AI Invisibility Cloak Developed
AI Invisibility Cloak Developed – A team from University of Maryland in the U.S. has developed an invisibility cloak that uses adversarial clothing patterns to evade video analytics.
The development is worth noting from the point of view of counter measures, but it’s also of interest from the point of view of potential threats.
Researchers used detection datasets – images from the field of view – to create patterns they said “suppress the objectness scores produced by a range of detectors (AI algorithms).
The research team loaded images from the COCO detection dataset and passed them through a detector. A gradient descent algorithm was then used to find the clothing pattern that most minimized the objectness scores (confidence in the presence of an object) for every object prior.
The pattern most effective at minimizing objectness was then reproduced as clothing, but it could be created as a screen in cardboard, poly, or, even be rendered onto a vehicle. What’s interesting, too, is that the presence of adversarial patterns seems to blind AI to clearly human aspects of the wearer – head, legs, people standing behind, etc.
As AI becomes increasingly important to security operations, staying abreast of threats and countermeasures is going to be vitally important for security managers, consultants and system designers.
#SEN #SENnews #security #electronics