Artificial Intelligence and machine learning have recently been one of the core focus areas for research at Google HQ. You might have heard of the company’s DeepMind neural network ‘AlphaGo’ beating the world champion Lee Sedol at the difficult game of Go. But, you wouldn’t have heard of this surprising feat achieved by a duo of Googlers in the field of AI and security.
According to a new research paper, Googlers Martín Abadi and David G. Andersen have demonstrated the capability of neural networks that can act as neurons. These also have the power to generate a human-independent encryption key to protect the data being exchanged between two different neural nets.
For a more realistic approach, the researchers have created three test subjects — Alice, Bob and Eve — who’re indeed neural nets which need to pass each other bite-sized notes. It is important to note that these neural networks are making their own encryptions using machine learning, without being taught specific cryptographic algorithms.
In this case, Alice and Bob are required to communicate using basic self-made encryption methods whereas Eve has to learn the info by eavesdropping and decoding the message herself. The researchers started the tests with a 16-bit long plain-text messages, with each bit being only a 0 or a 1. Alice has to take this original message, convert into a gibberish mess using her own encryption methods, while Bob has to try and decode the same using cipher keys.
Initially, both Alica and Bob were extremely poor at hiding any secret messages from Eve, who would swoop in and decipher the message easily. But over time the neural networks started getting tuned and about 15,000 attempts later, Alice figured out a specific encryption method(still basic) to send a message which can only be deciphered and read by Bob. In this attempt, Eve could guess just 8 of the 16 bits forming the message which is about half the probability of any event occurring in nature.
Though it sounds like an impressive exercise by the humanly-named AI neural nets created by Google, but it won’t really have any practical application in the near future. Alice has used machine learning to devise her own algorithm which the researchers are currently unaware of. Thus, this can only be an impressive feat for artificial intelligence as they’ve even started figuring out the sophistication of encryption methods.