From Pixels to Perception: Adversarial Camouflage Patterns Designed to Evade Detection Models and Human Observation

Camouflage techniques can effectively evade detection and tracking systems that rely on human visual perception or optical sensors during aerial reconnaissance of military targets. While the advent of deep neural network-based object detection models has enabled the identification of even well-camouflaged targets, these models are vulnerable to adversarial examples. Adding perturbations to camouflage patterns can attack the detection models, but this deteriorates the appearance of the camouflage and can be easily recognized by human vision. The purpose of this paper is to develop adversarial camouflage patterns that not only undermine object detection models but also deceive human perception. The research methodology involves synthesizing camouflage patterns for varied terrain environments using image processing techniques and feeding them into a custom GAN framework to create deceptive adversarial camouflage patterns. The experimental results demonstrate that the generated adversarial camouflage patterns can effectively fool both state-of-the-art object detection models and human observers, showcasing the effectiveness of the proposed approach over traditional camouflage designs.