Adversarial Learning in Unmanned Aerial Vehicles-aided Intelligent Transportation Systems

<p dir="ltr">With their inherent attributes such as mobility, flexibility, and adaptive altitude, Unmanned Aerial Vehicles (UAVs) can potentially enable Intelligent Transportation Systems (ITS) to be more efficient for data collection, data analysis, and communication networks. UAVs are equipped with machine learning and Deep Neural Network (DNN) models to enhance their capabilities for applications across various safety-critical domains, particularly traffic management in ITS. However, it has been demonstrated that these models are highly susceptible to adversarial attacks, which are defined as imperceptible changes to pixels of the actual input to statistically impact the decision of the machine learning and DNN models.</p><p dir="ltr">In this thesis, we explore the potential contributions of UAVs to ITS and examine how UAVs can enhance various aspects of ITS, such as traffic monitoring, congestion management, and emergency response. Furthermore, we analyze the pivotal role that machine learning techniques play in enabling this synergy. A thorough survey aiming to explore a quantitative understanding of widely used DNN models via a series of experiments and comparisons is presented. To investigate the influence of adversarial attacks on DNN-based aerial vehicle detection and traffic extraction from UAV videos, we propose a white-box adversarial threat model to periodically manipulate critical frames and pixels with the most impact on aerial vehicle detection and tracking. To protect UAV-based vehicle detection and tracking systems, we present adversarial detection and cleaning. To detect adversarial frames, multi-level temporal consistency analysis is utilised to capture both short-term and long-term dependencies in video sequences by detecting inconsistencies in frame predictions.</p>