High-level API

Moving forward the idea is to have two APIs:

• a high level API as close as possible to keras:

  • allowing user to more clearly express what they want to achieve with secure computation (along the lines of how TF2 wraps distribution up into a DistributionStrategy object)

  • making it easier for us to guide users in how to get good performance out of TFE, in some sense making it harder for them to shoot themselves in the foot

• a low level API similar to what exists currently:

  • gives full power but also the ability to shoot yourself in the foot by eg naively using expensive operations everywhere

See also Standardizing on Keras: Guidance on High-level APIs in TensorFlow 2.0

[mortendahl]

Note: let’s make sure that tfe.keras matches tf.keras as much as possible, ideally such that one can simply replace tf with tfe.

[mortendahl]

One small note: it seems safe to follow design patterns that users will be familiar with from TF already, even if there might potentially be better ways of doing things. One concrete example is how to make this Keras implementation agnostic to the underlying computation protocol, which can be done in at least two ways:

• define abstract Keras elements that can be created without knowing the computation protocol and later instantiated in an explicit call; one issue here is the Keras functional API which might require us to replicate a lot of dummy elements and patch up;

• an alternative is to take inspiration from TF distribution strategies and allow the use of contexts; to still be agnostic to the computation protocol the user can wrap model creation in a function

One potential issue with the second approach is how to know the entire model before committing to the protocol, which might be useful as part of picking techniques or (crypto) parameters. One solution that fits with the function-pointer approach is to have a special context (with dummy tensors etc) in which the model is constructed and from which a protocol suggestion can be derived.

cc @yanndupis @jvmancuso