Federated Learning Workshop AMLD2021

Preparations

We've tried to keep this workshop light on the setup/preparations side by using Google Colab for a majority of the materials and exercises that we will cover.

Therefore, the main requirement is access to a Google account which would be required to access the notebooks we prepared.

Please ensure you have access to a google account you can use, or create one in preparation for the session.

Don't worry about studying the notebooks below, these will be covered as integral part of the workshop and will remain available to you afterwards as well.

1. Interactive Examples: torch Federated Learning

1.1 Simulating Federated Learning using torch

This example introduces basic components that are part of the Federated Learning workflow. In the simulation we:

1. Train a baseline model on the CIFAR10 dataset.
2. Simulate a server & client and send a model from server to client to be trained.
3. Simulate a server & multiple clients and train a global model on multiple clients and perform FedAvg.

Please use the google colab notebook link below to access the relevant material:

PyTorch FL Simulation Notebook

2. Interactive Examples: Syft

2.1 Simple Syft Duet Intro

This example introduces basic primitives of a successful PySyft Duet Exchange:

• establishing a duet connection between data owner and client
• safely providing information about available datasets as a data owner
• issuing access requests as a client/data scientist
• approving/denying requests for data access as a data owner

Please use the google colab notebook links below to access the relevant materials:

• Data Owner: Data Owner Notebook

• Data Scientist: Data Scientist Notebook

2.2 Syft FL Example

This example showcases a basic federated learning exchange (with a data scientist modeling against data held by two separate data owners) using PySyft Duet.

Please use the google colab notebook links below to access the relevant materials:

• Data Owner 1: Data Owner 1 Notebook

• Data Owner 2: Data Owner 2 Notebook

• Data Scientist: Data Scientist Notebook

2.3 Opacus Example

This example discusses how to use opacus for differential privacy to ensure data privacy while training models collaboratively.

Please use the google colab notebook link below to access the relevant material:

• Opacus Example: Opacus Example Notebook