Slashdot Mirror
To Protect AI From Attacks, Show It Fake Data (technologyreview.com)
AI systems can sometimes be tricked into seeing something that's not actually there -- remember when Google's software "saw" a 3-D-printed turtle as a rifle. At an event earlier this week, Google Brain researcher Ian Goodfellow explained how AI systems defend themselves. From a report: Goodfellow is best known as the creator of generative adversarial networks (GANs), a type of artificial intelligence that makes use of two networks trained on the same data. One of the networks, called the generator, creates synthetic data, usually images, while the other network, called the discriminator, uses the same data set to determine whether the input is real. Goodfellow went through nearly a dozen examples of how different researchers have used GANs in their work, but he focused on his current main research interest, defending machine-learning systems from being fooled in the first place.
[...] GANs are very good at creating realistic adversarial examples, which end up being a very good way to train AI systems to develop a robust defense. If systems are trained on adversarial examples that they have to spot, they get better at recognizing adversarial attacks. The better those adversarial examples, the stronger the defense.
[...] GANs are very good at creating realistic adversarial examples, which end up being a very good way to train AI systems to develop a robust defense. If systems are trained on adversarial examples that they have to spot, they get better at recognizing adversarial attacks. The better those adversarial examples, the stronger the defense.
This sounds hokey, but I'm serious. If we create something that can learn on it's own it would eventually, even if we prevented it, be able to rewrite at least some of it's own routines and safeguards, even though that we didn't, necessarily, want it to be able to overwrite. As humans we are limited by our hardware capacity but as code running in "the cloud" able to allocate more resources to the growth of it's "mind" or even just to rewrite it's own code to be more efficient than a human would have been able to do it could grow and develop exponentially and potentially access everything the particular network or the internet has to offer to expand it's "mind" Combine that with millions of internet connected moving devices, industrial robots, etc ...... even if it isn't sentient, even if it is just glitchy and smart it doesn't sound like a good time for humanity.
Nick Bostrom: "Superintelligence. Paths, Dangers, Strategies." https://global.oup.com/academi...
I'll tell you what: you write a simple program that can reprogram itself and I will listen to this drivel. We can barely even create functional software. You guys think if you connect a million computers together then you will have AI. But really what you have is the Internet. And that ain't intelligent.
We are now programming AI to think critically more than we are teaching our children to do the same.
I welcome our robotic AI overlords.
Don't trust any concentration of power.