Encryption has long been a battle of wits, but they needn’t be human wits. Google‘s AI program, called Google Brain, recently announced a project in which neural networks competed with each other to invent better encryption algorithms. The contest consisted of networks nicknamed Alice and Bob passing messages while another network, called Eve, was listening in. Eve was at a disadvantage: She could see only the encrypted text, not the key that was used to encrypt it.
Competition between human cryptographers is familiar to anyone who saw The Imitation Game, the film about AI pioneer Alan Turing. But as Turing and many others have shown, codes can be cracked, and Eve often succeeded, too. That prompted Alice and Bob to redouble their efforts, and here’s where it got really interesting: Instead of using established cryptographic algorithms, created by humans, they kept inventing their own as they ratcheted up efforts to defeat Eve. This was quite a breakthrough. “Neural networks are generally not meant to be great at cryptography,” reads Google’s research paper.