This was, in fact, the reaction the scientists were hoping for. Using subtle alterations imperceptible to the human eye, they changed the objects in a way that would make them unrecognizable to artificial intelligence. The technique is referred to as an adversarial attack, a way to fool AI without being evident to humans.An AI algorithm identifying a turtle as a rifle or a baseball as coffee may seem innocuous and amusing, but the implications are worrisome. Dawn Song, a computer scientist at the University of California, Berkeley, and colleagues performed an experiment last year in which they put stickers on a stop sign to confuse a common type of image recognition artificial intelligence. It worked – the AI identified the stop sign as a 45 mile per hour speed limit sign. Imagine all the stickers and graffiti you’ve seen on street signs before, and then imagine autonomous cars trying to identify them – it could be disastrous.
Song also mentioned a trick in which a Hello Kitty was placed in an image recognition AI’s view of a street scene. The cars in the scene simply disappeared. If a world full of autonomous cars truly is coming in the future as many say, catastrophic accidents could be caused by hackers messing with image recognition AI’s vision. That’s why experiments such as these are performed by scientists – to bring the weaknesses of these systems to light so they can be improved.
The scariest thing is the way that hackers can potentially change images to fool AI without humans being aware. In what is called a white box attack, hackers can see the AI’s gradients, which describe how a slight change in an input image or sound can move the output in a certain direction. By knowing the gradients, the hackers can calculate how to slightly alter inputs, a little bit at a time, to create the wrong output, causing the AI to call a turtle a rifle, for example. It’s a big change to the artificial intelligence, yet so slight that humans can’t see it.
Artificial intelligence developers are working on techniques to combat these attacks. For example, one technique embeds image compression as a step in image recognition AI. This adds jaggedness to the gradients. But these methods can be cleverly outwitted, too. In another recent paper, a team of scientists analyzed nine image recognition algorithms from a recent conference. Seven of the algorithms relied on gradient obfuscation, and the team was able to break all seven using ways like sidestepping the image compression. Each case took no more than a couple of days to hack.
Another possibility is to train an algorithm with certain constraints that keep it from being confused, using verifiable, mathematical methods. Song is concerned about the real world limitations of these defenses, however.
“There’s no mathematical definition of what a pedestrian is, so how can we prove that the self-driving car won’t run into a pedestrian?” she said. “You cannot!”
These examples are unnerving reminders of how imperfect artificial intelligence is, and how easily corrupted it can be. Should we abandon the idea of autonomous cars altogether, then? Of course not – but the fact remains that with all technology, there is a constant race between hackers and those who try to safeguard against hacking. The fact that these weaknesses have been exposed means that developers and programmers can better prepare for such attacks, and possibly make their AI algorithms more foolproof by the time we actually see the streets filling up with autonomous cars.
If you’d like to read the full paper, titled “Synthesizing Robust Adversarial Examples,” about 3D printed items being used to fool AI, you can access it here. Authors include Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok.
Discuss this and other 3D printing topics at 3DPrintBoard.com or share your thoughts below.[Source: Science Magazine]