Skip to main content

That turtle is a gun! MIT scientists highlight major flaw in image recognition

Fooling Neural Networks in the Physical World
When is a rifle actually a 3D-printed turtle? When is an espresso actually a baseball? A fascinating, yet alarming new piece of research from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) shows that it’s possible to create objects that are able to trick Google image recognition algorithms into thinking they are looking at something else entirely.

In their paper, the team of MIT researchers describe an algorithm which changes the texture of an object just enough that it can fool image classification algorithms. The proof of what the team calls “adversarial examples” turns out to be baffling to image recognition systems, regardless of the angle the objects are viewed from — such as the 3D printed turtle which is consistently identified as a rifle. That’s bad news for security systems which use A.I. for spotting potential security threats.

“It’s actually not just that they’re avoiding correct categorization — they’re classified as a chosen adversarial class, so we could have turned them into anything else if we had wanted to,” researcher Anish Athalye told Digital Trends. “The rifle and espresso classes were chosen uniformly at random. The adversarial examples were produced using an algorithm called Expectation Over Transformation (EOT), which is presented in our research paper. The algorithm takes in any textured 3D model, such as a turtle, and finds a way to subtly change the texture such that it confuses a given neural network into thinking the turtle is any chosen target class.”

While it might be funny to have a 3D-printed turtle recognized as a rifle, however, the researchers point out that the implications are pretty darn terrifying. Imagine, for instance, a security system which uses AI to flag guns or bombs, but can be tricked into thinking that they are instead tomatoes, or cups of coffee, or even entirely invisible. It also underlines frailty in the kind of image recognition systems self-driving cars will rely on, at high speed, to discern the world around them.

“Our work demonstrates that adversarial examples are a bigger problem than many people previously thought, and it shows that adversarial examples for neural networks are a real concern in the physical world,” Athalye continued. “This problem is not just an intellectual curiosity: It is a problem that needs to be solved in order for practical systems that use deep learning to be safe from attack.”

Editors' Recommendations

Luke Dormehl
I'm a UK-based tech writer covering Cool Tech at Digital Trends. I've also written for Fast Company, Wired, the Guardian…
This AI cloned my voice using just three minutes of audio
acapela group voice cloning ad

There's a scene in Mission Impossible 3 that you might recall. In it, our hero Ethan Hunt (Tom Cruise) tackles the movie's villain, holds him at gunpoint, and forces him to read a bizarre series of sentences aloud.

"The pleasure of Busby's company is what I most enjoy," he reluctantly reads. "He put a tack on Miss Yancy's chair, and she called him a horrible boy. At the end of the month, he was flinging two kittens across the width of the room ..."

Read more
Digital Trends’ Top Tech of CES 2023 Awards
Best of CES 2023 Awards Our Top Tech from the Show Feature

Let there be no doubt: CES isn’t just alive in 2023; it’s thriving. Take one glance at the taxi gridlock outside the Las Vegas Convention Center and it’s evident that two quiet COVID years didn’t kill the world’s desire for an overcrowded in-person tech extravaganza -- they just built up a ravenous demand.

From VR to AI, eVTOLs and QD-OLED, the acronyms were flying and fresh technologies populated every corner of the show floor, and even the parking lot. So naturally, we poked, prodded, and tried on everything we could. They weren’t all revolutionary. But they didn’t have to be. We’ve watched enough waves of “game-changing” technologies that never quite arrive to know that sometimes it’s the little tweaks that really count.

Read more
Digital Trends’ Tech For Change CES 2023 Awards
Digital Trends CES 2023 Tech For Change Award Winners Feature

CES is more than just a neon-drenched show-and-tell session for the world’s biggest tech manufacturers. More and more, it’s also a place where companies showcase innovations that could truly make the world a better place — and at CES 2023, this type of tech was on full display. We saw everything from accessibility-minded PS5 controllers to pedal-powered smart desks. But of all the amazing innovations on display this year, these three impressed us the most:

Samsung's Relumino Mode
Across the globe, roughly 300 million people suffer from moderate to severe vision loss, and generally speaking, most TVs don’t take that into account. So in an effort to make television more accessible and enjoyable for those millions of people suffering from impaired vision, Samsung is adding a new picture mode to many of its new TVs.
[CES 2023] Relumino Mode: Innovation for every need | Samsung
Relumino Mode, as it’s called, works by adding a bunch of different visual filters to the picture simultaneously. Outlines of people and objects on screen are highlighted, the contrast and brightness of the overall picture are cranked up, and extra sharpness is applied to everything. The resulting video would likely look strange to people with normal vision, but for folks with low vision, it should look clearer and closer to "normal" than it otherwise would.
Excitingly, since Relumino Mode is ultimately just a clever software trick, this technology could theoretically be pushed out via a software update and installed on millions of existing Samsung TVs -- not just new and recently purchased ones.

Read more