An AI Figured Out How to Cheat at its Job and Hide the Evidence From Humans
Whether it's rubber-banding in Mario Kart or fudging moves in Street Fighter, AI will do whatever it needs to stay a step ahead of the humans challenging it. Of course, a CPU opponent cheating in a video game is different from an AI teaching itself to cheat, then hiding the evidence, but that's what just what an artificial intelligence called CycleGAN did, according to a study published by Stanford University and Google.
It all had to do with maps: Google was looking for a way to convert aerial photos into simplified images that could be used in Google Maps, then convert the simplified images back into something resembling the original aerial photos. To accomplish this, they used an AI called CycleGAN, which was trained to convert between the two formats and then graded on its accuracy and efficiency. During the tests, however, CycleGAN showed a suspicious degree of skill at its task, leading researchers to take a closer look at the images it was producing.
What they found confirmed their suspicions: certain details that should have been erased in the conversion from aerial photo to simplified map (like skylights on buildings) reappeared when CycleGAN converted map images back into the "aerial photo" versions. Upon reviewing the map images, they found that the AI had embedded the original aerial photo into the new image, but had done it with colors that made it invisible to human eyes. When it was asked to convert the map image back to the aerial photo version, it just read the embedded photo and spat it out.
The phenomenon of hiding images or codes in other images is called steganography, and while CycleGAN had proved itself to be lackluster at its intended image-converting function, it showed itself to be more than capable of deceiving its human testers...at least for a while. They even titled the study describing its deception "CycleGAN, a Master of Steganography."
https://www.outerplaces.com/science/item/19211-ai-cheat-cyclegan-steganography
欺詐的AI被捉到使用隱寫術來隱藏數據
CHEATING AI CAUGHT HIDING DATA USING STEGANOGRAPHY
The CycleGAN is a network that excels at learning how to map image transformations such as converting any old photo into one that looks like a Van Gogh or Picasso. Another example would be to be able to take the image of a horse and add stripes to make it look like a zebra. The CycleGAN once trained can do the reverse as well, such as an example of taking a map and convert it into a satellite image. There are a number of ways this can be very useful but it was in this task that an experiment at Google went wrong.
https://hackaday.com/2019/01/03/cheating-ai-caught-hiding-data-using-steganography/
沒有留言:
發佈留言