Train Syft Keras model with DP-SGD from TF Privacy

[yanndupis]

Description

We would like to be able to train a model with dp-sgd from TF Privacy. However these custom optimizers (e.g. DPGradientDescentGaussianOptimizer) are not supported out of the box in tf.keras.optimizers so they can't be easily exported to SavedModel.

Objectives/Key Results

• Figure out how to support custom optimizers like DPGradientDescentGaussianOptimizer
• Create a demo with this TF Privacy tutorial