Researchers at Google say they have created psychedelic stickers that can fool image-recognition software into seeing objects that do not exist.
In an example, the team produced colorful computer-generated patterns by sampling hundreds of photographs of a toaster; when the patterns were put next to another item--a banana--many neural networks saw the toaster instead.
The team says this method could be used to "attack" image-recognition systems, as these patches can be printed, added to a scene, photographed, and presented to image classifiers.
Even if the patches are small, they cause the classifiers to ignore other items in the scene and report a chosen target class. The researchers note this works because the computer-generated pattern is more "salient" to image-recognition software than real objects.
The team found the pattern consistently tricked software when it comprised at least 10 percent of a scene, while a photo of a real object was less likely to distract the software from another object.
From BBC News
View Full Article
Abstracts Copyright © 2018 Information Inc., Bethesda, Maryland, USA
No entries found