Generation of optical intelligent deception samples for UAVs

Abstract

Visible-light target detectors based on deep neural networks (DNNs) have been widely deployed in UAVs surveillance and low-altitude security control, posing challenges to UAVs operations. This paper employs a sample-adversarial attack method in UAVs models to address robustness issues against attacks from multiple viewpoints and at long distances through surface texture overlay and deep learning optimization. Specifically, we construct a 3D mesh model of the UAVs to reconstruct the complete structure of components such as the fuselage and wings; we initialize the adversarial texture as random noise and use a differentiable neural renderer to apply the texture to the entire surface of the UAVs model, ensuring texture conformity in non-planar areas; we utilize a segmentation network to extract a UAVs region mask, introduce a transformation function to map the rendered UAVs into the real-world scene, and generate adversarial samples that conform to the real environment; We employ a composite loss function for deep learning training, iteratively updating the textures via backpropagation of gradients until convergence. When the generated adversarial textures are applied to the drone and fed into a YOLO automatic recognition model, the detection results differ completely from those of the original drone image, thereby effectively deceiving intelligent target recognition systems in the physical world. Experiments demonstrate that this method achieves excellent intelligent deception effects in drone scenarios and can successfully evade detection by visible-light target detectors.