Getting Started¶

This guide assumes that you have followed the installation instructions in README.

Contents¶

Introduction to Tensorflow Encrypted API
Making private predictions
- Locally
- With GCP
Making private trainings

Introduction to TensorFlow Encrypted API¶

tf-encrypted has a simple api to make it easy for data scientists to make private predictions and trainings.

To define a machine learning model, tf-encrypted and TensorFlow follow a very similar API.

Tensorflow

Tensorflow Encrypted

layer0 = prot.matmul(x, w0) + b0
layer1 = prot.sigmoid(layer0)
layer2 = prot.matmul(layer, w1) + b1
prediction = layer2

layer0 = prot.dot(x, w0) + b0
layer1 = prot.sigmoid(layer0)
layer2 = prot.dot(layer, w1) + b1
prediction = layer2

To make private predictions and trainings, we have to define a ModelTrainer and a PredictionClient. The ModelTrainer is responsible for training the model, and then provides the encrypted weights. The PredictionClient will provide the private input that will be used to make a private prediction.

model_trainer = ModelTrainer(config.get_player('model-trainer'))
prediction_client = PredictionClient(config.get_player('prediction-client'))

We indicate tf-encrypted to perform secure computation with the Pond protocol as follows:

with tfe.protocol.Pond(server0, server1, crypto_producer) as prot:

Finally we can make a prediction as follows:

with tfe.protocol.Pond(server0, server1, crypto_producer) as prot:

  # get model parameters as private tensors from model owner
  w0, b0, w1, b1 = prot.define_private_input(model_trainer, masked=True)

  # get prediction input from client
  x = prot.define_private_input(prediction_client, masked=True)

  # compute prediction
  layer0 = prot.dot(x, w0) + b0
  layer1 = prot.sigmoid(layer0)
  layer2 = prot.dot(layer, w1) + b1
  prediction = layer2

  # send prediction output back to client
  prediction_op = prot.define_output(prediction, prediction_client)

Two servers and a crypto producer are doing the actual computation on encrypted data, with only the client being able to decrypt the final result.

Making Private Predictions¶

a. Locally¶

With Tensorflow Encrypted, you can very easily make private predictions with a pre-trained model saved as a protobuf file.

In the examples/ folder run the following line of code in your terminal to train a convolution network on MNIST dataset, then save the model as a protobuf file :

python3 mnist_deep_cnn.py

If you prefer to skip this step, we have saved the trained model mnist_model.pb in examples/test_data/.

Make private prediction on an MNIST input by running the following code snippet:

../bin/run test_data/mnist_model.pb test_data/mnist_input.npy

You have just made a prediction without revealing anything about the input!

b. GCP¶

NOTE It would be great to have an example as above with GCP where we import a protobuf file and run a prediction.

You can make private predictions with GCP as well. You can find a great example here.

Making Private Trainings¶

NOTE We currently don’t have an example. We are planning to train/create a simple logistic regression as a demo..