Federated Learning on myDRE with FlowerAI

Federated Learning on myDRE with FlowerAI

Setting Up a Federated Learning Environment on myDRE with Flower AI (github flower)

The system enables institutions to collectively train AI models on their local data without sharing sensitive information, enabling secure collaboration in medical research.

Workflow

  1. Each institution trains its local model on a supernode.
  2. Supernodes push model updates to the Aggregation server
  3. Supernodes pull updated global model parameters
  4. Process repeats to improve model performance
  5. FL training ends when the preperformance reached to a convergence.

Advantages

  1. Maintains data privacy and security
  2. Compliance with regulations
  3. Works within existing security constraints of healthcare institutions
  4. Allows collaboration without sharing raw data
  5. Enhanced model performance
  6. Reduced data transfer costs
  7. High scalability and robustness

Setup 

Deploy a Virtual Machine (VM) in your myDRE Workspace

Software

  1. Conda for package management (Miniconda is recommended)
  2. FlowerAI framework for federated learning

Configure your Network

  1. Turn off the proxy
  2. Opened specific domains and IPs with ports for Conda setup:
    1. Use ping cmd to get their IP (e.g. ping repo.anaconda.com)
      1. repo.anaconda.com
      2. docs.conda.io
      3. conda.anaconda.org
    2. Additionally, add the followings to the Domain-Allowlist:
      1. azure.archive.ubuntu.com
      2. packages.microsoft.com
  1. Open IP address and ports for Superlink communication, use ifconfig to find the IP address of the machine you host Superlink and the port is 9092.
  2. Only outbound traffic allowed (no incoming connections).

Architecture

  1. Supernodes: Client VMs at each institution (e.g., Inst. A, Inst. B, University Hospital).
  2. Aggregation Server: Aggregates the global model. Hosted on a secure server at University Hospital.
  3. Superlink server: Runs the superlink application that manages the FL clients and the security.
  4. Communication: Push-pull model initiated by nodes, replacing traditional bidirectional communication.

Security Measures

SSL certificates for client and server authentication
Encrypted communication between nodes and global model
Future plans for AAA framework (Authentication, Authorization, Auditing)

To know more about Flower,  join the Flower Slack community using the URL: https://flower.ai/join-slack/, then search for the channel.

    • Related Articles

    • myDRE-Learning

      The myDRE-Learning platform expands your knowledge on a wide range of topics, from the basics of myDRE to Virtual Machines, Workspaces, Rroles, Costs and much more. Accessibility is at the heart of our platform. You can learn at your own pace and on ...

    • Federated Learning on myDRE with VANTAGE6

      VANTAGE6 An open source privacy preserving Federated Learning Infrastructure for Secure Insight Exchange. ARCHITECTURE VANTAGE6 uses a client-server model, which is shown in the figure below. In this scenario, the researcher can pose a question and ...

    • Seminar FlowerAI x myDRE

      16th of September 2024, at 15:00 UTC (17:00 CET, 16:00 LON, 11:00 NY) Federated Learning and Analytics with Flower through myDRE workspaces Flower Labs and anDREa will be hosting an online seminar to demonstrate how to use the Flower framework for ...

    • Why is my support.mydre.org account different from my @mydre.org account?

      We get this question regularly now and here is why: Your @mydre.org account may look like an e-mail address, but it is not. It has no e-mail functionality and it is strictly your username on the mydre.org platform. For support.mydre.org however, you ...

    • myDRE - why

      Why myDRE was created in <1.5 minutes For more information From organization perspective anDREa FAQ From user perspective myDRE Or contact a colleague: Pascalle Broer