Google-Skynet is here

Google-Skynet is here

By Christopher D. Baker – 04/11/2016

Google Brain Creating Skynet? Can AI Hide Messages?

Can artificial intelligence (AI) create its own cryptographic algorithms? Google Brain, a research team set out to answer this question.

We learn within the paper they released [1] that this is the case, and that these AI neural nets can indeed create their own encrypted messages.

The study was conducted using three AI neural networks, named Alice, Bob, and Eve. Each assigned different tasks to resemble a potential real world scenario. Alice was tasked with sending encrypted messages to Bob, who was in turn tasked with the decryption of said messages.  Both using a shared secret key.

Eve’s purpose on the other hand was to attempt to crack the messages without any key.

Standard security testing scenario you may believe, but here’s the kicker, they were to make it up as they went.

Alice and Bob were tasked to only communicate successfully and to prevent eve from reading the messages. With no pre-programmed notion of known cryptosystems.  Allowing the two to develop their own encryption algorithms to thwart Eve.

With over 150,000 simGoogle Brain Creating Skynet? Can AI Hide Messages?ulations were run, with Eve never being fully successful at decryption. But, then again there were some runs that even Bob couldn’t decrypt the messages. They did evolve a system that Alice and Bob could communicate with little errors, while Eve showed some improvement from just random guessing, this was quickly superseded by Alice and Bob developing even stronger encryption methods that left Eve in the dust, as it were.

 

 

 

The researchers concluded, that neural networks can be used to protect communications, just by setting Alice to value secrecy as paramount. But cryptography is a far greater field than just symmetric encryption, and the researchers , Martín Abadi and David G. Andersen, said future study into the areas of stenography (hiding something within something else) and asymmetrical encryption may be forthcoming.   But they also concluded with a negative note, that Eve would unlikely be able to be a decent enough challenge for Alice and Bob, as neural networks are unlikely to be great at crypt-analysis, but perhaps in the field of metadata comprehension in traffic analysis.

References

[1] “LEARNING TO PROTECT COMMUNICATIONS WITH ADVERSARIAL NEURAL CRYPTOGRAPHY” https://arxiv.org/pdf/1610.06918v1.pdf (Visited: 04/11/2016)