Cross-Silo Federated Learning

The defining characteristic is a small cohort of known, reliable entities, typically 2 to 100 organizations. Each participant, or silo, is a legally distinct institution with a strong identity. This contrasts sharply with cross-device FL, which orchestrates millions of anonymous mobile devices. The established trust relationship allows for the use of semi-honest security models and simplifies client selection, as every silo is expected to be available and non-malicious in each training round.

Each silo is a stateful node with significant computational power, often including GPU clusters, and large, curated local datasets. Unlike stateless mobile phones, silos maintain their local dataset across training rounds. This enables complex local training procedures, such as executing multiple local epochs with large batch sizes. The resource abundance allows for the use of computationally intensive privacy mechanisms like homomorphic encryption or trusted execution environments (TEEs) without the stringent energy constraints of edge devices.

Communication occurs over dedicated, high-bandwidth, and secure channels, not the public internet. The primary threat model is an honest-but-curious aggregator server, not malicious end-users. This allows for the efficient use of secure aggregation protocols where the server computes the sum of encrypted model updates without inspecting individual contributions. The reliable connectivity eliminates the need for complex straggler mitigation or asynchronous protocols, allowing for simple, synchronous Federated Averaging (FedAvg) rounds.

Data is partitioned by sample space, not feature space. Each silo holds data for a different set of users or events, creating extreme statistical heterogeneity. A model trained on aggregated data from different hospitals will encounter vastly different patient demographics and equipment biases. This non-IID challenge is the central research problem, driving the need for advanced algorithms like FedProx or personalized federated learning to prevent local models from diverging from a useful global optimum.

The primary business motivator is compliance with data governance regulations like HIPAA, GDPR, or financial secrecy laws. Cross-silo FL provides a technical mechanism to legally train on sensitive data that cannot be moved. The architecture serves as a privacy-preserving data alliance, where the final global model's performance is the shared asset. This requires a robust governance framework, often involving a federated blockchain ledger for immutable audit trails and to manage the intellectual property rights of the jointly trained model.

The output is a single, high-performance global model that generalizes across all silos' data distributions, representing shared intellectual property. The goal is not personalization for a single silo, but a robust model that captures the union of all knowledge. This model is often validated on a held-out, centralized test set. The collaborative nature requires upfront agreement on model architecture and a federated data valuation process, often using game-theoretic tools like the Shapley value, to equitably compensate participants based on their data's marginal contribution to the final model's accuracy.

A cryptographic protocol that allows the central coordinating server to compute the weighted sum of encrypted model updates from multiple institutional silos without ever inspecting an individual participant's plaintext contribution. In a cross-silo setting, this is often implemented using secret sharing or homomorphic encryption to ensure that even the aggregator cannot reconstruct a single hospital's or bank's proprietary gradient vector. This is a critical trust enforcement mechanism when silos are mutually distrustful but must collaborate.

A data distribution characteristic where local datasets across silos are statistically heterogeneous—they are not independently and identically distributed. In a cross-silo context, this manifests as label distribution skew (e.g., one hospital treats predominantly cardiac cases while another focuses on oncology) or feature distribution skew (different MRI machine vendors). This heterogeneity is the primary cause of client drift, where local optima diverge from the global optimum, and requires specialized aggregation algorithms beyond simple FedAvg.

The fundamental optimization challenge in cross-silo FL arising from divergent local data distributions. Unlike cross-device FL where stragglers and communication are the bottleneck, cross-silo systems grapple with objective inconsistency: each silo's local loss landscape points toward a different minimum. Mitigation strategies include:

• FedProx: Adds a proximal term to local objectives to constrain divergence from the global model.
• SCAFFOLD: Uses control variates to correct for client drift during local updates.
• MOON: Applies contrastive learning at the model representation level to reduce feature space discrepancy.

The property of a distributed learning system that guarantees convergence to a correct global model even when a fraction of participating silos are faulty or actively malicious. In a cross-silo consortium of banks or defense contractors, a single compromised institution could upload a model poisoning update designed to sabotage the joint model or embed a backdoor. Byzantine-resilient aggregation rules—such as Krum, Trimmed Mean, or Median-based aggregators—filter out statistical outliers before weight averaging, ensuring robustness against arbitrary adversarial behavior.

A multi-tier learning architecture that introduces intermediate edge aggregation nodes between institutional silos and the central cloud orchestrator. This is particularly relevant for cross-silo deployments spanning multiple jurisdictions: a regional aggregator in the EU can perform local model fusion on data from hospitals within GDPR boundaries, and only the regionally-aggregated, privacy-compliant update is transmitted to the global server. This reduces wide-area network latency and provides a natural boundary for data sovereignty compliance.

A privacy-preserving distributed architecture where a deep neural network is physically partitioned between a client silo and the aggregation server. The client retains the initial layers and only transmits the smashed data (intermediate activations) to the server, which completes the forward and backward passes. Unlike federated learning, the raw data and full model architecture are never shared. This is advantageous in cross-silo settings where one party owns the labels and another owns the features—a common scenario in fintech risk modeling between a bank and a credit bureau.