EMLaaS: secure model serving

EMLaaS.md

EMLaaS: Secure model serving

Project 1 - Notebooks providing EMLaaS for image classification

This consists of: hosting a model (regular CNN??? I've tried resnet50 but not supported, I'll try other popular architectures) and being able to query it using a rest api.

• 1.1. host non-encrypted model

  • WIP demo: train model
  • WIP demo: host model
    • This demo was using jit.trace. I think we probably want to stick with plans for now. Working on it.

• 1.2. query encrypted model

  • Exactly like Syft Tutorial 11.
  • I have a MNIST model (MLP, 2 layers) that workers. One image takes around 30 seconds using local grid nodes. No hosting is implemented yet.
  • Next step is to use the model from the demo above.

• 1.3. host encrypted model

  • Grid gateway is responsible for model execution
    • We tell grid gateway which nodes to host the model shares (or it can automatically choose)
    • We ask grid gateway to run inference in our shares

• 1.4. implement fault tolerance

Project 2 - Notebooks providing EMLaaS for spam detection with RNNs

Same thing as project 1 but with RNNs instead of CNNs.

Project 3 - Web demo: skin cancer classification (don't think we have the time :sad:)

• Is it possible to support encrypt/decrypt using just JS? (I'll ask this in the weekly meeting)
  • Implement it
• [ ] Build web demo + interface

Challenges to support GPT-2

Implement 1.2. for GPT-2.

• Some operations may not be supported yet by Syft
• It could take minutes (hours??!!) to generate text (1-3 sentences)