Cross-Silo Federated Learning

Cross-Silo Federated Learning is a federated learning setting where a small number of organizations (2-100) collaborate to train a model — each organization (silo) has a reliable compute infrastructure, large local datasets, and participates in every training round.

Cross-Silo Characteristics

• Few Participants: Typically 2-100 organizations (hospitals, fabs, banks).
• Reliable: All participants are always available — synchronous training is feasible.
• Large Local Data: Each silo has substantial local datasets (unlike cross-device FL).
• Governance: Formal agreements, contracts, and compliance requirements between participants.

Why It Matters

• Industry Collaboration: Multiple semiconductor fabs can jointly train defect classifiers without sharing proprietary data.
• Regulatory: Each organization keeps data within its regulatory jurisdiction (GDPR, export controls).
• High Value: Each silo contributes unique, high-value data — collaboration yields significantly better models.

Cross-Silo FL is organizational collaboration — a few large organizations jointly learning from their combined knowledge without sharing raw data.