Google Brain Deep Learning Team Teaches Neural Networks To Invent Their Own Encryption
We often joke about certain advances in technology leading to Skynet scenarios where machines wage war with humans, but sometimes it feels inevitable. Take for example what a team of researchers from Google Brain, Google's deep learning project, have discovered. In our quest to advance machine learning capabilities, neural networks are now able to devise their own encryption schemes, which in turn could allow them to communicate in secret with each other.
Potential for human extinction aside, it's a rather fascinating thing. Neural networks are computer systems loosely modeled after the neural structure of the brain. Researchers Martin Abadi and David Andersen demonstrated that neural networks can devise a simple encryption technique using machine learning. What's more, they can do it without being taught specific cryptographic algorithms.
This is basic level encryption we're talking about, at least for now. The Google Brain team took three neural nets called Alice, Bob, and Eve and trained each one to be adept at its own role in the communication chain. Alice was in charge of sending a secret message, Bob had to decode it, and Eve was tasked with eavesdropping on the communication.
To prevent Eve or any other outsiders from being able to read the message, Alice converted her plain-text message into cipher text, basically a bunch of characters that look like a high-tech bowl of alphabet soup. If it were to be intercepted or otherwise spied by Eve, all she would see is a bunch of unintelligible text, while the original message would be kept secret.
The challenge was that Bob had to be able to decipher the message, so Alice and Bob started out with a set of numbers, essentially a key to decode the communication. Since Eve didn't have access to the key, she couldn't decrypt the communication like Bob could.
In the beginning, the neural nets didn't do a great job with this process. However, things improved over time as Alice came up with her own encryption scheme, and Bob figuring out how to decrypt it. After 15,000 rounds of this, Bob could successfully decrypt Alice's cipher text message back into plain text, while Eve could only work out 8 of the 16 bits comprising the message.
This is a big development as encryption was never a point of focus when developing neural nets.
Potential for human extinction aside, it's a rather fascinating thing. Neural networks are computer systems loosely modeled after the neural structure of the brain. Researchers Martin Abadi and David Andersen demonstrated that neural networks can devise a simple encryption technique using machine learning. What's more, they can do it without being taught specific cryptographic algorithms.
This is basic level encryption we're talking about, at least for now. The Google Brain team took three neural nets called Alice, Bob, and Eve and trained each one to be adept at its own role in the communication chain. Alice was in charge of sending a secret message, Bob had to decode it, and Eve was tasked with eavesdropping on the communication.
To prevent Eve or any other outsiders from being able to read the message, Alice converted her plain-text message into cipher text, basically a bunch of characters that look like a high-tech bowl of alphabet soup. If it were to be intercepted or otherwise spied by Eve, all she would see is a bunch of unintelligible text, while the original message would be kept secret.
The challenge was that Bob had to be able to decipher the message, so Alice and Bob started out with a set of numbers, essentially a key to decode the communication. Since Eve didn't have access to the key, she couldn't decrypt the communication like Bob could.
In the beginning, the neural nets didn't do a great job with this process. However, things improved over time as Alice came up with her own encryption scheme, and Bob figuring out how to decrypt it. After 15,000 rounds of this, Bob could successfully decrypt Alice's cipher text message back into plain text, while Eve could only work out 8 of the 16 bits comprising the message.
This is a big development as encryption was never a point of focus when developing neural nets.
