear risk factors from a vastly larger and more diverse patient cohort than any single hospital could access, enabling more accurate resource planning and early intervention without exposing individual patient histories.

Pharmaceutical companies and research hospitals in different jurisdictions form a federated consortium to train a graph neural network (GNN) for predicting molecular toxicity. Each partner contributes proprietary, highly sensitive chemical assay data. A secure aggregation protocol ensures the central server can only compute the sum of model updates, making it mathematically impossible to reverse-engineer any single partner's contribution. This accelerates the identification of viable drug candidates while protecting intellectual property and complying with divergent regulations like GDPR and local data sovereignty laws.

Clinicians seeking patients for a rare disease clinical trial use a federated query engine across a network of academic medical centers. Instead of moving data to a central warehouse, the query is distributed to each site's local FHIR-compliant data node. Each node returns only aggregate counts of matching patients, preserving privacy. Once a sufficient cohort size is confirmed, the same infrastructure can be used to train a federated survival analysis model to identify prognostic biomarkers, dramatically accelerating research for conditions where no single institution has enough data.

A topology where datasets share the same feature space but differ in sample space. This is the most common paradigm in healthcare, where multiple hospitals record the same clinical measurements (labs, vitals) for different patients.

• Use Case: Three hospitals training a readmission prediction model using identical EHR fields.
• Key Assumption: Feature alignment across all nodes.
• Aggregation: Typically uses Federated Averaging (FedAvg).

A topology where datasets share the same sample space but differ in feature space. This enables collaboration between a hospital and an insurance company holding different attributes for the same patients.

• Mechanism: Uses entity alignment and split neural networks.
• Privacy: Employs homomorphic encryption for secure gradient computation.
• Challenge: Requires overlapping identifiers without exposing them.

A multi-tier topology introducing intermediate edge aggregators between clients and the central server. This reduces communication latency and improves scalability for large hospital networks.

• Architecture: Client → Edge Aggregator → Central Server.
• Benefit: Reduces wide-area network traffic by aggregating updates regionally.
• Example: A state-level aggregator combining updates from county hospitals before sending to a national model.

A peer-to-peer topology where clients communicate model updates directly with each other without a central aggregation server. This eliminates the single point of failure and trust bottleneck.

• Protocol: Uses gossip algorithms or blockchain consensus.
• Resilience: No central server to attack or fail.
• Trade-off: Increased communication complexity and convergence time.

A privacy-preserving architecture where a deep neural network is partitioned between a client and a server. Only intermediate activations (smashed data) and gradients are exchanged, never raw data.

• Configuration: Client holds initial layers; server holds remaining layers.
• Advantage: Client never shares raw data or full model parameters.
• Use Case: Resource-constrained hospitals training large vision models.

A fully decentralized framework using blockchain-based coordination to enable secure, resilient collaborative training without a central coordinator. Smart contracts manage membership and model aggregation.

• Consensus: Permissioned network validates updates.
• Security: Immutable audit trail of all contributions.
• Application: Multi-institutional genomic analysis with strict governance requirements.