The machines are getting smarter folks, as they are now at a point where they can keep secrets from us. A team from Google Brain that is part of the company's deep learning project worked on artificial intelligence capable of protecting messages from prying eyes.
What we have here is human-independent AI performing an encryption task that makes it difficult for us to crack. Just imagine computers encrypting our messages and data by their lonesome, and we can't have access to it. That's what Google is working on, and the company has experienced some success thus far.
Google employees Martín Abadi and David G. Andersen are the men behind this exciting project. They created three test subjects - all neural networks - known only as Alice, Bob and Eve. The task is to have them pass notes to each other that are encrypted to a point where humans are unable to gain access.
According to a report from New Scientist, Abadi and Andersen assigned tasks to each AI. Alice needed to send a message that only Bob can decipher and read, while Eve is tasked with how to decode the message. Not an easy task, even for an AI, because at first both Alice and Bob were terrible at hiding their secrets, but things got better over time.
"We ask whether neural networks can learn to use secret keys to protect information from other neural networks. Specifically, we focus on ensuring confidentiality properties in a multiagent system, and we specify those properties in terms of an adversary," wrote Abadi and Andersen. "Thus, a system may consist of neural networks named Alice and Bob, and we aim to limit what a third neural network named Eve learns from eavesdropping on the communication between Alice and Bob."
AIs Becoming Smarter
Alice was able to create her own encryption, and Bob managed to figure out how to decrypt it. This was all achieved after more than 15,000 tries. It reached a point where Bob could decrypt Alice's text messages back into normal plain text, and Eve could only figure out eight of the 16 bits that form the message.
Here's something rather interesting. You see, researchers have no idea how the encryption works. Machine learning can provide the solution without much fuss, but when it comes down to understanding the whole thing, that's when problem arises.
Also, one cannot guarantee security would be tight if these practices are used in the real world, so as it stands, AI is not yet ready to create encryptions from scratch, but that day will come as they learn.
Google Brain is pretty awesome, for because of it, the Google Translate feature has gotten a huge boost in accuracy.
