A team of researchers from Google have shown that machines are learning their own way how to protect their messages through encryption. In this way, they are being able to send secret messages to other machines, while avoiding that anyone intercepting these messages can possibly understand them.

The research made by Martin Abadi and David Andersen shows that computing systems that are based on artificial neurons, known as neural networks, can use encryption techniques and learn it without being taught specific cryptographic algorithms. This step was described by the researchers as very interesting, mostly because these networks are not usually intended to be good at cryptography.

In the experiment which was conducted, there were three neural networks that were called Alice, Bob and Eve. Alice had to send a secret message to Bob, while Bob had to decode this message. At the same time, Eve's job was to try to intercept the message and once it did, try to understand it.

To make sure that Eve could not understand the message, Alice had to convert the original text message into a ciphered text. However, the text needed to be decipherable by Bob through a set of numbers called keys which were previously agreed between Bob and Alice.

Through the experiment, the neural networks started learning how to send secret messages. In fact, in the beginning the encryption was relatively poor, but then Alice started developing its own strategy while Bob worked out how to decrypt it.

The researchers explain that it took 15,000 times for Bob to be able to convert Alice's message in plain text, while at this time Eve only managed to guess 8 of the 16 bits which formed the message.