globus-labs / FLoX-prototype

Python library for serverless Federated Learning experiments.
Apache License 2.0
14 stars 1 forks source link

Enhance privacy preservation #17

Open nikita-kotsehub opened 2 years ago

nikita-kotsehub commented 2 years ago

Currently, endpoints send back the model weights, which isn't secure. Here some ideas we can implement:

  1. Send back just the updates
  2. Encode the position of the model updates and send back only the weights that changed (assuming only a few weights change)
  3. Add a mask to each individual model update such that upon aggregate the masks are negated
  4. Add encryption to the weights before they enter the funcX pipeline