A comparative analysis of UAV computer vision systems in physically realizable adversarial AI scenarios

Le résumé fourni par la source

Computer vision systems with object recognition are seeing broad adoption in uncrewed aircraft systems for aided target recognition. One operationally relevant implementation pairs electro-optical cameras with computer vision models to recognize camouflage patterns worn by soldiers, enabling friend-versus-foe identification. Adversarial camouflage patches offer a means of defeating these models by inducing a measurable signature change when worn by a subject. This work presents a performance assessment of adversarial camouflage patches against a camouflage recognition model using EO imagery collected from a Skydio X2D UAS. Field experiments were conducted across multiple altitudes, illumination conditions, and viewing geometries using dynamic human subjects wearing two adversarial patch types at two uniform locations, compared against a control group with no patch worn. EO video was recorded to native onboard storage and processed post-flight using a common object-detection pipeline. Results indicate a general increase in adversarial patch effectiveness at higher altitudes and under cloudy as opposed to sunny conditions. The effect of viewing angle on patch performance varied considerably across patch types and worn locations. Overall, the digitally color-matched patch demonstrated generally superior performance relative to the glare reduction patch.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.