Google’s AI made an encryption system and no-one understands it yet
Researchers from Google’s deep learning project are teaching computers to keep secrets.
In a paper published by Martín Abadi and David G Andersen of Google Brain, a project is described wherein three neural networks are able to develop a system of encryption – independent of humans.
The three neural networks, named Alice, Bob and Eve, were each given a specific role in the communications. Alice was trained to deliver a secret message to Bob, Bob’s job was to decode the message, and Eve had to eavesdrop on the conversation.
Alice was instructed to convert a plain-text message into what looked like gibberish, so that Eve (and anyone other reader except from Bob) would be unable to make sense of it. Bob and Alice were both given a pre-agreed cipher key, which could be used to decode the message. To begin with, however, they were both pretty shoddy at sending messages without Eve being able to understand.
As New Scientist reports, they got much better at it. After 15,000 attempts, Eve was only able to get eight bits of the messages correct. Seeing as the messages only consisted of 16 bits of either 1 or 0, this amounts to the neural network essentially making a random guess about the message’s content. Alice managed to send secret messages to Bob.
Due to the nature of machine learning, the researchers don’t yet know the precise details of how Alice’s encryption system functions. This means they can’t be sure of the security of the method, and so practical applications are somewhat limited. It’s also worth noting that the encryption methods devised by Google’s AI are incredibly basic compared to the human-created systems.
Still, it marks a thought-provoking movement for computing systems that, the researchers admit, “are generally not meant to be great at cryptography”. Abadi and Andersen write that future work could consider steganography, pseudorandom-number generation, or integrity checks, and that neural networks could use these techniques to make sense of metadata. For more information, you can have a look at the paper here.
“Computing with neural nets on this scale has only become possible in the last few years, so we really are at the beginning of what’s possible,” Joe Sturonas of encryption company PKWare told New Scientist.
Jacob Ginsberg, senior director of encryption company Echoworx, told us there may be advantages in using AI to automate aspects of encryption:
“When you consider that humans are consistently the weakest point in a security chain, there’s both financial and operational value in automating encryption between systems,” said Ginsberg. “It has the potential to dramatically increase security. It’ll be interesting to see how this technology develops over the next few years and what the adoption levels among businesses will be.”