simple-average

Simple Average

As a simple starting point, this example illustrates how TF Encrypted can be used to securely compute the average of inputs from a set of input providers.

The set of players consists of five inputters, a result receiver, and the three compute servers needed by the default Pond protocol. The flow between them is as illustrated below, where encrypted inputs are sent to the compute servers calculating an encrypted average that is finally sent to the result receiver and decrypted:

Computation

For this simple example the inputs are simply random tensors of a pre-specified shape, computed locally by each inputter using the following function:

def provide_input() -> tf.Tensor:
    return tf.random_normal(shape=(10,))

These values are then brought into the secure computation using tfe.define_private_input:

inputs = [
    tfe.define_private_input('inputter-0', provide_input),
    tfe.define_private_input('inputter-1', provide_input),
    tfe.define_private_input('inputter-2', provide_input),
    tfe.define_private_input('inputter-3', provide_input),
    tfe.define_private_input('inputter-4', provide_input)
]

Here, all values in inputs are encrypted tensors that we next tell the compute servers to calculate the average of using:

result = tfe.add_n(inputs) / len(inputs)

and resulting in another encrypted tensor result.

Similarly to what we did for the inputs we define a function to be run locally on the result receiver after decryption has taken place:

def receive_output(average: tf.Tensor) -> tf.Operation:
    return tf.print("Average:", average)

result_op = tfe.define_output('result-receiver', result, receive_output)

Finally, to run the computation we follow the normal TensorFlow pattern of creating and using a session:

with tfe.Session() as sess:
    sess.run(result_op)

Note that we are using tfe.Session here instead of tf.Session.

Running

The example may be run using either an insecure local configuration, using different threads for the players:

python examples/simple-average/run.py

or remotely using different network hosts for the players as defined in a config.json configuration file:

python examples/simple-average/run.py config.json

See the documentation for more details.