Researchers develop patch to deceive surveillance cameras

A team composed of three Belgian researchers has developed a tool that makes users undetectable for person recognition software, reported cybersecurity specialists from the International Institute of Cyber Security (IICS).

Wiebe Van Ranst, Simen Thys and Toon Goedeme,
from the KU Leuven University, presented their research entitled “Cheating
Automated surveillance
cameras: a patch to avoid individual detection” in a recent
computer security event.

In broad terms, cybersecurity experts described
the used method, which consists of a signaling that can defend the carrier
against Darknet, an open source neural network framework with support for You
Only Look Once (YOLO), real-time object detection system.

The use of the so called “adversarial
images”, to deceive the detection systems that work with machine learning,
increasingly call the attention of the cybersecurity community. Although the
research work in this area is not scarce, the Belgian experts say that their
method goes further, because previous research does not address something as
diverse as people.

“The goal is to bypass the detection of
security systems that use people recognition just as people enter the camera’s
visual field,” the investigators say.

The researchers focused on YOLOv2; to achieve
their goal, they added a set of data to images to return bounding boxes that
surround the person’s image in the detection algorithm. “In a fixed
position relative to these bounding boxes, then we apply the current version of
our patch to the image”, is explained in the document.

“The resulting image is sent to the
detector and measured the number of times the algorithm manages to detect the
person after this process; the optimizer then changes the pixels in the patch
to increase the detection protection level”.

The result of this process appears on the screen
in the form of a multicolor patch of about 40cm². Researchers hope to extend
their work to other neural network architectures like Faster R-CNN, and believe
that they could print this pattern on a T-shirt that would render the user  “virtually invisible” to the surveillance
cameras detection algorithms.

Comments are closed.