Google has a trio of neural networks called Alice, Bob, and Eve. These neural networks were trained to optimize and perfect their parts of a communication process. Alice was to send a message to Bob and Bob was to decode that message. Eve was only there to throw a wrench in the works and read the message Alice sent without permission. To keep the message secret Alice and Bob did something that is more than a bit disturbing if you have ever seen Terminator.
Alice and Bob ended up creating their own form of encryption using machine learning without having to be taught any specific cryptographic algorithms. Researchers Martin Abadi and David Andersen say that the encryption processes the computers created is very basic compared to those that humans have created.
Despite the simplicity of the encryption process, this is a big deal because the researchers say that neural networks “are generally not meant to be great at cryptography.” Bob and Alice started with an agreed set of numbers for their key. During the process Alice was able to develop her own encryption strategy and Bob was able to slowly learn to decipher the text. It apparently took 15,000 tries before Bob was able to turn the encrypted text that Alice sent into plain text.
During that time Eve, the eavesdropper, was only able to guess 8 out of the 16-bits in the secret message. The researchers say that this is the same percentage you could get from pure chance with a binary system. The catch is that the researchers on the project have no idea how the encryption method works so they can’t give any security guarantees. Essentially these two neural networks have learned to encrypt their own messages and their human creators have no idea how to decipher the text themselves. This is how you get Skynet.
SOURCE: New Scientist