Skip to main content

Flower Integration

syft-flwr integrates the Flower federated learning framework with the SyftBox protocol for privacy-preserving distributed training.

Architecture​

Flower Client​

The Flower client runs on each participant's machine:

  • Trains models on local private data
  • Sends model updates (not raw data) to server
  • Receives global model updates
  • Operates within SyftBox datasite

Flower Server​

The server coordinates federated training:

  • Aggregates model updates from clients
  • Broadcasts updated global model
  • Implements aggregation strategies
  • Runs as a SyftBox application

Communication Flow​

┌──────────────┐                    ┌──────────────┐
│ Client 1 │ │ Client 2 │
│ (Private │ │ (Private │
│ Data) │ │ Data) │
└──────┬───────┘ └──────┬───────┘
│ │
│ Model Updates │
│ (via SyftBox) │
│ │
└────────────┬──────────────────────┘
│
┌─────▼──────┐
│ Server │
│ (Aggregator)│
└────────────┘

Core Components​

Bootstrap Process​

Initializes federated learning setup:

  • Configures Flower server
  • Registers clients
  • Sets up communication channels

Aggregation Strategies​

Supported strategies for combining model updates:

  • FedAvg: Federated Averaging (default)
  • Custom strategies can be implemented

File Mounts​

Manages file synchronization for FL:

  • Model checkpoints
  • Training configurations
  • Results and metrics

Templates​

Pre-configured setups for common FL scenarios:

  • Classification tasks
  • Regression tasks
  • Custom workflows

Key Modules​

flower_client.py​

Client-side training logic and model updates.

flower_server.py​

Server-side aggregation and coordination.

strategy/fedavg.py​

FedAvg aggregation implementation.

run_simulation.py​

Local simulation for testing FL workflows.

grid.py​

Distributed grid setup for production deployments.

note

Detailed API documentation and usage examples coming soon

See Also​