Hierarchical Federated Learning is a multi-tier federated learning architecture that introduces intermediate aggregation layers — instead of all clients communicating directly with a central server, clients first aggregate within local groups (e.g., within a site), then group aggregates are sent to the global server.

Hierarchical Architecture

• Edge Level: Devices/sensors within a single machine or department aggregate locally.
• Site Level: Department-level models aggregate within a fab or facility.
• Global Level: Site-level models aggregate at the organization or cross-organization level.
• Aggregation: Each level can use different aggregation strategies (FedAvg, FedProx, robust aggregation).

Why It Matters

• Communication: Reduces long-distance communication — most aggregation happens locally.
• Scalability: Scales to thousands of clients by distributing the aggregation load.
• Natural Structure: Maps to organizational hierarchies (sensors → machines → fabs → enterprise).

Hierarchical FL is aggregation in tiers — mirroring organizational structure for scalable, communication-efficient federated learning.