Google researchers say they can fool image recognition tools by distracting them with shiny stickers. It’s an alternative to trying to disguise the object the system is looking for.
This isn’t the same approach as a recent experiment that used carefully placed modifications to make one thing look like another, including 3D-printing replica turtles that consistently made image recognition software ‘see’ rifles.
Instead the researchers took hundreds of photos of a particular object and used them to create images that look psychedelic to the human eye but are designed to combine all the signals an AI system looks for when recognizing that object.
In one example, a small sticker with an image based on the most ‘toastery’ aspects of toasters was enough to make the system see a toaster, even though the sticker was placed next to a much larger banana which a human would see as the main object in the scene. (The sticker only took up a tenth of the scene.) While this worked, the system was able to spot the banana after the researchers replaced the sticker with a large photograph of a single toaster.
It shows how, even though computers use the same technique as humans when looking for a specific object (namely, check for characteristics you’ve learned are consistently present in that object), but that such systems haven’t yet got close to the same level of ability.
The researchers say that in theory the method could be used to fool image recognition systems such that they fail to spot what is there because they are distracted by the composite image.
(Image Credit: Tom Brown)