An Endogenous Security Defense Strategy for Adversarial Robustness in UAV Target Recognition

Unmanned Aerial Vehicles (UAVs), particularly rotorcraft, are extensively utilized in applications such as inspection and security surveillance, where the YOLO series algorithms have demonstrated excellent performance in real-time target detection. However, vision-based target recognition systems are vulnerable to adversarial example attacks. When confronted with carefully crafted adversarial examples, traditional single YOLO models suffer a drastic decline in recognition rates, posing significant security risks. This paper constructs a model library comprising diverse target detection algorithms, specifically incorporating non-YOLO series algorithms to enhance heterogeneity and reduce the impact of adversarial example transferability. This constitutes a key strategy for improving the robustness of UAV ground target recognition under adversarial scenarios. The Dynamic role in this multi-model fusion, enhancing the overall system security and…