A Pluggable Solution For Robust UAV Tracking Against Attacks

Mengjie Jia, Yanyan Li, Houbing Herbert Song, Jiawei Yuan

Abstract: unmanned aerial vehicles (UAVs) are increasingly integrated into vehicular systems for applications such as autonomous surveillance, traffic monitoring, and navigation. These applications rely on real-time object tracking, which has been significantly enhanced by deep-learning (DL)-based models. However, DL-based trackers remain highly vulnerable to adversarial attacks, where imperceptible perturbations can severely degrade tracking accuracy and reliability. To address these challenges, we propose a pluggable defense solution designed to enhance the robustness of UAV tracking systems without modifying existing tracking architectures. Our approach leverages a dual-level optimization strategy to mitigate adversarial perturbations at both feature and decision levels, ensuring resilient tracking performance. Implemented as a preprocessing stage, our solution can be seamlessly integrated with various UAV tracking systems. We evaluate our approach against multiple adversarial attacks across three widely used UAV tracking benchmarks, such as UAVTrack112, UAV123, and UAVDT. Experimental results demonstrate that our pluggable solution effectively restores tracking accuracy and improves robustness under various adversarial attacks without sacrificing tracking performance in original (attack-free) scenarios. Real-world tests on a UAV platform validate the efficiency and practicality of our method. Comprehensive results indicate our solution can strengthen UAV tracking in real-world applications and ensure reliable performance in adversarial environments.

External IDs:doi:10.1109/jiot.2025.3593838