Adversarial Neural Cryptography in Theano

Last week I read Abadi and Andersen's recent paper [1], Learning to Protect Communications with Adversarial Neural Cryptography. I thought the idea seemed pretty cool and that it wouldn't be too tricky to implement, and would also serve as an ideal project to learn a bit more Theano. This post describes the paper, my implementation, and the results. The authors set up their experiment as follows. We have three neural networks, named Alice, Bob, and Eve.