Cross-Silo Federated Learning

The topology involves a limited number of identified, reliable organizations, typically ranging from 2 to fewer than 100. Unlike massive, anonymous cross-device fleets, every participant is known and authenticated.

• Identity Management: Relies on institutional Public Key Infrastructure (PKI) rather than anonymous device IDs.
• Stable Connectivity: Participants are expected to be almost always available for training rounds, eliminating the need for complex device selection logic.
• Contractual Trust: Collaboration is often governed by legal agreements (e.g., Data Use Agreements) that complement technical privacy safeguards like differential privacy.

Each silo possesses enterprise-grade hardware and large, curated, domain-specific datasets. This contrasts sharply with cross-device FL, where clients are resource-constrained smartphones.

• Local Training Capacity: Silos can perform full local epochs of training on powerful GPU clusters before sharing updates.
• Data Volume: A single hospital silo may hold millions of structured electronic health records or whole-genome sequences.
• Stateful Clients: Silos maintain persistent local state, enabling advanced optimization techniques like local momentum or adaptive learning rates across rounds.

Cross-silo FL typically employs a centralized hub-and-spoke architecture where a designated aggregation server coordinates the training lifecycle.

• Aggregation Server: A trusted or semi-trusted central node that initializes the global model, distributes it to silos, and aggregates returned updates using algorithms like Federated Averaging (FedAvg).
• Round-Based Synchronization: Training proceeds in discrete synchronous rounds, where the server waits for all or a quorum of silos to respond before updating the global model.
• Secure Aggregation Protocols: The server can aggregate encrypted model updates without ever inspecting individual contributions, often using Secure Multi-Party Computation (SMPC).

Data across silos is almost always non-Independently and Identically Distributed (non-IID) , presenting a core statistical challenge.

• Label Distribution Skew: Hospital A may specialize in oncology while Hospital B focuses on cardiology, leading to vastly different diagnostic label distributions.
• Feature Distribution Skew: Genomic sequencing pipelines, equipment calibration, and population demographics vary by institution, causing systematic biases in the input features.
• Mitigation Strategies: Techniques like FedProx (proximal regularization) and personalized federated learning are employed to stabilize convergence and prevent local models from diverging too far from the global consensus.

Cross-silo FL integrates a layered stack of Privacy-Enhancing Technologies (PETs) to provide defense-in-depth beyond the basic data locality guarantee.

• Differential Privacy: Calibrated noise is added to model updates before transmission, providing a mathematical guarantee against membership inference attacks.
• Homomorphic Encryption: Allows the aggregation server to perform mathematical operations directly on encrypted model updates, ensuring the server never sees raw gradients.
• Trusted Execution Environments (TEEs) : Sensitive aggregation logic can run inside hardware-isolated enclaves, protecting computation from the host operating system and cloud provider.

This topology is specifically architected to satisfy data residency and sovereignty mandates in highly regulated sectors.

• GDPR Alignment: Raw personal data never leaves the jurisdiction of the data controller, addressing cross-border transfer restrictions.
• HIPAA Compliance: Protected Health Information (PHI) remains within the covered entity's secure enclave; only de-identified mathematical updates are shared.
• Auditability: Consortium agreements define clear audit trails for model lineage, data provenance, and participant contributions, satisfying clinical governance requirements.

Individual institutions lack sufficient cases of rare Mendelian diseases to train statistically powerful models. Cross-silo FL allows a consortium of specialized children's hospitals to jointly train a variant pathogenicity classifier on a combined cohort that is orders of magnitude larger than any single site's dataset.

• Each hospital retains local control over its whole-exome sequencing data
• The aggregated model learns allele frequency patterns and causal variant signatures across diverse populations
• Enables discovery of novel gene-disease associations that would be statistically invisible in isolated silos

Traditional GWAS requires pooling individual-level genotype and phenotype data, creating privacy and regulatory barriers. A cross-silo Federated GWAS architecture computes association statistics locally at each biobank and shares only summary-level allele effect sizes and standard errors.

• Preserves compliance with GDPR and institutional IRB protocols
• Enables meta-analysis across multi-ethnic cohorts to improve polygenic risk score portability
• Supports dynamic participation where new biobanks can join the consortium without re-negotiating data-sharing agreements

Predicting rare drug-gene interactions requires exposure data across diverse genetic backgrounds. A cross-silo FL network of hospital systems collaboratively trains a model to predict severe adverse drug reactions from HLA typing and pharmacogene variants.

• Each hospital contributes labeled cases of drug-induced liver injury or Stevens-Johnson syndrome
• The global model learns to identify high-risk pharmacogenomic haplotypes without exposing individual prescriptions
• Enables preemptive genotyping recommendations before administering high-risk medications like carbamazepine or allopurinol

Tumor somatic mutation profiling for rare cancers suffers from small sample sizes at any single cancer center. Cross-silo FL enables National Cancer Institute-designated centers to jointly train a driver mutation classifier on combined whole-genome sequencing data.

• Each center retains its protected health information (PHI) behind its own firewall
• The federated model learns mutational signature patterns across diverse tumor types
• Supports federated survival analysis to correlate genomic features with treatment outcomes without sharing patient timelines

Polygenic risk scores (PRS) trained on homogeneous populations exhibit poor portability across ancestries. A cross-silo FL consortium of biobanks representing diverse genetic ancestries jointly calibrates a trans-ancestry PRS model.

• Each biobank computes local PRS weights on its population-specific cohort
• The federated aggregation produces a model that generalizes across European, African, East Asian, and admixed populations
• Addresses the critical health equity gap in genomic risk prediction without requiring centralized mega-biobanks

Pre-trained genomic foundation models like DNABERT or Enformer require fine-tuning on domain-specific regulatory data. A cross-silo consortium of research hospitals collaboratively fine-tunes a Federated Enformer model to predict chromatin accessibility and gene expression from DNA sequence.

• Each institution contributes its ATAC-seq and RNA-seq tracks as local training labels
• The federated fine-tuning preserves the pre-trained weights while adapting to tissue-specific epigenomic patterns
• Enables joint modeling of promoter-enhancer interactions across cell types without centralizing raw sequencing reads