WIP: Hijacking Attacks on UAV Follow-Me Systems in Realistic Scenarios

Modern vision-based object tracking is a vital component of Unmanned Aerial Vehicle (UAV) systems. It enables advanced applications such as follow-me, which allows a drone to automatically track and follow a subject. While a wealth of research explored the vulnerabilities of object tracking algorithms, there lacks a comprehensive analysis on whether the vulnerabilities can be exploited on real UAV systems, considering challenges including physical constraints, real-world uncertainties, and limited attacker's knowledge. To bridge the knowledge gap, we design a hijacking attack that deceives the UAV follow-me mode to track a wrong subject by leveraging existing object tracking attacks. We thoroughly analyze its feasibility in real-world scenarios. With insights from the study, we are able to improve the attack success rate on the UAV follow-me application from 47% to 95% by leveraging inaccuracies of sensor measurements and instability of the gimbal camera, which indicates a realistic system exploit.