Skip to main content

Google's neural networks develop their own encryption methods

In an effort to demonstrate how AI could be used to boost encryption, researchers at Google taught two neural networks how to communicate with one another while keeping their conversation secret from a third.

Researchers at the company's deep learning initiative, Google Brain, have successfully taught two neural networks given the nicknames “Alice” and “Bob” to secretly communicate with one another while keeping the details of their conversations from one called “Eve.” Last week the team behind this endeavor, published a paper detailing the process of the experiment and its results.

Neural networks operate by searching through copious amounts of data in order to find patterns and connections that are in turn used in their future computations. Google Brain was effectively able to prove that artificial intelligence (AI) could be harnessed to deal with data security and the increased difficulty of encrypting and decrypting critical messages.

The researchers began their experiment by having Alice send Bob a 16-digit encrypted message that was made up of ones and zeroes. These two neural networks began communicating with one another using a shared key but as their correspondence continued, the way in which their messages were encrypted began to change as well.

The rival neural network tasked with deciphering the messages, EVE was able to decrypt the first 7,000 messages that Alice and Bob sent to one another. However, eventually she was unable to continue to do so as the neural networks continually changed tactics with the aim of keeping their messages secret from Eve.

Martin Abadi and David G. Andersen, two of the researchers who worked on the project, highlighted how they were able to teach neural networks how to encrypt and decrypt messages, saying: “We demonstrate that the neural networks can learn how to perform forms of encryption and decryption, and also how to apply these operations selectively in order to meet confidentiality goals.

"While it seems improbable that neural networks would become great at cryptanalysis, they may be quit effective in making sense of metadata and in traffic analysis.” 

Image Credit: Asif Islam / Shutterstock

Anthony Spadafora
After living and working in South Korea for seven years, Anthony now resides in Houston, Texas where he writes about a variety of technology topics for ITProPortal.