A Robust UAV Tracking Solution in the Adversarial Environment

Mengjie Jia, Yanyan Li, Jiawei Yuan

Abstract: Unmanned aerial vehicles (UAVs) are ideal platforms for object tracking due to their high mobility and advanced sensing capabilities. Recent advancements in AI have enhanced UAV tracking by integrating deep learning models with UAV systems, but they also introduce security concerns due to the vulnerabilities of deep learning models to adversarial attacks. To address this challenge, we propose a new UAV tracking solution integrating a reconstruction module with an anomaly detection module to enhance the robustness of UAV tracking systems against attacks. Our reconstruction module processes video frames to mitigate adversarial impacts without compromising tracking performance on clean frames. The anomaly detection module employs a reference generator to dynamically construct adversarial reference samples for feature map comparisons to effectively detect attacks. We evaluated our solution against state-of-the-art attacks on three benchmarks. The results show that our solution improves the tracking performance under attack conditions, achieving an average precision of 97.4% and a success rate of 96.3% of the original tracking. Additionally, our method achieves a 98.9% attack detection rate with a 4.23% false positive rate in anomaly detection. The evaluation results demonstrate the effectiveness of our approach in enhancing the robustness and reliability of UAV tracking systems in the adversarial environment.