Federated Learning architecture

IBM Federated Learning has two main components, the aggregator and the remote training parties. 

Aggregator

The aggregator is a model fusion processor. The aggregator:

• runs as a platform service in regions Dallas, Frankfurt, London, or Tokyo.
• starts with a Federated Learning experiment. The admin manages the aggregator.

Party

A party is a user that provides model input to the Federated Learning experiment aggregator. The party can be:

• on any system that can run the Watson Machine Learning Python client and compatible with Watson Machine Learning frameworks. Note: the system does not have to be specifically Cloud Pak for Data. For a list of system requirements, see Set up your system.
• running on the system in any geographical location. You are recommended to locate each party in the same region where the data is to avoid data extraction out of different regions.

This illustration shows the architecture of IBM Federated Learning.

A Remote Training System is used to authenticate the party's identity to the aggregator during training.

User workflow

1. The data scientist:
   1. Identifies the data sources.
   2. Creates an initial "untrained" model.
   3. Creates a data handler file. They might intersect with a training party entity.
2. A party connects to the aggregator on their system, which can be remote.
3. An admin controls the Federated Learning experiment by:
   1. Configuring the experiment to accommodate remote parties.
   2. Starting the aggregator.

This illustration shows the actions that are associated with each role in the Federated Learning process.

Parent topic: Get started