The computers have a secret, and some people are worried.
According to a story from New Scientist, researchers working on the Google Brain Project announced recently that computer systems they created based on a system of artificial neurons have not only created a basic encryption technique, but have learned how to keep it a secret.
In a paper titled, “Learning to protect communications with adversarial neural cryptography,” researchers with the Google Brain Project, which is a deep learning research venture at Google, reported that two neural network systems they created worked together to come up with a message, encrypt the message and then decode it.
In the paper, researchers said, “'We ask whether neural networks can learn to use secret keys to protect information from other neural networks.” That question was answered when the neural network that created the encryption system did just that.
Researchers with the Google Brain Project use deep learning, a branch of machine learning based on a set of algorithms, to conduct research.
In deep learning, computer systems called neural networks use algorithms, or specific rules to be followed in calculations, to try to teach themselves how to do certain tasks.
Here’s how the encryption processed worked:
Researchers Martín Abadi and David Andersen created three “neural networks,” or computing systems using something like an artificial neuron. A neuron in the human body is a nerve cell responsible for transmitting information to other nerve cells, gland cells, or to muscle.
The neural networks created by Abadi and Anderson have names: Alice, Bob and Eve. Each of the networks had a specific job. For Alice, it was to send a secret message to Bob; for Bob, it was to decode the message that Alice sent; for Eve, it was to eavesdrop on Alice’s message and try to decode it.
>> Got a question about the news? See our explainers here
According to the paper, something interesting happened. Alice eventually, and on her own, created an encryption system, despite the fact she was not taught how to build out such a system. The paper emphasized that the encryption system Alice came up with is very basic, but the fact that Alice was able to create something no one thought neural networks could achieve was remarkable in itself.
What Alice created was a 16-bit message with each bit representing either a “1” or a “0.” She took the original message and mixed it up (cipher text) before sending it to Bob. Alice and Bob had an agreed upon a set of numbers that was the key used to decipher the message, according to the researchers.
Next, Bob did his part by converting Alice’s cipher text message back into plain text. It took 15,000 attempts, but Bob got it.
Then came Eve. Eve’s job, remember, was to try to figure out what Alice and Bob were saying. Eve managed to get half of the message figured out, but researchers say Eve’s attempt was more akin to someone getting it right by guessing the numbers.
Scientists did see something notable from the research – the networks were not very good, initially, at creating, decoding, or trying to decipher a coded message, but with practice, they not only created a system, but kept that system secret from the developers. As a post in New Scientist pointed out, the way the machine learning works prevents even the researchers from figuring out what kind of encryption method Alice came up with.
The neural networks are considered a form of artificial intelligence, which encompasses the theory that computer systems will be able to perform tasks that normally require human intelligence.
Famed scientists Stephen Hawking, director of research at the Centre for Theoretical Cosmology within the University of Cambridge, warned at a conference in London last week that artificial intelligence could develop a will of its own that is in conflict with that of humanity. Or, he said, if care is taken to avoid certain risks, it could be helpful to humanity.
"Alongside the benefits, AI will also bring dangers, like powerful autonomous weapons, or new ways for the few to oppress the many," Hawking said.
To read the research paper, click here.
About the Author
