Slashdot Mirror
Google's AI Created Its Own Form of Encryption (engadget.com)
An anonymous reader shares an Engadget report: Researchers from the Google Brain deep learning project have already taught AI systems to make trippy works of art, but now they're moving on to something potentially darker: AI-generated, human-independent encryption. According to a new research paper, Googlers Martin Abadi and David G. Andersen have willingly allowed three test subjects -- neural networks named Alice, Bob and Eve -- to pass each other notes using an encryption method they created themselves. As the New Scientist reports, Abadi and Andersen assigned each AI a task: Alice had to send a secret message that only Bob could read, while Eve would try to figure out how to eavesdrop and decode the message herself. The experiment started with a plain-text message that Alice converted into unreadable gibberish, which Bob could decode using cipher key. At first, Alice and Bob were apparently bad at hiding their secrets, but over the course of 15,000 attempts Alice worked out her own encryption strategy and Bob simultaneously figured out how to decrypt it. The message was only 16 bits long, with each bit being a 1 or a 0, so the fact that Eve was only able to guess half of the bits in the message means she was basically just flipping a coin or guessing at random.ArsTechnica has more details.
Would you agree there are varying levels of intelligence in the natural world?
Cute, but kinda disappointing. Basically, the "AI" kept banging on, randomly trying crap "cyphers" until they made something that a third "AI" couldn't break by randomly flipping bits until the text was decoded.
This isn't AI. This is more like an old game I played where you trained pseudo AI warriors by setting them loose on a battlefield and letting them learn by themselves how to fight and survive.
Essentially, they started as really stupid bots that couldn't even walk in a straight line. To teach them how to fight, you'd set up an objective (say, go to flag) and let them wander around by themselves. The game would "reward" your bots for completing or coming close to the objective. The reward came in the form of "fitness" points. At the end of a pre-determined time, the bots with the lowest fitness would be killed, and new bots would be spawned.
The bots that were spawned would have "programming" similar to the fit bots that survived the previous round, but with small-ish changes in their programming (for example, instead of always turning left every time it hits a wall, it might decide to go right with 50% probability).
Over thousands of iterations of randomly trying stuff, they'd eventually learn how to walk on a straight line. Then you'd teach them how to avoid obstacles by placing walls around the battlefield (and watch in dismay as your top of the line warriors walk straight into a wall for the first few hundred generations or so), and how to fight by rewarding them for killing enemy bots.
Once they were ready, you could set up battles and capture the flag type games with your bots.
It was kinda fun, but mainly it was a cute demonstration of natural selection in action (the, so called, genetic algorithms). You could learn a few things, like, for example, that brutally culling your bot herd by setting unreasonable objectives (reach objective flag in 5 seconds), and manually killing off anyone that doesn't meet your unreasonable criteria, would not necessarily produce more effective fighters, because you'd not be rewarding good fighters, you'd be rewarding people that rush straight into the objective, that would be killed by slower, more deliberate actors.
The game was called NERO: Neuro Evolving Robotic Operatives. I haven't played it in ages so I can't say how well it plays right now. You can find it here (I think): http://nerogame.org/
How about EI for "Electronic Intelligence"?
Okay, so an AI program created an encryption method that might be hard to break.
But what if it turns out to be an extremely inefficient method?
The whole goal of encryption research is to develop the fastest algorithm that offers a given level of protection against attack.
If algorithmic speed was not a goal of this AI approach, then it's not likely that the resulting algorithm will be practically useful.