DUAL ENSEMBLE ENHANCED PHYSICAL ATTACKS ON AERIAL DETECTION

Recently, increasing attention has been devoted to adversarial attacks against deep neural networks (DNNs), and numerous adversarial attack methods have been proposed. However, existing attack approaches are typically trained using a single model, resulting in limitations in attack performance and transferability. To address these challenges, a dual ensemble enhanced physical attack (DE²PA) method is proposed, which employs background adversarial patches to implement physical attacks. Specifically, in the DE²PA, model integration masks are adopted to divide the adversarial patch into two segments. The pixels of each segment are individually optimized by dual object detection models to enable the generated adversarial patches to integrate the attack characteristics learned from different models, which contributes to enhancing the attack performance and transferability of the adversarial patches. Several proportionally scaled experiments are conducted in physical scenes to evaluate the effectiveness of the DE²PA.The experimental results demonstrate that the adversarial patches generated by DE²PA exhibit superior attack performance and transferability.