Skip to main content

Data Owner: Privacy & Security

Once a Federated Learning (FL) job proposal is approved, the next step is to run the FL client on your machine. This guide walks you through executing client-side training and explains how your data remains protected throughout the process.

1. Core Privacy Guarantees​

  • Compute-to-Data: The most fundamental protection. Your raw data never leaves your device; the code is brought to the data instead.
  • Zero Raw Data Access: The Data Scientist cannot see your private file path or read the raw contents. They only interact with the Mock Data you provided.
  • End-to-End Encryption: All model updates are encrypted while in transit across the SyftBox network.

2. Security Controls​

  • Human-in-the-Loop: No code can execute without your manual review of the job proposal.
  • Read-Only Environment: The FL client is granted "Read-Only" access to your specific dataset folder, preventing it from modifying or deleting your original files.
  • Access Revocation: You can stop a running job at any time or delete a dataset to instantly revoke all network access to that metadata.

3. Advanced Protections (PETs)​

syft-flwr is designed to support additional Privacy-Enhancing Technologies:

  • Differential Privacy: Noise can be added to model updates to prevent "reverse-engineering" of the underlying data from the weights.
  • Secure Aggregation: Protocols that ensure the aggregator can only see the sum of updates from all participants, but never an individual client's update.

Congratulations! You have completed the full Data Owner guides track for syft-flwr.