School of Computing Graduate Theses

Permanent URI for this collection


Recent Submissions

Now showing 1 - 5 of 512
  • Item
    Cooperative Perception of Autonomous Vehicles for Semi-Connected Environment
    Hakim, Bassel; Computing; Noureldin, Aboelmagd; Hefeida, Mohamed
    Connected Autonomous Vehicles (CAV) holds great promise for improving road safety by reducing human error and lowering accident rates. However, the contemporary CAV industry requires implementing cooperative perception systems as proactive means to detect potential accidents and enable CAVs to optimize their future trajectories and decisions. Despite numerous studies, significant challenges persist due to limited communication resources and the absence of essential safety features in many CAV systems. This research addresses these challenges by developing accident-aware cooperative perception systems capable of thriving in resource-constrained network environments. To achieve this objective, three innovative solutions are presented. Firstly, a methodology for optimizing the selection of critical information within limited cellular network resources is introduced. This approach operates at the CAV’s end, where information is selected based on its significance to specific receivers, and at the base-station, which must deal with resource limitations and high message volumes. Remarkably, the base-station transmits more than 95% of message value while discarding 65% to 95% of messages in various experiments with varying available resource blocks. Additionally, this thesis proposes a Game Theory-Based Transmission algorithm (GTBT) to mitigate redundancy, offering the advantage of decentralized redundancy mitigation implementation without additional communication costs. Secondly, this research introduces two adaptive information clustering techniques, specifically Spatial safety-aware Clustering (SC) and Temporal Clustering (TC), which optimize safety-related information by adapting standard generation rules. These techniques significantly enhance safety relevance by up to 12.5X, while simultaneously reducing communication payload by up to 41% across different environments. Furthermore, this research shows that the proposed SC approach can extend existing methods, further reducing payload by more than 20%. Thirdly, this research presents a theoretical benchmark solution, named Holistic COoperative Perception solution (H-COP), a versatile tool that can guide industry leaders, researchers, and policymakers toward safer and more effective CAV systems. As a novel reference point, this study demonstrates the usefulness of H-COP for evaluating various methods and exposing additional limitations in existing standards, specifically, the high number of replicas beyond the maximum allowable senders. Notably, STC effectively eliminates these unnecessary messages, a task that state-of-the-art techniques struggle to accomplish.
  • Item
    Improving the Accessibility of Digital Games Using Partial Automation
    Cimolino, Gabriele; Computing; Graham, Nick
    Partial automation is a games accessibility technique that can improve the accessibility of digital games to novices and players with disabilities. Under partial automation, the player shares control of the game with an AI copilot that performs gameplay actions that the player has difficulty controlling. Through a series of studies, we found that this technique can extend a game's accessibility to players who might be unable to play otherwise. This thesis contributes an exploration of the design space of partial automation, including many examples of partial automation created for use in our studies, as well as empirical evidence that partial automation can improve a game's accessibility.
  • Item
    Hierarchical Deep Reinforcement Learning with Cross-Attention and Planning for Autonomous Roundabout Navigation
    Montgomery, Bennet L.; Computing; Muise, Christian; Givigi, Sidney
    Autonomous vehicle control is an important sub-field of autonomous vehicle research. Many challenges remain to improve the safety and performance of autonomous vehicle control systems in urban driving environments. One such urban driving environment is the roundabout junction, which presents its own unique challenges to potential solutions to autonomous vehicle control. This thesis proposes and tests a vehicle control agent as a candidate solution for urban roundabout navigation. The vehicle control agent is based on a hierarchical deep reinforcement learning architecture with a superior network selecting short-term lane-change behaviour and a subordinate network selecting longitudinal acceleration values. The road sequence followed by the agent is selected by a route planner based on Dijkstra’s algorithm. The proposed agent learns to navigate the roundabout environment safely, reaching the goal state in 100% of validation scenarios after training. The agent also outperforms an agent based on the Krauß-following model in 2 out of 5 tested metrics and matches the performance of the Krauß-following model in the remaining 3 metrics.
  • Item
    Combined Reinforcement Learning and Symbolic Logic for Autonomous Cyber Operations
    Kerr, Ryan H.; Computing; Ding, Steven
    Autonomous Cyber Operations (ACO) promise to help alleviate the strain posed by the shortage of cyber security talent. Training Reinforcement Learning (RL) agents to perform network penetration testing and engage in red versus blue team exercises could result in agents that are capable of defending networks 24/7 or potentially be used to train new cyber talent. However, the partially-observable, non-deterministic nature coupled with a large action space makes the network cyber domain challenging for many Artificial Intelligence (AI)-driven approaches, hindering their performance and overall utility. Current research in ACO are largely between RL and AI planning, however, the solution to addressing the aforementioned challenges may lie in their intersection. This thesis integrates the two paradigms and establishes a neuro-symbolic agent training system through an interactive symbolic logic engine. Three methods are examined for accelerating agent training, namely: action masking, plan-based reward shaping, and using the logic engine as an environment driver. The results show that action masking is highly effective at reducing the number of steps to convergence, while the logic-based environment provides a significant per-step performance improvement to speed up training. These results highlight that a hybrid approach is a viable, and perhaps even necessary, method for developing and improving ACO agents.
  • Item
    A Privacy-Preserving Analytics Pipeline for De-Identified Primary Care Data
    Pepin, Ian; Computing; Zulkernine, Farhana; Alaca, Furkan
    Data breaches in the healthcare industry are at an all-time high. The average breach in the healthcare industry reached US$10.10 million in 2022, which is highest among all industries for the 12th consecutive year [106]. Although the healthcare industry is one of the more highly regulated industries, initial attack vectors such as phishing, compromised credentials, or insider threats are at the root of many breaches today. Potential vulnerabilities to these attack vectors can be dangerous for individuals or organizations that share medical data with others. This research aims to address the challenges in the secured sharing and processing of clinical text data. The research objectives include evaluation and comparison of de-identification tools for clinical notes, and the assessment of Secure Multi-Party Computation (MPC) protocols and frameworks to perform computations on encrypted medical data. The thesis makes several contributions in the area of secured analytics of sensitive data. First, we compare the features and performance of five state-of-the-art de-identification tools for free-text clinical notes, highlighting the strengths and weaknesses of each one. Next, we propose a de-identification pipeline that removes most of the manual work associated with this type of task. Finally, we build a solution that involves MPC, specifically Secret Sharing, to allow multiple parties to jointly evaluate functions on their encrypted inputs without revealing the unencrypted data to anyone. We evaluate the performance of the framework against the same framework for the analysis of unencrypted medical data. The contributions of this thesis benefit researchers and medical professionals by demonstrating the feasibility of our proposed methods in privacy-preserving secured data analytics.