Google Brain has done something quite crazy again.
Using three roles to create an encryption and decryption machine learning system - Alice, Bob, Eve.
Alice and Bob need to communicate with each other, while Eve tries to steal data from them.
These three roles are neural networks.
Alice: The loss function is such that the more accurate Eve's guesses are compared to random guessing, the greater the loss.
Bob: The loss function results in a greater loss the further the guess is from the truth.
Eve: The loss function results in a smaller loss the more accurate the guess.
This forms a generative adversarial network, allowing Alice and Bob to communicate with each other, creating an encryption and decryption network, and preventing Eve from obtaining correct plaintext information.
This is an important example of mutual learning leading to mutual growth.
The paper in the link mentions experiments conducted using TensorFlow on a single GPU: https://techcrunch.com/2016/10/28/googles-ai-creates-its-own-inhuman-encryption/