In new research released last week, a team of MIT computer science students managed to trick Google’s Cloud Vision artificial intelligence into thinking that a picture of four machine guns was probably a helicopter. They did it by carefully manipulating the underlying pixels of an original image, changing it in ways that were imperceptible to humans but completely disorienting for the AI.
The team demonstrated several other tricks, including convincing Cloud Vision that a group of skiers were actually a dog. They did it all without access to the vision system’s underlying code, a so-called “black box” scenario. The research points towards potential vulnerabilities in the systems behind technology like self-driving cars, automated security screening systems, or facial-recognition tools.
To fool the system, the researchers manipulated the original image pixel-by-pixel, changing it in ways humans couldn’t detect but which, bit by bit, altered what Cloud Vision saw. It sounds not too different from brute-force password hacking, in which a malicious algorithm plugs in letters and numbers until it finds the combination that opens your email.
Get Data Sheet, Coins2Day’s technology newsletter.
Speaking with Wired, one of the researchers said that this sort of randomized hack can actually help us better understand how artificial intelligences think. Google and other big tech firms, meanwhile, are working to address these sorts of attacks, hopefully before their real-world applications become more widespread.
One significant qualifier is that this particular trick relies on digital alteration of 2-D images, while something like a self-driving car draws on much richer, less easily manipulated visual data. But lo-fi, real-world hacks have also been used to trick AI vision systems—for instance, when carefully-placed stickers were recently used to make an AI misread traffic signs.
The research hasn’t completed peer review yet, but a preliminary version can be read here.
