Myujth

I'm attempting to deploy trained neural networks to multiple platforms by using TVM to compile the models to WebGL.

We're hoping to sandbox the execution of these models within our app since

1. much of the data they operate on is sensitive and we'd like to be able to assure users (and ourselves) that the mysterious JavaScript we're downloading can't exfiltrate their data¹




2. we don't want bugs in this code to interfere with the overall application

Ideally, we'd like an API like this in JS, C, C++, or Rust:

predictions, err = runInSandbox('myModel.js', canUseNetwork=False,
 canUseFileSystem=False, input=mySensitiveData)

Some options I've looked into include:

• Docker. This solves sandboxing, but can't run on Android or iOS. It would also be nice to avoid installing Docker as part of installing our application.


• 
  Puppeteer. This also seems to not run on Android or iOS, and is also undesirable to bundle with an application.


• 
  Running the WebGL code in an iframe. This doesn't prevent network access, though maybe there's some clever trick to do so?


• Running an embedded JavaScript or Wasm VM. This would be ideal, but I haven't found one that will run WebGL.

Any ideas for how to do this across platforms would be greatly appreciated. Approaches for particular platforms---especially Android and iOS---are also welcome, since using different approaches for different platforms is okay if there's no alternative.

1 Users are allowed to audit our native client-side source code, but it is not realistic for them to audit every change to our models / data processing pipeline.