AI can now invent new forms of encryption

One of the primary requirements in any online communication is keeping the transmission secure: that is, only the person to whom the information is sent to should be able to read it, and nobody else. To accomplish this, digital data is encrypted using one of several complex techniques. To further this research, Google recently conducted an experiment where they set up an interesting three-way challenge that involved artificial intelligence.

The experiment was conceptualized by the Google Brain team based out of their Mountain View headquarters (which is different from their DeepMind team, known for creating the AlphaGo AI supercomputer that famously defeated the world champion Go player.) The setup was simple: they began by creating three simple neural networks--AI implementations that can ‘learn’ based on experience. Two of these, nicknamed Alice and Bob, were tasked with sending a secure message where Alice was the sender and Bob the recipient. The third, Eve, had to attempt to ‘eavesdrop’ and decode the message by any means possible--basically, a hacking experiment that pitted AI against AI.

The communicators Alice and Bob had the advantage of having the shared key that would be used to both encrypt and decrypt the message--a common technique used in the real world, commonly known as symmetric cryptography or shared secret encryption. Because a third party--Eve, in this case--does not have access to this key, it would not be generally possible to decrypt the message.

Being neural networks, each of these elements were left to ‘find their own way’--nothing was defined beforehand. The first two, Bob and Alice, were left to choose the best possible encryption method before sending the message. The endpoint to this ‘game’ would be when Eve the ‘hacker’ was able to guess the original text to within an acceptable degree.

What then transpired was a cat and mouse chase involving key generation, where Alice would encrypt the original plaintext message that Bob would be successful at decrypting, when during the time Eve would try to intercept it and guess the decryption key.

While the results were too varied to be conclusive as to which AI won and lost, it became clear that the neural networks were capable of inventing altogether new types of algorithms for encrypting communications on their own. Such experiments could aid future AI systems to make better analyses of large amounts of data and communications traffic.