Researchers from the Google Brain project have already taught AI to create, but now they have created something even more interesting: they implemented an artificial encryption algorithm created independently of humans. According to the report, Google employees, Martín Abadi and David G. Andersen, deliberately allowed three test subjects - the neural networks Alice, Bob, and Eve - to transmit messages to each other, encrypted using a method that was created by the networks themselves and which they improved during the experiment.
Each AI got its own task: Alice had to create a secret message that only Bob could read, and Eve had to find a way to intercept and decrypt this message. The experiment began with a simple text message that Alice turned into an unreadable set of characters that only Bob could figure out using his key. At first, Alice and Bob were bad at keeping their secret, but after 15 thousand attempts, Alice came up with a reliable encryption strategy, and Bob simultaneously learned to select the reverse cipher. The message was only 16 bits long, taking the value 0 or 1, so the fact that Eve was able to guess only half of the characters means that she was just "flipping a coin."
Of course, the impersonation of these three neural networks is a little confusing: due to the principles of machine learning, the researchers themselves do not know which encryption method Alice developed, so there may not be any practical application of the results.
Anyway, this is a very interesting study, but you shouldn't worry about the machines communicating behind our backs yet.
Ivan Biryukov
