快猫短视频

EVER had an in-joke only you and your best friend knew the punchline to? According to computer scientists, artificial brains could use this kind of secret knowledge to deliver coded messages.

Neural networks are computer systems designed to mimic the brain. Just as the brain鈥檚 nerve cells are wired together in a complex mesh, neural nets consist of a web of electrical switches, or a computer simulation of these connections.

When neural networks tackle a problem, connections that are ultimately successful become stronger than those that give a wrong answer. The more lessons a network trains with, the better it learns which pathways to follow to find the right answers.

What happens, then, when two different neural networks are used to train each other? Wolfgang Kinzel of the Institute for Theoretical Physics in Wurzburg, Germany, and Ido Kanter of the Minerva Center in Ramat-Gan, Israel, tried it with old-fashioned hardware networks, and found that the two met in the middle, becoming mirror images.

In each lesson, the scientists asked the computers to categorise unique, random pieces of information with the aim of getting the same answer as their partner. After each round, they compared each other鈥檚 results. In a surprisingly short time, the two networks became aligned so that their properties are equal and opposite at every point. Connections that flowed one way in one network went in the opposite direction in its partner.

From there, it is a simple step for one of the pair to reverse all its weightings so the two networks end up identical. They would have the same weightings, without ever having told each other what they were.

The researchers realised that this phenomenon could be useful in cryptography. At present, computers that need to exchange information securely use codes or 鈥渒eys鈥� based on huge numbers. But one weakness of this system is that the sender has to secretly tell the receiver what the key is before they can start exchanging messages. An eavesdropper who hears the key will be able to decode any subsequent communications. But synchronised neural networks could use their hidden weightings as the key.

Imagine two friends talking in public surrounded by eavesdroppers. If the friends share an in-joke, the spies鈥攏ot having shared the same unique experiences鈥攚ill have a hard time figuring out what is going on. Similarly, synchronised networks will jump to the same conclusion, given the same limited information.

Immediate applications might include anything that needs to send information rapidly and securely, such as mobile phones, video conferencing and Internet communication. Kinzel even speculates that living organisms might be using the same principle to transfer information between different parts of the nervous system.

The technique could be quite powerful, says computer engineer Don Wunsch at the University of Missouri in Rolla. 鈥淚 could see it becoming an alternative when users need to create a cheap and fast encryption with a minimum of shared communication, when security is of moderate, but not life-and-death, concern.鈥�