Google's Neural Networks Invent Their Own Encryption

Google Logo New AH 2

The tech industry is moving at such a rapid pace that it’s always hard to guess what the future truly holds, but one thing can be reasonably presumed – whatever it is, it will have something to do with artificial intelligence. As no AI is truly intelligent without the ability to learn and evolve, scientists and engineers have recently resorted to developing deep learning. Without going into too many technical details, deep learning is an advanced form of machine learning based on a computing model which has been designed to function as a normal brain. The basic unit of this model is appropriately called a neuron, which is why the model as a whole is referred to as a neural network. Each neuron computes a given set of inputs, and when we start connecting them together, we get a neural network.

Now, neural networks aren’t artificial intelligence per se, but they’re a key stepping stone towards creating a “real” AI which isn’t designed to perform a particular function but can instead learn and adapt just like a human would. While we probably still have a long way to go until we reach that goal, some important advancements are already being made on an almost daily basis. We’re currently still at a stage in which the most powerful neural networks are being programmed in a rather narrow manner so that they can focus on learning to perform a single task as efficiently as possible. Cue Google Brain deep learning research project which recently managed to teach neural networks to create their own cryptographic algorithms and use them to encrypt their internal communications.

More specifically, Google’s researchers programmed two neural networks and called them Bob and Alice. The duo was then tasked with creating a secret cryptographic key which it would use to encrypt their communications channel. Bob and Alice proved to be up to the job because they managed to come up with a surprisingly advanced form of encryption. The programmers then threw in another neural network called Eve into the mix. Eve’s task was decrypting the initial encryption key developed by Bob and Alice. Interestingly enough, while Eve failed to do that, it did show signs of improving its decryption algorithm. However, Bob and Alice weren’t just sitting idly waiting for their key to be cracked; they were actively improving it faster than Eve was making progress.


Based on these results, the Google Brain team concluded that much like humans, neural networks are better at encrypting than decrypting stuff. That’s the gist of it, but if you’re interested in more technical details, you can read the full research paper titled “Learning to protect communications with adversarial neural cryptography” by following the source link below.