Federated Learning Protocols
What You'll Learn
You'll understand how federated learning systems coordinate client selection, aggregation, and communication to train models collaboratively without centralizing data.
What Are Federated Learning Protocols?
Federated learning protocols define how the distributed training process actually happens across many devices or organizations. While the architecture (from the previous lesson) describes what components exist, the protocol specifies the rules for interaction: which clients participate in each round, how their updates combine, and how communication flows between clients and the central server.
Think of it like a democratic voting system: the protocol determines who votes (client selection), how votes are counted (aggregation), and how many voting rounds occur before reaching a decision (communication rounds).
Core Protocol Components
Client Selection
The server doesn't use all clients in every training round—that would be impractical with millions of devices. Instead, it samples a subset (e.g., 100 out of 10,000 devices) each round. Selection can be:
- Random: Fair but ignores device capabilities
- Weighted: Prioritizes clients with more data or better connectivity
- Availability-based: Only selects currently active, charged devices
Aggregation Strategies
Once clients send model updates, the server combines them:
- FedAvg (Federated Averaging): Weighted average based on each client's dataset size
- FedProx: Adds regularization to handle heterogeneous client data
- Secure aggregation: Uses cryptography so the server never sees individual updates
Communication Rounds
Training happens in synchronized rounds:
- Server broadcasts current model
- Selected clients train locally
- Clients send updates back
- Server aggregates and updates global model
- Repeat until convergence
Each round balances progress against communication costs.
Key Takeaway: Federated learning protocols orchestrate client selection, aggregation methods, and iterative communication rounds to enable privacy-preserving distributed training at scale.