No, it’s not a deleted Q gadget from some late-stage Pierce Brosnan 007 movie. Researchers really have created a patch that could effectively disguise aerial vehicles from A.I. image recognition systems designed to autonomously identify military objects.
The technology, developed by researchers at the Netherlands Organisation for Applied Scientific Research, is capable of consistently fooling the state-of-the-art YOLO (You Only Look Once) real-time object-detection system. And, potentially, others as well. It could be used to help defend fighter planes from enemy drones.
“We have shown that a relatively small patch, roughly 10% of the size of the plane, is effective in camouflaging the whole plane against automatic detection,” Ajaya Adhikari, one of the researchers on the project, told Digital Trends. “These small patches seem to be a more practical solution for camouflage than covering the whole plane.”
The high-tech patches are a different twist on camouflage: A type of disguise that is intended to fool machine, rather than human, vision. In recent years, research into the field called “adversarial A.I.” has continued to grow. Adversarial A.I. is capable of exploiting vulnerabilities in the way that A.I. systems look at images and classify them. Previous examples include work by researchers who were able to get an image-recognition system to classify a 3D-printed turtle as a gun and a baseball as an espresso simply by tweaking their surface pattern.
“To the best of our knowledge, we are the first who have explored adversarial A.I. techniques for camouflage in aerial surveillance,” Richard den Hollander, the other lead researcher on this latest project, told Digital Trends. “The results of our work show that adversarial camouflage can be considered as a potential alternative to traditional camouflage when using deep learning models for automatic analysis.”
In an abstract describing their work, the researchers note the following: “Our results show that adversarial patch attacks form a realistic alternative to traditional camouflage activities, and should therefore be considered in the automated analysis of aerial surveillance imagery.”
Don’t expect the military to start slapping these patches on planes and drones just yet, though. The investigators said that more research is still needed to validate the approach. This will include performing field tests with a printed adversarial patch on actual objects in aerial views, along with investigating the effect of camouflage on other detection models.
A paper describing the work, titled “Adversarial Patch Camouflage against Aerial Detection,” is available to read online.
