Google researchers created a bunch of optical illusion stickers that trick artificial intelligence programs to see something that’s not there in the first place. The team designed a pattern which was able to fool their AI software that it was looking at a toaster while it was looking at a banana.
This test stands as a fine example of how these stickers or images could be used as possible means of attacking image recognition systems, which also includes hacking the system used in self-driving cars to avoid collisions. There are chances that hackers could place a sticker on road sign to fool an AI vehicle to think it is a stop signal on a high-speed road which could a lot of damage. It can even go as far as people slipping past AI security by simply wearing mask or glasses with certain patterns.
The researchers created a very powerful design from the sample of hundreds of toaster photographs. The image created was then placed near a banana. The machine registered seeing a toaster and not a banana. The team also offers printable versions which people can take home and fool AI.
The researcher team, led by Tom Brown further classified in their paper that these pictures can be printed out, photographed, added to any scene and then presented to image classifiers. Even if the picture is small, the classifiers ignore the other items in the scene to report a target class.
The whole exploit works by exploiting computers to prioritize on certain pixels in the image that they’re analyzing. The researchers added that these stickers work because the pattern they created is much more noticeable than the real world objects.
‘While images may contain several items, only one target label is considered true, and thus the network must learn to detect the most ‘salient’ item in the frame,‘ they added.