Google Researchers Create AI That Builds Its Own Encryption

Alice and Bob have figured out a way to have a conversation without Eve being able to overhear, no matter how hard she tries.

But Alice and Bob aren't humans who can whisper. They're artificial intelligence algorithms created by Google engineers, and their ability to create an encryption protocol that Eve (also an AI algorithm) can't hack is being hailed as an important advance in machine learning and cryptography.

The researchers, Martin Abadi and David G. Andersen, explained in a paper published this week that their experiment is intended to find out if neural networks—the building blocks of AI—can learn to communicate secretly.

"Neural networks are generally not meant to be great at cryptography," the engineers wrote. So they created multiple versions of Alice and Bob to see which ones were best at keeping their transmissions hidden from Eve. That would be easy if Alice and Bob were given a human-created encryption protocol to run, but what about one that they had to create themselves?

As the Abadi and Anderson wrote, "instead of training each of Alice and Bob separately to implement some known cryptosystem, we train Alice and Bob jointly to communicate successfully and to defeat Eve without a pre-specified notion of what cryptosystem they may discover for this purpose."

Recommended by Our Editors

Crypto Wars: Why the Fight to Encrypt Rages On

After multiple experiments powered by a computer with a single GPU (Abadi and Anderson explain that the type of hardware used is not important) Eve was not able to understand anything about Bob and Alice's transmissions.

The engineers acknowledge that their experiment simply proves that neural networks can create their own encryption, not that they're necessarily good at it. "While it seems improbable that neural networks would become great at cryptanalysis, they may be quite effective in making sense of metadata and in traffic analysis," they wrote.