Adversarial Attacks Against AI-Powered Object Detection in Low-Altitude Defense

Deep learning-based object detection technology serves as the perceptual core of low-altitude security systems, directly impacting the reliability of critical applications such as drone inspection and aerial logistics. To evaluate the robustness of low-altitude visual systems under adversarial conditions, with the widely deployed YOLOv8 detector as research subject, a joint adversarial attack method was proposed. By co-optimizing multi-task losses, including target localization and classification, the method generated adversarial patches capable of inducing false negatives and false positives in the detector. Experiments conducted on a dataset containing various aircraft targets (MAR20) demonstrated that the proposed method achieved an average attack success rate of 78.86 against YOLOv8 series models, significantly outperforming random noise attacks. Cross-model transfer experiments further revealed that the adversarial patches maintained an average attack success rate exceeding 65 across different variants of YOLOv8, indicating strong generalization capability. These results systematically expose the adversarial vulnerabilities inherent in YOLOv8's joint optimization mechanism. From an attacker's perspective, this study empirically revealed the severe vulnerabilities present in low-altitude security visual systems under joint adversarial optimization, providing an important reference for building robustness evaluation frameworks, designing active defense strategies, and ensuring the secure deployment of critical low-altitude infrastructure.