Skip to main content

Data Owner: Reviewing FL Job Proposals

You act as the final security gate for your data. In a Federated Learning (FL) workflow, no code runs on your private assets without your explicit manual review and authorization. This guide explains how to use the RDS-Dashboard to audit and approve incoming job requests.

Overview​

When a Data Scientist submits an FL project, it appears as a Job Proposal on your datasite. You have the power to inspect the code, verify the resource requirements, and decide whether to allow the computation to proceed.

1. Finding Pending Proposals​

All incoming requests are queued in your dashboard:

  1. Open your RDS-Dashboard (typically at http://localhost:8000).
  2. Navigate to the Jobs tab in the sidebar.
  3. Look for jobs with the status PENDING_REVIEW.

2. Auditing the Project​

Before approving, you must perform a "Human-in-the-Loop" audit of the request.

Identity & Context​

  • Submitter: Verify the email address of the Data Scientist.
  • Target Dataset: Ensure the job is requesting access to the correct dataset (e.g., your Diabetes Training Data).

Code Inspection​

Click the "View Code" button to open the secure file explorer. Focus your audit on these key files:

  • client_app.py: This script interacts with your private data. Ensure it performs only model training (e.g., calculating gradients) and does not contain print() statements or code that exports raw data rows.
  • pyproject.toml: Review the project dependencies. It should only include trusted libraries like syft-flwr, torch, or scikit-learn.

3. Approving and Running​

Once you are confident in the code's safety:

  1. Authorize: Click the Approve button. This signals to the SyftBox network that you trust this specific version of the code.
  2. Execute: Click Run Job.
  3. Local Isolation: SyftBox will spin up a secure, isolated execution environment (sandbox) to run the training.

4. Monitoring Progress​

Switch to the Monitoring or Logs tab to see the live output.

  • Round Updates: You will see the local training rounds starting and completing.
  • Veto Power: If you see any unexpected behavior in the logs, you can click Terminate at any time to immediately kill the execution environment and delete the temporary sandbox.

5. Security Guarantees​

  • Compute-to-Data: The code comes to your data; your data never leaves your machine.
  • Read-Only Access: The execution environment has read-only access to your private data folder. It cannot modify or delete your original files.

Next Step: Proceed to the Running FL Clients guide to understand how the client-side training process works once you've approved a job proposal.