CONFIRMED: EAS Doctoral Proposal Defense by Mengjie Jia
Title: Secure and Robust UAV Tracking Systems Location: Zoom https://umassd.zoom.us/j/98095026817?pwd=3pHJ1Yt2arNEGoy5ilG45wH9uLFTNH.1 Meeting ID: 980 9502 6817 Passcode: 881612 Advisor: Dr. Jiawei Yuan, Department of Computer and Information Science Committee Members: Dr. Gokhan Kul, Department of Computer and Information Science Dr. Long Jiao, Department of Computer and Information Science Dr. Ruolin Zhou, Department of Electrical & Computer Engineering Abstract Visual object tracking is a fundamental research area in computer vision, which plays a critical role in robotic applications such as autonomous navigation and surveillance. Deep learning-based visual object trackers have significantly enhanced UAV tracking systems, but they also introduce security concerns due to the vulnerabilities of deep learning models to adversarial attacks. Traditional template-based trackers rely on annotated templates for initialization, limiting their applicability in real-time scenarios like UAV surveillance, where object descriptions are often provided in natural language. Template-free trackers address this limitation by leveraging multi-modal inputs and combining visual and semantic guidance. However, both methods remain susceptible to adversarial attacks. Template-based methods are vulnerable to visual perturbations on input frames, and the adoption of multi-modal inputs introduces new vulnerabilities to template-free methods, such as misaligned embeddings between visual and textual inputs. This research investigates the vulnerabilities of DL-based UAV tracking systems and proposes defense mechanisms to enhance their robustness. The research will be conducted in three directions: (1) enhance the robustness of template-based tracking methods to adversarial perturbations through input reconstruction (2) conduct in-depth security analysis and exploration of template-free object tracking methods, focusing on multi-modal vulnerabilities (3) design robust template-free tracking systems by integrating semantic references from textual descriptions and leveraging motion features across consecutive frames. For further information please contact Dr. Jiawei Yuan at jyuan@umassd.edu
See description for location
> See Description for contact information
https://umassd.zoom.us/j/98095026817?pwd=3pHJ1Yt2arNEGoy5ilG45wH9uLFTNH.1