Subliminal Images of Pasta Easily Fooled Google's Bots

Is this a video about gorillas, or pasta?

|
Apr 4 2017, 3:23pm

THANAKRIT PRANGTHONG/Shutterstock

As humans, we're pretty good at telling the difference between a video showing dumb robots failing to climb up the stairs and a video of a group of tigers chasing a drone. But can a machine tell what's in a YouTube video?

As it turns out, a subliminal image of a pasta bowl inserted every two seconds within a video about Gorillas and the famous scientist Jane Goodall tricks Google's artificial intelligence bots into thinking the video is about spaghetti, according to a new experiment.

Read more: Inside Wikipedia's Attempt to Use Artificial Intelligence to Combat Harassment

A group of researchers from University of Washington played around with Google's Cloud Video Intelligence API, a new tool that's still in private beta that uses machine learning to automatically recognize and label what's inside a video. By inserting random pictures at regular intervals—every two seconds—the researchers successfully tricked the tool into mislabelling videos. That way, the tool categorized a video about tigers as a video of a car because the researchers put a picture of an Audi in it, as the they explained in their paper.

"The API completely fails to correctly understand the video content," the researchers, who called their technique "image insertion attack," wrote.

The subliminal pictures used by the researchers to trick Google's new API into mislabelling videos. (Image: University of Washington)

Google did not respond to a request for comment in time for publication.

Every day, hundreds of hours of videos get uploaded to the internet, so a tool that automatically detects what's inside a video, labels it correctly, and annotates its content so that humans can use text search to go through it would be immensely useful. For example, it'd be much easier to go through surveillance video to find suspects, find the exact moment a certain play happened during a basketball game, or filter out videos with inappropriate content that get uploaded to YouTube or Facebook.

With their simple experiment, University of Washington researchers Baicen Xiao, Radha Poovendran and Hossein Hosseini proved that we're still far away from a future where bots are as smart as we are—at least when it comes to telling pasta and gorillas apart.

Subscribe to pluspluspodcast , Motherboard's new show about the people and machines that are building our future.