What It Is
Each participant trains the model locally and sends only model updates, such as weights or gradients, to a coordinator. The coordinator combines them into an improved global model and sends it back for the next round. The result is a model that learns from many locations while the underlying data stays where it was created.
Key Points
- Privacy: raw data never leaves its source.
- Efficiency: far less bandwidth than uploading entire datasets.
- Compliance: helps meet data residency and confidentiality needs.
- Challenges: uneven data between sites and unreliable device connections.
Why It Matters
Updates can still leak information, so techniques such as secure aggregation and differential privacy matter. It is a strong fit for industrial and regulated settings, where moving raw data is costly, slow, or not permitted.
How ClearLeaff Applies It
We design federated learning architectures with PyTorch and ONNX for edge environments. Multi-site industrial clients improve shared models while their data stays on premises.