Mnist with Differential Privacy

[Fabien-GELUS]

This example is a Substra implementation of on the Classification_Privacy tutorial from Tensorflow_Privacy.
The structure of this example is inspired from my previous MNIST example, I simply tweaked the model and the train method.

This tutorial uses tf.keras to train a CNN to recognize handwritten digits with the DP-SGD optimizer provided by the TensorFlow Privacy library. TensorFlow Privacy provides code that wraps an existing TensorFlow optimizer to create a variant that implements DP-SGD.
The algorithm also measure the differential privacy guarantee after training the model: You will see in the console the value Epsilon (ϵ) - This is the privacy budget.

[RomainGoussault]

Nice work @Fabien-GELUS 👍 !

[natct10]

Awesome! We tested it with a 0.8038 perf!

[bowni]

Hey there - is there anything preventing from merging? @natct10 @RomainGoussault
I'd say let's move forward, it is good for the morale 😃

[bowni]

Really nice job @Fabien-GELUS 💪 It feels really great so see an example, even very simple, of remotely training a model via Substra with using differential privacy to enhance the privacy-preserving promise!
(cc @RomainBey @RomainGoussault @natct10 @mattthieu @camillemarini @ClementMayer)