DataEncryptionLocked

You read that right: artificially intelligent neural networks are now coming up with their own encryption techniques to send secret messages to one another. Two Google Brain researchers reported the successful construction of a human-independent, purely AI-created encryption that has already been used to keep other AIs – and their human interlocutors – from eavesdropping on the conversation.

Considering the frequency with which AIs and chatbots decide they’ll wipe out humanity or simply keep people in a “human zoo” you might find this terrifying and start scurrying for the nearest underground bunker.

The reality is a little more banal than global takeover and the subjugation of the human populace.

But before you start stockpiling canned goods, you can relax. While it’s true that the machines have managed to have a conversation about which we know nothing, the reality is a little more banal than global takeover and the subjugation of the human populace.

Here’s how it went down: the researchers took three neural networks, named Alice, Eve and Bob and tasked two of them (Alice and Bob) with passing messages the third (Eve) could not read. A simple numeric cipher was shared with Alice and Bob, but early attempts at secrecy were easily deciphered by Eve.

As the researchers describe:

We do not prescribe specific cryptographic algorithms to these neural networks; instead, we train end-to-end, adversarially. We demonstrate that the neural networks can learn how to perform forms of encryption and decryption, and also how to apply these operations selectively in order to meet confidentiality goals.

Indeed, by the time 15,000 messages had been shared between Alice and Bob, things had changed. Alice was able to successfully encrypt communications, Bob was able to successfully decrypt them and Eve was unable to do any better than guess at their contents. The messages were binary, either a 1 or a 0, and Eve only managed to get 50% correct, equivalent to flipping a coin.

Alice was able to successfully encrypt communications, Bob was able to successfully decrypt them and Eve was unable to do any better than guess at their contents.

The significance of the result comes down to the fact that Alice and Bob were able to develop their encryption method on the fly, simultaneously developing an encryption and decryption strategy that only they could make sense of even while Eve was intercepting their messages.

So while the messages were a far cry from Skynet, they do prove that AIs can successfully invent an encryption method that they have developed themselves and that even other AIs can’t decrypt.

Unfortunately, the encryption method the AIs developed cannot even be studied, as the machine learning simply spits out an answer rather than showing its working. This also means the commercial applications are relatively limited, as no security guarantees could be offered for the encryption algorithm. This means that we, as mere humans, are unlikely to benefit from AI-created encryption.

Do you use encryption in your daily life? Where do you see AI-created encryption going?