A team of Google researchers has created a psychedelic stickers that can fool the image recognition software to see objects that are not there.
Using a toaster as an example, the team has produced colored computer models generated by the sampling of the hundreds of photos of the device.
When the models were put next to another element, such as a banana, many networks of neurons seen a toaster in the place.
The team says the method could be used to “attack” the image recognition systems.
“These contradictory, the patches can be printed, added to any scene, photographed and presented to the image classifiers,” the researchers said.
“Even when the plots are small, they cause the classifiers to ignore the other elements of the scene and report to a chosen target class.”
The researchers said that their trick has worked, because the computer generated pattern was more “salient” to the image recognition software that of real objects.
Rather than changing the appearance of elements to hide them, the researchers found their toaster-inspired patterns “distracted” image recognition software.
“While the images may contain multiple elements, only the label target is considered true, and so the network must learn to identify the most “salient” of the element in the framework,” they wrote in their research paper.
The model systematically deceived image recognition software when it has taken at least 10% of the same scene.
By comparison, a photo of a real toaster was much less likely to distract the software from another object in the scene, even in larger sizes.