Adversarial Vehicle Camouflage via Photo-Realistic Neural Rendering

Jan 1, 2026·
Adarash Kumar
· 0 min read
Abstract
We present a method for generating adversarial vehicle camouflage that fools object detectors while maintaining photo-realistic appearance. Using a neural renderer trained on the CARLA driving simulator, we optimise vehicle textures to minimise detection confidence across diverse viewpoints and environmental conditions. Our approach achieves significant detection suppression on EfficientDet-D0, with strong transfer to YOLOv5, SSD, and Faster R-CNN, while producing patterns that resemble natural environmental effects.
Type
Publication
Submitted to IROS 2026
publications
Authors
PhD Researcher — AI for Autonomous Vehicle Testing
Adarash Kumar is a final-year PhD researcher working on AI and computer vision for self-driving car testing. His research focuses on adversarial robustness of object detection systems in autonomous driving, using simulation environments and neural rendering to evaluate and improve safety-critical perception models.