The goal of this research is to create Data Vending Machines that are capable of interacting with confidential data, preventing the provider to ever see what is being worked on and at the same time giving the user the ability to provide the DVM with contextual information or sensitive data for processing.

Currently the only way to process confidential data is to either:

One is potentially viable but you need access to the hardware and models, the other is just nuts. While you can process a lot of things locally on consumer hardware your needs might be bigger or just occasional where you can’t build a cluster at home. The other issue is that you can’t rent an algorithm (model) for one run. Confidential computing is the third way. Both parties provide their part (model & data) to an encrypted blob called a virtual machine and the owner of the data gets the encrypted result. There are still some trust assumptions to solve.

Ideally for maximum privacy you’d push the entire VM to the provider but that can be a bit heavy on the overhead, hence the exploration of solving this with confidential containers and DVMs.

Use cases

Currnet high level thinking

1) Deploy a new confidential container for each job 2) Establish secure communication channel with the user

Potential issues


1) SEC-01 first demo - “something works”