Electrical and Computer Engineering

Permanent URI for this collectionhttps://uwspace.uwaterloo.ca/handle/10012/9908

This is the collection for the University of Waterloo's Department of Electrical and Computer Engineering.

Research outputs are organized by type (eg. Master Thesis, Article, Conference Paper).

Waterloo faculty, students, and staff can contact us or visit the UWSpace guide to learn more about depositing their research.

Browse

Now showing 1 - 20 of 2129

Polynomial Controllers for Optimal Trajectory Matching with Stability Guarantees
(University of Waterloo, 2025-11-04) Kitaev, Alexander
We formulate a trajectory matching problem in which a set of reference trajectories for a plant is given, and a control law that causes the plant’s trajectories to be as close as possible to the reference trajectories is desired. These trajectories might be generated by an implicit controller such as a model predictive control (MPC) algorithm or manually chosen by a user. This thesis presents a nonconvex optimization approach for solving the trajectory matching problem that generates explicit polynomial controllers. The value of this approach is that the explicit control laws it generates are simpler to implement, and can be used for stability analysis. Additionally, the method presented in this thesis guarantees local stability of the generated controller by ensuring local contractivity towards the generated trajectories. This thesis presents several theoretical results that justify the method described here. Firstly, a proof that the local contractivity constraints used to ensure local stability can be expressed as a set of matrix inequalities is presented, which turns an infinite set of constraints into a finite one. Secondly, a theorem that describes how symmetries in the trajectory matching problem correspond to symmetries in its solution is presented and proven, which enables a reduction in the control design problem size and resulting solution. Finally, this thesis demonstrates the method it describes on two example problems motivated by real-world applications. The first of these is stabilization and disturbance recovery for a single-machine infinite-bus (SMIB) power system, and the second is a lane change manoeuvre for Dubin’s vehicle, a simple vehicle model. In each case, the reference trajectories are generated by MPC.
An Empirical Study of Privacy Leakage Vulnerability in Third-Party Android Logs Libraries
(University of Waterloo, 2025-10-21) ZHAO, YIXI
Mobile logging libraries are essential tools for debugging and monitoring Android applications, yet their privacy implications remain largely unexplored. This paper presents the first large-scale empirical study of privacy risks in Android logging practices, analyzing 48,702 applications from Google Play to identify sensitive data leakage through third-party logging frameworks. Our findings reveal that while logging library adoption is limited (3.4% of applications), nearly half (49.3%) of logging-enabled applications exhibit privacy leaks, creating significant security vulnerabilities. Three dominant libraries—Timber (35.2%), SLF4J (35.1%), and Firebase (29.4%)—account for 99.7% of all verified privacy leakage instances. We identify distinct logging patterns across frameworks, with SLF4J showing balanced log level distribution, Timber concentrating heavily on DEBUG levels (78.5%), and Firebase dominated by Analytics Events (98.0%). Our analysis reveals that privacy violations predominantly stem from indirect data flows (62.5%) requiring intermediate processing steps, with most leaks occurring through moderate-complexity paths of 2-4 statements. User-info sources dominate privacy leaks (69.7%), while user-input sources represent a substantial portion (30.3%), highlighting GUI components as significant risk vectors. Longitudinal analysis of application updates demonstrates that privacy leaks tend to improve over time, indicating growing developer awareness of privacy concerns, though persistent vulnerabilities underscore the need for systematic privacy protection measures. Our study contributes the largest dataset of third-party logging-based privacy violations to date, a reproducible analysis pipeline for future research, and actionable insights for developers and library maintainers. These findings emphasize the critical need for practitioners to recognize both user information and user input as significant privacy threats when implementing third-party logging frameworks in Android applications.
Radio Resource Management of Hybrid Beamforming Systems
(University of Waterloo, 2025-10-14) Quan, Yuan
Beamforming and massive multi-input multi-output (MIMO) are two of the key technologies that enable high capacity and spectrum-efficient communications in 5G and beyond systems. Codebook-based HBF (Hybrid Beamforming) wherein ABF (Analog Beamforming) vectors are chosen from pre-designed codebooks and, optionally, DBF (Digital Beamforming) can be performed on the selected ABF vectors, results in lower hardware cost, training overheads, and complexity for real-time operations over FDBF (Full Digital Beamforming). We study RRM (Radio Resource Management) for the DL (Downlink) and the UL (uplink) of codebook-based HBF systems assuming proportional fairness. Our study focuses on the practical multi-channel case without assuming that the number of RF (Radio Frequency) chains at the BS (Base Station), K, is larger than the number of UE (User Equipment) in the cell, U. Indeed, in the highly practical yet underexplored case where K
Controlling Light with Photon Subtraction via the Single-Photon Raman Interaction
(University of Waterloo, 2025-10-14) Pasharavesh, Abdolreza
This dissertation leverages deterministic photon subtraction based on the single-photon Raman interaction (SPRINT) to engineer multiphoton quantum fields and design quantum optical platforms for applications ranging from non-Gaussian quantum light generation to photon-number-resolving (PNR) detection and photon number splitting (PNS) attacks on quantum key distribution (QKD). The work is structured into four main parts. In the first part (Chapter 2), we evaluate the performance of the subtraction scheme using system parameters that are technologically accessible according to the current state of the art. We analyze the photon subtraction process in a configuration where the transitions of a Λ-type emitter are selectively coupled to the stationary modes of a bimodal cavity, which are in turn coupled to distinct waveguide modes. Using the input-output formalism of quantum optics and quantum trajectory methods, we investigate single- and multiphoton transport in the system. The results indicate that success rates approaching unity are achievable with currently reported coupling rates for cold atoms trapped in crossed optical-fiber cavities as well as for solid-state platforms based on quantum dots. In the second part (Chapter 3), we explore the capability of the photon subtraction scheme to generate non-Gaussianity in initially Gaussian input fields. Using a photon subtractor with the emitter directly coupled to a chiral waveguide, we show that for both squeezed vacuum and coherent light input pulses, the Wigner function of the output field clearly reveals its non-Gaussian character following photon subtraction. Furthermore, we propose a measurement-based scheme on the subtracted photon which can lead to conditional generation of quantum states resembling Schrodinger’s kitten state directly from coherent input light with fidelities above 99%. This result is particularly noteworthy, as coherent pulses, unlike the squeezed vacuum inputs commonly used in previous studies, are readily available experimentally. The last two parts of the dissertation explore the possibilities arising from cascading multiple photon subtractors. In the third part (Chapter 4), we investigate the operation of a PNR detector composed of a cascade of waveguide-coupled Λ-type emitters, which deterministically demultiplexes incoming photons among single-photon detectors. We present a closed-form expression for the detector’s precision in the linear regime and predict how correlations generated by nonlinear photon-photon interactions influence this precision. We compare the performance of the proposed PNR detector with that of a conventional PNR scheme based on spatial demultiplexing via beamsplitters. Our results indicate that the proposed scheme can outperform conventional detectors under realistic conditions, making it a promising candidate for next-generation PNR detection. In the fourth part (Chapter 5), we present a specialized photon subtraction scheme that enables the deterministic extraction of single photons from multiphoton states while leaving input single-photon states unaltered. The proposed device consists of a two-way cascade of two Λ-type emitters coupled via a chiral waveguide. We analyze the interaction of this system with traveling few-photon pulses of coherent light and use these results to highlight how this two-emitter extension improves the original deterministic single-photon subtraction when it comes to implementing undetectable PNS attack on a QKD channel. Finally, in Chapter 6, we demonstrate how this two-emitter approach can be extended to an n-emitter cascade, resulting in a photon subtractor that selectively extracts photons from an input light stream based on their arrival time sequence. We show that this photon subtractor enables the generation of high-fidelity and modal purity multiphoton Fock states. The application of these Fock-state pulses in optical interferometry is investigated, highlighting their potential for quantum metrology at the Heisenberg limit. These results introduce novel applications of SPRINT-based photon subtraction in areas ranging from non-Gaussian photonics, to PNR detection, QKD, and quantum metrology.
Approaching Memorization in Large Language Models
(University of Waterloo, 2025-10-08) Cheng, Xiaoyu
Large Language Models (LLMs) risk memorizing and reproducing sensitive or proprietary information from their training data. In this thesis, we investigate the behavior and mitigation of memorization in LLMs by adopting a pipeline that combines membership inference and data extraction attacks, and we evaluate memorization across multiple models. Through systematic experiments, we analyze how memorization varies with model size, architecture, and content category. We observe memorization rates ranging from 42% to 64% across the investigated models, demonstrating that memorization remains a persistent issue, and that the existing memorization-revealing pipeline remains valid on these models. Certain content categories are more prone to memorization, and realistic usage scenarios can still trigger it. Finally, we explore knowledge distillation as a mitigation approach: distilling Llama3-8B reduces the extraction rate by approximately 20%, suggesting a viable mitigation option. This work contributes a novel dataset and a BLEU-based evaluation pipeline, providing practical insights for research on LLM memorization.
Enhancing Social Learning in Humanoid Robots Taught by Non-Expert Human Teachers
(University of Waterloo, 2025-09-30) Aliasghari, Pourya
Tools that assist with daily tasks are valuable. For example, with the aging population in Canada and worldwide, there is a growing demand for ways to help older adults perform daily activities independently. Socially intelligent robots can promote independence by assisting with routine tasks. While advanced robots may be capable of performing various specialized operations, it is not feasible for their designers to program them in advance to effectively carry out multi-step, complex tasks requiring high-level planning and coordination, `out of the box' in new environments and for users with diverse preferences. To successfully integrate into domestic environments, robots must learn new task knowledge from human users. Many of our own skills as human beings have been acquired through social learning, i.e., learning via observation of or interaction with others, throughout our lifetime. Social learning for robots enables the transfer of skills without the need for explicit programming, allowing users to teach robots via natural, intuitive, and interactive methods. This thesis targets three key challenges in the social learning of robots: enabling non-expert humans to teach robots without external help, enabling robots to learn and perform multi-step tasks, and enabling robots to identify the most suitable teachers in their social learning. The first phase of my research examines whether or not participants with no prior experience teaching a robot could become more proficient robot teachers through repeated human-robot teaching interactions. An experiment was conducted with twenty-eight participants who were asked to kinesthetically teach a Pepper robot various cleaning tasks across five repeated sessions. Analysis of the data revealed a diversity in non-experts' human-robot teaching styles in repeated interaction. Most participants significantly improved both the success rate and speed of their kinesthetic demonstrations after multiple rounds of teaching the robot. The second phase introduces a novel, biologically inspired imitation approach enabling robots to understand and perform complex tasks using high-level programs that incorporate sequential regularities between sub-goals that a robot can recognize and achieve. To learn a new task, the system processes demonstrations to identify multiple possible arrangements of sub-goals that achieve the overall task goal. For task execution, the robot determines the optimal sequence of actions by evaluating the available sequences based on user-defined criteria, through mental simulation of the real task. This learning architecture was implemented on an iCub humanoid robot, and its effectiveness was evaluated across multiple scenarios. In the third phase, I propose an attribute for identifying the most suitable teachers for a robot: human teachers’ awareness of and attention to the robot’s limitations and capabilities. I investigate the impact of this attribute on robot learning outcomes in an experiment with seventy-two participants who taught three physical tasks to an iCub humanoid robot. Teachers’ awareness of the robot’s visual limitations and learning capabilities was manipulated by offering the robot’s visual perspective and by placing participants in the robot’s position when labelling actions in demonstrations. Participants who could see the robot’s vision output paid increased attention to ensuring that task objects in their demonstrations were visible to the robot. This emphasis on attention resulted in improved learning outcomes for the robot, as indicated by lower perception error rates and higher learning scores. I also propose a metric for robots to estimate the potential for receiving high-quality demonstrations from particular human teachers. These findings demonstrate the feasibility of non-experts adapting to robot teaching through repeated exposure to human-robot teaching tasks, without formal training or external intervention, and also contribute to understanding factors in human teachers that lead to better learning outcomes for robots. Furthermore, I propose a robot learning approach that accommodates variations in human teaching styles, enabling robots to perform tasks with greater flexibility and efficiency. Together, these contributions advance the development of multifunctional and adaptable robots capable of operating autonomously and safely in human environments to assist individuals in various daily activities.
Design of Phantoms and Targeted Agents for Ultrasound Contrast Imaging of Microvessels and Fibrin Clots
(University of Waterloo, 2025-09-22) Shang Guan, Han Yue
Venous thromboembolisms (VTE) are common cardiovascular events including deep vein thrombosis and pulmonary embolism. Early diagnosis and management of VTEs can significantly improve patient quality of life and mortality. Such clinical insights made possible through medical imaging are critical for clinicians to manage patient risk and treatment efficacy. As the linchpin in point of care medical imaging, Ultrasound (US) is the workhorse modality in patient diagnosis and treatment monitoring within the clinic. However, specific detection of the molecule fibrin which is the current clinical treatment target for VTEs remains elusive for US in the clinic. In this dissertation, I designed an ultrasound contrast agent with fibrin targeting properties. The baseline contrast agent was used to test a novel in vitro flow phantom with tissue mimicking characteristics to validate a contrast specific super-resolution ultrasound algorithm. The contrast agent was further modified to target fibrin leveraging the existing fibrin targeting properties of a therapeutic agent and demonstrated to be adherent to a fibrin surface as well as enhancing the detection of fibrin surfaces when imaging with a clinical scanner. Finally, a comparison study was performed examining the fibrin targeting ligand options in ultrasound literature and to further investigate the robustness of using anti-fibrin therapeutics for pure imaging applications in more physiologically realistic flow scenarios. This work aims to raise current developments in ultrasound molecular imaging to address the existing clinical need to detect fibrin as a treatment target. Baseline MBs were shown to be versatile enablers for specialized contrast imaging applications. Moreover, the significant improvement to ultrasound signal using fibrin targeting contrast confirms that ultrasound molecular imaging can be a robust solution to improve fibrin detection. Finally, the insights gained through comparing the existing MB targeting strategies provides new directions to consider for future development of fibrin targeting contrast.
Vulnerabilities in Maximum Entropy Inverse Reinforcement Learning under Adversarial Demonstrations
(University of Waterloo, 2025-09-18) Alipanah, Arezoo
Reinforcement Learning (RL) has emerged as a powerful paradigm for solving complex sequential decision-making problems. However, its effectiveness is fundamentally dependent on the availability of a well-specified reward function, the design of which is often a significant challenge. Inverse Reinforcement Learning (IRL) offers a compelling solution to this problem by enabling an agent to infer an underlying reward function from expert demonstrations. This approach has become a cornerstone of imitation learning, allowing machines to acquire sophisticated behaviors by observing human experts. A critical assumption underpinning most IRL research is that the demonstrators, while potentially suboptimal, are acting in good faith. This thesis challenges that assumption by formally investigating a significant yet underexplored security vulnerability: the susceptibility of IRL algorithms to intentionally malicious demonstrators. We address the scenario where an adversary seeks to corrupt the learning process by strategically injecting a small number of deceptive demonstrations into a training dataset, with the goal of degrading the performance of the final deployed policy. This research formalizes the problem of adversarial demonstration attacks within the IRL framework. The adversary’s objective is to design a malicious policy that generates trajectories capable of manipulating the inferred reward function. To ensure the attack remains covert, the malicious demonstrations must be statistically similar to the genuine expert demonstrations. We introduce a similarity constraint, based on the expected feature counts of trajectories, that forces the adversarial behavior to remain within a plausible, non detectable margin of the expert’s behavior. The core of our investigation is to determine whether such a constrained, malicious policy can be systematically designed and to quantify the extent of performance degradation it can induce on a policy learned from the corrupted reward function. To address this problem, we propose a novel optimization-based framework for generating the adversarial policy. The framework models the adversary’s strategy as a constrained optimization problem over the space of state-action occupancy measures. The objective is to find a policy that minimizes the expected cumulative reward according to the true, ground-truth reward function, thereby maximizing the performance loss of the agent that will learn from it. This minimization is subject to two key sets of constraints: (1) the feature-matching similarity constraint that ensures the deceptive nature of the attack, and (2) the standard Bellman flow constraints that ensure the resulting occupancy measure corresponds to a valid policy under the environment’s dynamics. A time-varying stochastic policy is then extracted from the solution to this optimization problem, providing a concrete method for generating the malicious demonstration trajectories. The effectiveness of this framework is empirically validated through a series of controlled simulation studies targeting the widely-used Maximum Entropy (MaxEnt) IRL algorithm. Our experiments are conducted in two distinct grid-world environments: ‘CliffWorld‘, which represents a safety-critical task with significant negative rewards, and ‘Four Rooms‘, a more complex navigation environment with a larger state space. We systematically evaluate the impact of varying the fraction of injected malicious data and the strictness of the similarity constraint. The performance of our proposed adversarial method is benchmarked against both a baseline of expert-only demonstrations and a scenario where random, non-strategic noise is injected into the dataset. The results of our investigation reveal a significant vulnerability in MaxEnt IRL. We demonstrate that injecting even a small fraction of malicious demonstrations, as little as 10%, can cause a disproportionately severe degradation in the performance of the deployed policy. This performance drop is substantially greater than that caused by injecting an equivalent amount of random noise, confirming the targeted nature of our adversarial generation framework. The conclusions underscore the need for the development of robust defense mechanisms and adversarially-aware IRL algorithms to ensure the safe and reliable deployment of learning agents in real-world, high-stakes applications.
Nonconvex Trajectory Optimization using Trajectory Sensitivities: Application to Personalized Autonomous Driving
(University of Waterloo, 2025-09-15) Wu, Xiaofei
Autonomous Driving (AD) has been studied in the past decade and has been gradually deployed in everyday life. A key factor in increasing people’s level of acceptance is trust, which may be enhanced by personalized autonomous driving. One way to design personalized autonomous vehicles is by mimicking the driver’s own driving style while driving safely. Many existing works explore learning-based approaches to achieve this goal. However, the performance of these methods is highly dependent on sample efficiency, and it is usually difficult to enforce safety guarantees. To mitigate these difficulties, this thesis proposes an autonomous vehicle control framework in the form of a parameterized nonconvex trajectory optimization problem with a bilevel structure, where the upper-level models the driving style of a target driver and the lower-level performs vehicle motion planning. Therefore, the focus of this work is the formulation of this parameterized nonconvex trajectory optimization problem and its solution methods, discussed under an application scenario of personalized autonomous vehicles. The lower-level of the bilevel programming problem solves a trajectory optimization problem. The nonlinear dynamic of the vehicle model leads to challenging nonconvex trajectory optimization problems. Many existing approaches formulate them as multistage programs and rely on derivatives of each stage to obtain a local approximation at each iteration, in which case the quality of approximation when solving the optimization program has significant impact on convergence behavior. In this work, we develop a novel approach for obtaining improved local approximations when solving nonconvex trajectory optimization problems. By performing an input-to-state reformulation of system dynamics, we use trajectory sensitivities, which are derivatives of the entire system trajectory with respect to control inputs, to form local approximations. This novel approximation method, when used to solve optimization problem and to linearize the constraints, results in less approximation error than the traditional approach, while the latter has accumulating numerical errors for multi-stage planning problems. Local convergence guarantees for the proposed method are presented for nonconvex optimization problems with input-affine inequality constraints. The method is applied to generate trajectories for an autonomous vehicle that are not dynamically feasible and is extended to include a scenario with static obstacles that introduces nonconvex constraints. The upper-level of the bilevel programming problem models the driving style of the target driver by minimizing the difference between human driving data and motion planning results. The decision variables are weight factors that characterize the driving style and are used to parameterize the lower-level objective, hence affecting its planning results. We adopt a gradient-based approach to solve this problem. However, differentiability is not guaranteed given the bilevel structure and the nonconvex lower-level solution mapping, so we use subgradient "descent" to generalize gradient descent for non-differentiable functions. The quotation marks suggest the fact that subgradient methods are not necessarily monotone. Therefore, a projected subgradient update algorithm is adopted to solve the upper-level problem. When learning-based approaches may fail in rare or unseen scenarios, our proposed method with an embedded vehicle model will continue to work. In addition, the optimization framework with dynamical and safety constraints ensures driving safety. The lower-level motion planner has been simulated on a variety of reference paths and compared with the traditional sequential quadratic programming with conventional first-order Taylor approximation to outperform in approximation accuracy, allowable trust-region radius, iterations to converge, and total solver time. Furthermore, out method is less prone to failure when handling multiple obstacles with a complex reference. The upper-level problem is also simulated to solve tracking problems and obstacle avoidance problems, demonstrating its ability to mimic the driving style of a target driver.
Mining Time Series for Maximal Coverage with Matrix Profiles and Constraint Optimization
(University of Waterloo, 2025-09-12) Nagar, Neeraj
Time Series Data (TSD) is essential in many areas of modern data analysis because they reflect how different processes change over time. Understanding these data can still be challenging. One of the main challenges is that the underlying states of a system are often hidden, making it harder to interpret patterns and draw reliable conclusions. This thesis addresses the critical task of mining recurrent patterns from systems whose internal states are neither directly observable nor controllable. It introduces a novel unsupervised approach explicitly designed to maximize coverage in TSD. The research proposes a structured approach comprising three key steps: firstly, generating candidate patterns and their occurrences using an advanced Matrix Profile (MP) algorithm known for its efficiency and accuracy in detecting subtle recurrent patterns; secondly, translating these candidate patterns into a constraint-based model incorporating group constraints to enforce selection of all instances of the same pattern and exclusion constraints to prevent overlapping occurrences; and thirdly, selecting an optimal subset of core patterns using either a constraint solver to ensure optimal selection on shorter Time Series (TS), or scalable greedy heuristic methods that offer practical efficiency for larger or more complex datasets, thereby effectively balancing optimality with computational feasibility. The effectiveness of the proposed approach is demonstrated through rigorous evaluations on real-world power consumption TSDs representing computer activities, alongside controlled synthetic datasets, using metrics such as coverage efficiency, computational time, and pattern compactness—striving to maximize data representation with minimal redundancy. Our experimental results show that the proposed method improves how efficiently data is covered, striking a practical balance between capturing key patterns and avoiding unnecessary repetition. This work contributes to the advancement in unsupervised pattern mining and has useful applications in areas such as forecasting, system health monitoring, anomaly detection, and policy verification.
Convex Reparameterizations for Efficient Mixed H2/Hinf Feedback Control
(University of Waterloo, 2025-09-12) Fang, Zhong
The design of controllers with mixed H2/Hinf cost functions remains a challenging problem in control theory, with pervasive applications across diverse engineering fields. The main difficulties arise from nonconvexity and infinite dimensionality of the associated optimization problem for the design. Recently, several new approaches were developed to tackle nonconvexity by reparameterizing the variables to transform the optimization into a convex but infinite-dimensional formulation incorporating additional affine constraints in the design problem. For state feedback design, system level synthesis is focused since input output parameterization is primarily intended for output feedback. To make the problem tractable, and to address limitations of historical approximation methods, a new Galerkin-type method for finite-dimensional approximations of transfer functions in Hardy space with a selection of simple poles was recently developed. However, prior applications of this simple pole approximation resulted in a design problem that required an additional approximation of a finite time horizon to compute H2 and Hinf norms for the closed-loop response. This finite horizon resulted in increased suboptimality, degraded performance, and increased problem size and memory requirements. To address these limitations, this thesis presents a novel control design framework that combines the frequency domain convex reparameterization affine constraints with a state space formulation of the H2 and Hinf norms using linear matrix inequality. This state space formulation eliminates the need for a finite time horizon approximation, and results in a convex and tractable semidefinite program for the control design. Suboptimality bounds are provided for the method which guarantee convergence to the global optimum of the infinite dimensional problem as the number of poles approaches infinity with a convergence rate that depends on the geometry of the pole selection. The recently developed convex reparameterization methods have been challenging to adapt to continuous time control design in practice, because they typically rely on finite dimensional approximations for tractability that lead to numerical ill-conditioning or even closed-loop instability. In this work, the hybrid state space and frequency domain control design method is adapted to develop the first practical and tractable continuous time control design based on these convex reparameterizations that does not suffer from ill-conditioning and that guarantees closed-loop stability for stabilizable plants. Approximation error bounds are established for the first time for the simple pole approximation in continuous time. These bound the error based on the geometry of the pole selection, and show that this error goes to zero as the number of poles approaches infinity. These bounds are particularly challenging to obtain compared to the discrete time case due to the noncompactness of the domain of integration for computing the H2 and Hinf norms in continuous time. These approximation error bounds are then used to develop suboptimality guarantees of an analogous nature to those in discrete time. This is the first time that suboptimality bounds with zero asymptotic error have been developed for a control design method using these recent convex reparameterization approaches in continuous time. Again, the noncompactness represents a major challenge that must be overcome to establish these results. There exist several recently developed convex reparameterizations for output feedback control design in discrete time (including system level synthesis and input output parameterization). However, all of these methods currently lack rigorous suboptimality guarantees that establish convergence to the solution of the infinite dimensional problem as the approximation dimension approaches infinity. This is largely due to the additional complexity introduced by the output feedback case compared to state feedback. This work develops novel output feedback control design methods using four different convex reparameterization approaches, each with different benefits and trade-offs, that result in convex and tractable control design formulations. Moreover, a single unified approximation theory is developed that simultaneously establishes suboptimality bounds for all four methods that recovers analogous results to the state feedback setting. In particular, they show a convergence rate to the global optimum that depends on the geometry of the pole selection in a similar fashion to the state feedback case. The novel methods are applied to design controllers for power converter interfaced devices to provide frequency and voltage regulation to the power grid. Practical case studies demonstrate the ability of the methods to match desired dynamic behavior for these services. We consider multi-controller scenarios involving distributed energy resources where multiple power converters must coordinate to provide grid services while respecting physical and engineering constraints, including state, input, and output limits for each device, enabling coordinated control of dynamic virtual power plants. Case studies involving power converter voltage and frequency regulation, as well as multi-controller coordination in the IEEE 9-bus system, demonstrate superior performance.
Planning Under Uncertainty: Informative and Stochastic Path Planning via Gaussian Processes
(University of Waterloo, 2025-09-09) Iskandar, Avraiem
Autonomous systems operating in real-world environments often face uncertainty due to incomplete or noisy information about their surroundings. Effective planning under such conditions requires models that can represent uncertainty and support informed decision-making. This thesis addresses planning under uncertainty in environments characterized by spatially correlated stochastic processes. We model the agent's belief over these environments using Gaussian Processes (GPs), enabling Bayesian inference from sparse observations by capturing spatial correlations. We study two problems. The first is informative path planning (IPP) in continuous space, where an agent collects measurements along a trajectory to reduce uncertainty about an unknown spatial field, subject to budget and obstacle constraints. We propose a hierarchical framework that integrates global graph-based planning with local continuous trajectory optimization to improve scalability and performance. The second problem is stochastic path planning in random fields (SPPRF), where an agent seeks to minimize expected traversal cost in an environment with uncertainty induced by a stochastic process. The agent maintains a GP-based belief over the cost field and adapts its plan through belief updates as new data becomes available. Our solutions to these problems combine methods and tools from stochastic processes, path planning, and trajectory optimization. In summary, the contributions of this thesis extend the planning capabilities of autonomous agents in environments governed by spatially correlated uncertainty, encompassing both informative exploration and uncertainty-aware planning.
Learning Human-Aware Strategies for Legible and Predictable Robot Navigation
(University of Waterloo, 2025-09-08) Moskalenko, Anna
A challenge in human-robot interaction, particularly in dynamic and crowded environments, is to design navigation that is both legible and predictable. For people to feel comfortable in the same environment as robots, their movements must be legible - ensuring quick understanding of the robot’s intentions - and predictable, meaning they align with human expectations. In this thesis, we introduce a learning-based navigation system that leverages a vector reward function to capture the dual objectives of legibility and predictability. Rather than relying on manually designed transitioning rules or fixed weighting parameters such as alpha and beta, the reward function is learned from expert demonstrations generated by a planner (LPSNav) that blends between these objectives using a continuous parameter. Our approach does not attempt to replicate the planner’s handcrafted logic, but instead generalizes the emergent patterns in its trajectories through a supervised learning framework inspired by the structure of maximum entropy IRL. The proposed architecture includes: an LSTM-based Robot Motion History Encoder, a CNN-based Environment Encoder, an LSTM-based Human History Encoder, and a two-channel Reward Analysis Engine. The system was evaluated both in simulation using the nuScenes dataset and in real-world trials using the Clearpath Jackal robot. In the user study, participants interacted with the robot in predefined navigation scenarios, and their feedback was used to assess the perceived clarity and predictability of the robot’s actions. Results show that our method generates more human-like trajectories and outperforms baseline models in scenarios requiring socially acceptable motion, such as intersections, sidestepping, and close-proximity passing. We provide both quantitative metrics (e.g., ADE and FDE) and qualitative visualizations demonstrating smooth and socially-aware navigation behavior. This work represents a step toward safer and more intuitive human-robot coexistence, offering a practical solution for real-world robotic deployment.
Modeling and Experimental Studies of Electrolyte for Zinc Battery Systems
(University of Waterloo, 2025-09-08) MASHAYEKHI, ALIREZA
Zinc-based batteries have emerged as promising alternatives to lithium-ion systems due to their low cost, inherent safety, and sustainability. However, challenges in electrolyte stability, dendrite formation, and limited lifetime have constrained their practical deployment. Addressing these barriers requires both experimental innovation and data-driven approaches for electrolyte and materials design. This thesis aims to advance zinc battery technologies through four interconnected directions: (i) development of novel electrolyte systems, (ii) optimization of zinc deposition and cycling performance, (iii) machine learning methods for early-stage lifetime prediction, and (iv) accelerated discovery of functional ionic liquids. Molten salt-derived Zn(TFSI)2 electrolytes were investigated for zinc–oxygen batteries across a wide temperature range. Electrochemical, spectroscopic, and microscopy analyses revealed the structural evolution of zinc oxide nanosheets during cycling and highlighted that Zn(TFSI)2·8H2O suppresses parasitic reactions more effectively than Zn(TFSI)2·18H2O, enabling improved reversibility and energy storage potential. Complementary studies on hydrated ZnCl2 electrolytes demonstrated how temperature and hydration level influence zinc nucleation, morphology, and cycling stability. ZnCl2·10H2O achieved 99.2% coulombic efficiency over 50 cycles, while higher operating temperatures increased discharge capacity from 17 mAh g-1 at -10 °C to 72 mAh g-1 at 40 °C. Artificial intelligence approaches were developed to classify and predict battery lifetime from early cycling data. Machine-learned models achieved up to 96% accuracy after only two cycles and 98% with additional data, while human-selected electrochemical features showed strong generalizability across chemistries. Deep learning methods reached 99.5% accuracy with extended cycling data but proved less transferable to systems with distinct degradation profiles. In parallel, convolutional and generative adversarial neural networks were applied to accelerate the discovery of ionic liquids for zinc batteries. These models improved property prediction and successfully generated new candidate electrolytes with enhanced performance at room temperature. Overall, this work provides an integrated experimental–computational framework for electrolyte optimization, electrochemical performance improvement, and AI-driven materials discovery. The findings pave the way toward more durable, efficient, and scalable zinc-based energy storage technologies.
Enhancing Power Fuzzing: Synthetic Side-Channel Data Generation, Optimal Sampling, and Noise Mitigation
(University of Waterloo, 2025-09-02) Vakulenko, Olha
Embedded systems increasingly dominate critical applications, driving the need for advanced testing and validation methodologies capable of uncovering hidden or undocumented behaviours. Traditional fuzzing approaches, which rely on observable outputs or system crashes, often fail to reveal the internal operations of embedded devices. Powertrace-based fuzzing provides a non-intrusive alternative by analysing a device’s power consumption during operation. Achieving robust and reliable fuzzing performance requires researchers to overcome significant challenges in signal acquisition, noise mitigation, and classification reliability. This thesis addresses these challenges by introducing several key improvements to the PowerFuzzer framework. First, it develops SigVarGen, a modular synthetic signal generation framework that produces realistic idle-state and active signals under controlled noise, drift, and timing variations. SigVarGen enables comprehensive algorithm development and stress testing across diverse simulated conditions, bridging the theoretical model design and empirical validation gap. Second, it presents SR\&OS, a dynamic calibration algorithm that optimises sampling rate and trigger offset selection. SR\&OS leverages adaptive binary search and statistical response detection to capture meaningful system responses despite variable latencies and noise conditions. The thesis also performs a detailed risk assessment of typical noise sources in side-channel measurements and ranks mitigation strategies based on their effectiveness and practical feasibility. It identifies practical denoising techniques, such as trace averaging, singular spectrum analysis, and independent component analysis, as effective methods for improving signal quality. Furthermore, it evaluates signal quality metrics and validates comparative power and correlation-based indicators as efficient predictors for adaptive acquisition termination. Together, these developments create a more robust and scalable framework for detecting undocumented behaviours in embedded systems through powertrace analysis. Experimental validation using synthetic datasets and real-world embedded targets demonstrates improvements in calibration accuracy and acquisition efficiency. The findings lay a foundation for future advancements in hardware fuzzing frameworks, mainly targeting embedded environments.
Resource Allocation and Task Scheduling for Integrated Sensing and Communications
(University of Waterloo, 2025-08-27) Hu, Shisheng
Integrated Sensing and Communications (ISAC) has emerged as a promising paradigm for future Sixth-Generation (6G) wireless networks. In this paradigm, wireless networks can have both Sensing and Communication (SAC) capabilities using shared network resources. ISAC not only enables the provisioning of SAC services but also has the potential to enhance their performance: end-user devices can offload sensing data collection tasks to Access Points (APs) co-located with edge servers and upload raw or preprocessed sensing data to powerful edge servers for high-performance processing; meanwhile, APs can leverage contextual information about communication tasks obtained through sensing, such as the visibility of virtual content in mobile Augmented Reality (AR) streaming, to enhance communication efficiency. The interesting issue is to efficiently utilize network resources to optimize SAC service performance in the presence of high and spatiotemporally varying service demands. However, the main technical challenges are: 1) how network resource allocation and SAC task scheduling are proactively determined to enable efficient coordination between APs and mobile end-user devices for achieving satisfactory service performance; 2) how an end-user device adaptively offloads computation-intensive Deep Neural Network (DNN)-based sensing tasks to an edge server to optimize task performance, under dynamic task arrival, task processing, and server workload statuses; and 3) how an AP efficiently acquires contextual information about communication tasks through sensing individual mobile AR users and dynamic environment for resource-efficient mobile AR streaming. In this thesis, we develop efficient resource management schemes for ISAC, including resource allocation and task scheduling, to address the above three technical challenges. First, we investigate proactive resource management for ISAC, determining the reservation of radio and computing resources, the active probability of mobile devices for communications, and the partitioning of sensing regions. To cope with the non-stationary spatial distributions of mobile devices and sensing targets, which can result in the drift in modeling the distributions and ineffective resource management decisions, we construct Digital Twins (DTs) of the network slices for individual SAC services. In each DT, a drift-adaptive DNN-empowered statistical model and an emulation function are developed for the spatial distributions in the corresponding slice, which facilitates closed-form decision-making and efficient validation of a resource management decision, respectively. Numerical results demonstrate that the proposed scheme can significantly enhance service satisfaction ratios and reduce resource consumption compared to benchmark schemes. Second, we investigate task offloading for DNN-based sensing data processing. Particularly, we consider that an end-user device stochastically generates and adaptively offloads DNN-based sensing data processing tasks to an AP co-located with an edge server. To adapt to the dynamic on-device and edge server workload status, leveraging the multi-layer and multi-exit architecture of the considered DNNs, an offloading decision for each sensing task is made on whether and when to stop on-device task processing and offload the task to the edge server to complete the processing. Two DTs are constructed to evaluate all potential offloading decisions for each sensing task, which provides augmented training data for a machine learning-assisted decision-making algorithm, and to estimate the task processing status at the device, which avoids frequently fetching the status information from the device and thus reduces the signaling overhead. Simulation results demonstrate the outstanding performance of the proposed task offloading scheme in terms of balancing sensing result accuracy, delay, and energy consumption. Third, we propose an efficient resource allocation scheme in sensing-assisted mobile AR streaming. In specific, we consider that the position and surrounding environment of an AR user can be captured via sensing to extract contextual information for AR streaming, i.e., the visibility of virtual content. The goal is to minimize the overall radio resource consumption for delivering virtual content visible to the AR user by properly determining the radio resource allocation for user positioning and environment mapping. To this end, we first develop a mathematical model to estimate the content visibility uncertainty and the content delivery resource consumption. We then generate a reference resource allocation decision that guides a deep reinforcement learning-based decision process to efficiently adapt to non-stationary user and environment dynamics. Trace-driven simulations demonstrate that the proposed scheme significantly reduces radio resource consumption for delivering virtual content visible to an individual AR user, compared to benchmark schemes. In summary, we have proposed efficient resource management schemes for ISAC that optimize SAC service performance with efficient resource utilization and practical operational complexity. The research results from the thesis provide valuable insights into the design of scalable and adaptive ISAC systems that seamlessly unify sensing, communications, and intelligence in the future 6G.
Defect characterization, advanced etching techniques, and large field-of-view imaging for metalenses
(University of Waterloo, 2025-08-25) Zhu, Chenxu
Metalenses, made of nanostructures, have attracted significant interest for their ability to overcome the limitations of traditional lenses. This thesis provides a thorough study on the design, fabrication, and characterization of high-performance metalenses using metasurface technology. It underscores the interactions among fabrication imperfections, advanced nanofabrication techniques, and imaging performance across a large field of view (FOV). Defects in metalenses introduced during fabrication can significantly affect their optical performance. Chapter 3 carried out a unique experimental analysis using real-fabricated devices to investigate the impact of different defects, including inclined sidewall angle, uniform critical dimension bias, non-uniform bias due to the proximity effect, and the notching effect in etching, on parameters such as focal spot intensity, focal length, and Strehl ratio (SR). Our research indicates a 47% decrease in intensity when the sidewall is positively tapered by 4˚, a 15% intensity reduction with a 30 nm deviation in structure size, minimal intensity and focusing quality deterioration with the non-uniform bias, and a 13% intensity decrease along with a significant drop in focusing quality when the notching effect occurs. Our study provides critical insights into the tolerances required for optimal metalens performance and offers recommendations for time and cost savings. Chapter 4 introduces an innovative etching technique that utilizes a three-step C4F8/SF6 plasma etching process with varying gas ratios at different depths. By maintaining the plasma after each step, this continuous three-step process offers enhanced flexibility for tuning the etching of high aspect ratio (HAR) nanostructures, resulting in smooth and vertical profiles. By using the optimal gas ratio, metalens nanostructures with diameters of 71 nm and heights of 1 μm were successfully created, with feature size variation kept to less than 10 nm. This proposed continuous multi-step approach significantly improves the controllability of silicon nanopillar etching, which is crucial for achieving precise phase control and effectively manipulating light in metalens applications. Building on these advancements, Chapter 5 details the design and fabrication of a single-layer metalens that achieves a broad FOV of 120° at a near-infrared wavelength of 940 nm. The metalens features a simple structure consisting of a metallic aperture paired with a single-layer metasurface. To optimize the phase profile of the metalens, a ray-tracing algorithm was utilized. The resulting enhanced metalens demonstrates commendable focusing efficiency (35-61%) and diffraction-limited imaging (SR > 0.8), positioning it as a promising candidate for applications in imaging, sensing, and integrated photonic systems.
Towards Secure and Scalable Blockchain Systems: From Game-Theoretic Oracle Networks to AI-driven Smart Contract Repair
(University of Waterloo, 2025-08-25) Nassirzadeh, Behkish
The adoption of blockchain technologies in security-critical and high-throughput domains remains limited by persistent challenges in scalability, reliability, and automated vulnerability mitigation. This thesis presents a cohesive body of work that addresses two fundamental limitations of modern blockchain systems: the difficulty of ensuring safe, efficient execution in smart contracts and the lack of robust mechanisms for secure data connectivity through decentralized oracle networks (DONs). To address the first challenge, we introduce a suite of tools, GasGauge, GasGuard, and GasGaugeAI, that advance the detection, analysis, and automated repair of gas-related Denial-of-Service (DoS) vulnerabilities in Ethereum smart contracts. GasGauge leverages static-dynamic analysis to model safe loop bounds and identify Out-of-Gas (OOG) risks. We examine how emerging AI methods, particularly large language models (LLMs) and program synthesis tools, provide a scalable path forward for developing self-healing blockchain systems. GasGuard builds on this foundation by integrating a fine-tuned LLM to insert guard conditions that prevent unsafe execution automatically. Finally, GasGaugeAI extends the pipeline with a novel multi-LLM framework that classifies gas-dependent vulnerabilities, generates Foundry-based test cases, synthesizes function-level repairs, and validates fixes iteratively. Across hundreds of real-world contracts, these systems demonstrate the potential of AI-guided repair to drastically reduce manual auditing efforts and prevent exploitable gas exhaustion patterns. Beyond contract-level vulnerabilities, this thesis tackles the broader problem of trustworthy data connectivity in decentralized applications. We propose CountChain, a game-theoretic decentralized oracle network for secure aggregation in counting systems. Built on this foundation, AdChain applies DON principles to online advertising, mitigating discrepancy fraud through incentive-aligned protocols. Our experiments demonstrate that CountChain and AdChain provide both scalability and provable security against rational adversaries. Together, the tools, systems, and theoretical insights presented in this thesis contribute to the vision of blockchain infrastructures that are both secure and scalable by design, bridging the gap between automated repair and game-theoretic connectivity.
Detecting Unchecked Exception-Related Behavioural Breaking Changes with UnCheckGuard
(University of Waterloo, 2025-08-25) Sharma, Vinayak
The ubiquitous use of third-party libraries in software development has enabled devel- opers to quickly add new functionality to their client software. Unfortunately, library usage also carries a cost in terms of software maintenance: library upgrades may include breaking changes, in which client expectations about library behaviour are no longer met in new library versions. Behavioural breaking changes can be particularly insidious, and in their full generality, could require sophisticated program analysis techniques to (approximately) detect. In this work, we present our UnCheckGuard tool, which detects a class of behavioural breaking changes—those related to exceptions thrown by Java libraries. UnCheckGuard analyzes both sides of the library/client duet. On the library side, UnCheckGuard creates a list of new exceptions that may be thrown by methods in a library’s public API, includ- ing by its transitive callees. On the client side, UnCheckGuard identifies client methods that call library methods with new exceptions. To reduce false positives, UnCheckGuard additionally filters out new exceptions that cannot be triggered by particular clients, using taint analysis. It therefore can be used by client developers as a tool to screen library updates for relevant incompatibilities. We have evaluated UnCheckGuard on 302 libraries and 352 library-client pairs drawn from the DUETS collection and found 120 libraries with newly-added exceptions, as well as 1708 callsites to library methods which, when upgraded to the latest version, may introduce a behavioural breaking change in the client due to a newly added unchecked exception. These findings highlight the practical value of UnCheckGuard in identifying exception-related incompatibilities introduced by library upgrades.
Semantically Consistent Alignment for Novel Object Discovery in Open-Vocabulary 3D Object Detection
(University of Waterloo, 2025-08-22) Chow, Adrian
3D object detection is a fundamental task in the autonomous driving perception pipeline, where identifying and localizing objects within the surrounding environment is critical for safe and robust decision-making. However, traditional 3D object detectors are limited by their reliance on a closed set of training categories, rendering them incapable of recognizing novel or out-of-distribution objects encountered in open-world driving scenarios. To address this limitation, the field of open-vocabulary (OV)-3D object detection has emerged, aiming to generalize beyond predefined label sets by leveraging vision-language models (VLMs) to align 3D object proposals with semantically rich 2D language-informed features. Despite promising results, a major challenge in OV-3D object detection lies in achieving robust cross-modal alignment between 3D and 2D features, which is often compromised by noisy annotations, occlusions, and resolution inconsistencies that disrupt semantic coherence. In this thesis, we present OV-SCAN, a novel framework for Open-Vocabulary 3D object detection that enforces Semantically Consistent Alignment for Novel object discovery. OV-SCAN introduces a two-stage strategy: (1) discovering precise 3D annotations for novel objects using vision-language supervision, and (2) filtering out semantically inconsistent or low-quality 3D–2D training pairs that arise from annotation errors and sensor limitations. We validate the effectiveness of OV-SCAN through comprehensive experiments on autonomous driving benchmarks, where our framework consistently outperforms existing methods in the OV-3D object detection task. Overall, OV-SCAN underscores the critical role of semantic consistency in cross-modal alignment and demonstrates its potential as a scalable solution for discovering and localizing novel objects in real-world autonomous driving scenarios.

Browse

Recent Submissions