Google Brain researchers say they have developed artificial intelligence (AI) technology that can form encryptions and communicate securely.
The researchers found by teaching an AI to value secrecy, neural networks can learn to protect their communications without being prescribed a set of cryptographic algorithms.
Two neural networks, Alice and Bob, were given a shared secret key, and a separate network, Eve, was tasked with decrypting the communication between Alice and Bob. All three robots had the same neural network architecture but had no connection other than Alice and Bob's shared key. The key and plaintext were entered into the first layer of Alice's neural network, while the key and Alice's ciphertext output were input for Bob; Eve received just the ciphertext. Each network was provided a loss function, and Eve and Bob's plaintext had to be as close to the input plaintext as possible, while Alice's loss depended on Eve's accuracy, creating an adversarial generative network among the robots.
Alice and Bob created a system in which they could communicate with few errors, but although Eve showed some improved degree of accuracy, Alice and Bob responded by improving their cryptography technique.
Researchers say future work might focus on steganography, as well as asymmetric (public-key) encryption.
From Ars Technica
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA