Google Neural Networks Build Encryption
by Laurie Sullivan, Staff Writer @lauriesullivan, November 1, 2016
Google’s Brain team taught neural networks built on machine learning to create their own form of encryption. A research paper explains how.
The paper, written by Martín Abadi and David G. Andersen, details “whether neural networks can learn to use secret keys to protect information from other neural networks.”
Each network was given a name and assigned a task. The networks named Alice and Bob passed each other notes using an encryption method that each created themselves once they recognized the sensitivity of the document being shared. A third AI neural network named Eve tried to intercept and decrypt the messages.
Even more interesting, each were assigned a task. Abadi and Andersen assigned letters for each action such as “P,” which might signify a confidential message to Alice. The goal was for Alice and Bob to communicate clearly and hide their communication from Eve. Eve would learn to eavesdrop on the communication between Alicfe and Bob and decode the message.
It took about 15,000 attempts for the AIs to archive their goal, according to the paper. The message — which was a mere 16 bits long, with each bit a 1 or a 0 — meant the AIs did more guessing than predicting.
The study also set out to determine whether neural networks can learn what to protect and what to encrypt. The project did demonstrate that neural networks can learn to protect the confidentiality of their data from other neural networks. They can discover forms of encryption and decryption without being taught specific algorithms for these purposes.
Abadi and Andersen found that “neural networks may be useful not only for cryptographic protections but also for attacks. While it seems improbable that neural networks would become great at cryptanalysis, they may be quite effective in making sense of metadata and in traffic analysis.”
New Scientist called attention to the paper, which can be downloaded here.
MediaPost.com: Search Marketing Daily
(32)
