Thesis/Project Final Defense Schedule
Join us as the School of STEM master’s degree candidates present their culminating thesis and project work. The schedule is updated throughout the quarter, check back for new defenses.
View previous quarter schedules
Select a master’s program to navigate to candidates:
Master of Science in Computer Science & Software Engineering
AUTUMN 2024
Monday, October 14
YANG YU
Chair: Dr. Yang Peng
Candidate: Master of Science in Computer Science & Software Engineering
9:00 A.M.; Join Yang Yu’s Online Defense
Project: Design and Development of a Scalable Communication Plane for Efficient TinyML Model Deployment on IoT Devices
This project focuses on designing and developing an efficient model deployment service for TinyML on IoT devices. The system addresses the critical challenges of scalability, end-to-end communication, and model deployment in resource-constrained environments. The proposed solution provides flexibility for various IoT use cases by incorporating both device-as-client and device-as-server operation modes. The device-as-client mode allows edge devices to check with a central server for new models periodically. In contrast, the device-as-server mode enables real-time model deployment via a lightweight HTTP server running on the edge devices.
A key feature of the system is the Over-The-Air (OTA) download update mechanism, which allows remote models to be updated, thus reducing the need for physical device access. The combination of HTTP server processing of requests and the integration of heartbeat monitoring and Redis-based data management enhances the system’s robustness. It ensures efficient communication between devices and servers. Performance evaluations have shown that the system can handle large-scale deployments while maintaining low-latency, energy-efficient operation. Future work may include further optimizations for scalability, model versioning, and enhanced security protocols to better support more extensive and complex IoT systems.
Friday, October 25
SHENG WANG
Chair: Dr. Yang Peng
Candidate: Master of Science in Computer Science & Software Engineering
1:00 P.M.; Join Sheng Wang’s Online Defense
Project: TinyML Deployment Optimizer: A Management System Solution for IoT and Embedded Ecosystems
With the rapid advancement of machine learning and the widespread adoption of Internet of Things (IoT) devices, deploying artificial intelligence on resource-constrained devices has become an inevitable topic. This paper addresses the critical challenge of efficiently deploying and managing machine learning models on small, resource-constrained Internet of Things and embedded devices, known as Tiny Machine Learning. It presents a novel, comprehensive management system streamlining the deployment process, ensuring compatibility between models and devices in resource-limited environments. It introduces a comprehensive platform that integrates model conversion, compatibility checking, and deployment tracking, brings tools for converting standard machine learning models to formats suitable for embedded system devices, and prepares device firmware for over-the-air updates. The platform’s architecture, built on a microservices framework and utilizing in-memory data storage, ensures scalability and real-time responsiveness. The system demonstrates improvements in deployment efficiency and management capabilities throughout the rigid simulations. It shows significant success in model-device compatibility matching, considerably reducing model deployment time compared to manual methods. The results demonstrate the robustness of the methodology in tackling critical challenges related to the deployment and management of TinyML systems. This work contributes to TinyML by providing a scalable, user-friendly solution that addresses the complexities of managing machine learning deployments on resource-constrained devices, paving the way for widespread adoption of TinyML and enabling applications ranging from smart home devices to advanced industrial sensors.
Wednesday, November 27
JACOB CHESNUT
Chair: Dr. Kelvin Sung
Candidate: Master of Science in Computer Science & Software Engineering
3:30 P.M.; Join Jacob Chesnut’s Online Defense
Project: Achieving Real Time Ray Traced Rendering for a Virtual Reality Headset Using Foveated Rendering
Virtual reality (VR) is a system which supports immersive observation and interaction of 3D virtual scenes. Ray tracing is a rendering method which is capable of producing high-quality images of 3D virtual scenes. Ray tracing in VR allows realistic lighting effects to be combined with immersive 3D. However, rendering to VR hardware using a ray tracing renderer is an expensive task which can be especially challenging to meet the performance requirements for real-time rendering. This is because of the high-resolution screens of most VR headsets and the high cost per-pixel computations of ray tracing.
Foveated rendering, using the user focusing position to inform the distribution of rendering resources across the screen, is a technique which can reduce the cost of rendering and allow the use of both VR and ray tracing. This method has its own challenges including the need for eye-tracking hardware, the need to balance quality drop-off and performance, and the need to adjust to each user. Modern VR systems have eye tracking capabilities, and as such are a good choice for supporting foveated rendering systems.
This project explores the use of adaptive resolution foveated rendering in order to provide real-time ray tracing for VR systems. To better address differences between users, my system supports the use of user calibration and postprocessing effects to overcome common quality issues resulting from foveated rendering. The final system is capable of real-time ray tracing of a moderately complex environment to a VR device. The final image is of high quality with virtually unnoticeable effects from foveated rendering. Future work includes improvements to handling of dynamic objects and fine tuning of runtime performance by balancing CPU and GPU processing.
Monday, December 2
HAIHAN JIANG
Chair: Dr. Annuska Zolyomi
Candidate: Master of Science in Computer Science & Software Engineering
1:15 P.M.; Join Haihan Jiang’s Online Defense
Project: Development and Evaluation of a User-Centric Online Pediatric Portal for the Developmental Disability Community
This project centers on the redesign of the Developmental and Behavioral Pediatrics (DBP) online portal, aiming to address the limited accessibility and user engagement of its current static design. Guided by principles of user-centered design (UCD) and human-computer interaction (HCI), the project seeks to create a dynamic, responsive web platform tailored to the diverse needs of clinicians, educators, and families. Key objectives include improving navigation, enhancing accessibility, and introducing features like personalized content and advanced search functionality.
The redesigned portal will leverage modern web technologies, such as React, to transform static resources into an interactive and intuitive platform. The project integrates iterative stakeholder feedback to refine its architecture and user interface, emphasizing content organization, usability, and stakeholder satisfaction. It also aspires to set benchmarks in pediatric health information dissemination, balancing technical excellence with user engagement.
Deliverables include a responsive portal design, user role-based access paths, and robust administrative features, ensuring content management and sustainability. By evaluating the portal’s impact on user trust, accessibility, and information retrieval efficiency, this project contributes significantly to advancing digital resources for the neurodevelopmental disability community.
Wednesday, December 4
SIRI CHANDANA PRIYA BADDULA
Chair: Dr. Brent Lagesse
Candidate: Master of Science in Computer Science & Software Engineering
8:45 A.M.; Join Siri Chandana Priya Baddula’s Online Defense
Project: Enhancing Hidden Device Localization and Room Layout Estimation
The widespread adoption of IoT devices has brought significant privacy concerns, particularly with hidden streaming devices misused in private spaces such as homes, Airbnb rentals, and hotel rooms. Traditional localization methods require time-intensive physical traversal and network traffic analysis, often taking 5 to 30 minutes. This work introduces an innovative system that eliminates the need for physical traversal, leveraging LiDAR technology to capture environmental dimensions and density, combined with Channel State Information (CSI) for enhanced accuracy. Processed through a Transformer Neural Network (TNN), the system achieves an average localization precision of around 70% with a reduced localization time of just 30 seconds. Tested across varied environments, the system demonstrates adaptability and precision, though challenges remain in dense spaces. This novel approach sets a new benchmark for efficient, accessible, and privacy-conscious localization of hidden devices, offering a significant advancement in safeguarding personal spaces.
SRAVYA KODI
Chair: Dr. Yang Peng
Candidate: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Join Sravya Kodi’s Online Defense
Project: Latency Reduction for Data Processing in Vehicular Edge Computing using Reverse Offloading
Vehicular Edge Computing (VEC) has emerged as a transformative paradigm to address the growing demand for low-latency and high-throughput data processing in connected and autonomous vehicular networks. As vehicular applications become increasingly sophisticated, efficient data processing at the network edge becomes imperative. This project implements a reverse offloading-based scheme to reduce data processing latency in VEC environments. The proposed scheme strategically redistributes computational tasks from roadside units (RSU) to nearby vehicles with sufficient computing power and capacity. By intelligently offloading processing tasks closer to the data source, this approach avoids the latency caused by RSU overloading, and it further reduces the communication delay associated with sending data to remote cloud servers during RSU overloading time. To validate the effectiveness of reverse offloading, we have conducted extensive simulations in various scenarios, considering factors such as communication overhead, vehicle resource availability, and varying network conditions. The obtained results indicate a significant reduction in end-to-end latency, showcasing the practical viability of the proposed approach. This project contributes to the ongoing efforts in optimizing data processing in VEC environments, paving the way for enhanced real-time applications and services in connected vehicular ecosystems.
Thursday, December 5
LUO LENG
Chair: Dr. Munehiro Fukuda
Candidate: Master of Science in Computer Science & Software Engineering
1:15 PM.; Join Luo Leng’s Online Defense
Project: An Incremental Enhancement of MASS GUI
This research enhances the Multi-Agent Spatial Simulation (MASS) framework’s application usability and development workflow, addressing the challenge of fragmented workflow in distributed computing simulations. The project transforms multiple separate interfaces—InMASS for simulation execution, Cytoscape for visualization, and a web-based monitoring interface—into a unified platform, reducing operational complexity by consolidating three separate tools into one integrated system. Key contributions include enhanced agent visualization in quadtree-based 2D space, implementation of octree data structure for 3D space simulation, and development of interactive MASS simulation capabilities within Cytoscape. The integration leverages Cytoscape’s OSGi framework and JShell’s REPL features to encapsulate web browser and CLI functionalities as plugins, enabling seamless interaction between components while reducing development complexity and learning curve. Notable technical implementations include comprehensive agent tracking functionality, SSH-tunneled remote access capabilities, and real-time cluster monitoring features. Compared to existing open-source agent-based modeling systems, our enhanced MASS framework offers distributed computing capabilities, while maintaining an intuitive, integrated development environment. The results show that time spent on common simulation tasks (modifying agent behaviors, adjusting input, and configuring output) was reduced by 60% while enhancing user experience in developing and monitoring distributed agent-based simulations.
Friday, December 6
TYSON HEO
Chair: Dr. Erika Parsons
Candidate: Master of Science in Computer Science & Software Engineering
1:15 P.M.; Join Tyson Heo’s Online Defense
Project: EMG Dance: Proof of Concept VR Rhythm Game for Phantom Limb Pain Rehabilitation
There are various challenges for patients undergoing Physical Therapy (PT) that may result in longer recovery time, for instance, lack of patient adherence, access limitations, geographical location, and financial barriers. These issues can become particularly problematic when PT exercises are monotonous and patients feel bored and/or are not encouraged to complete the exercise routines. The focus of our work targets lower-limb amputation patients, and it aims to be part of a larger initiative to involve the use of new technologies for the gamification of PT, which can potentially help improve the recovery of patients suffering from neurological conditions. The larger umbrella project is dubbed the “Smart Neurorehab Ecosystem” and it’s an ongoing effort involving the University of Washington (UW) Bothell CSSE, UW Seattle Neuroscience, and Rehabilitation Medicine at Harborview Medical Center (UWHM).
This capstone project aims to develop a gaming platform that allows for the creation of strategies to help health professionals start to address some of the aforementioned issues. At the same time, this software platform will allow for the collection of data that can be used to understand: 1) how to improve the game design to better serve amputee patients and 2) and how to track changes that could reflect a patient’s progress in their rehabilitation. This game is first being tested through the involvement of healthy subjects, as their usage can be used to debug the software and evaluate functionality that would be relevant in a real patient setting. This information can be used in the future to better tailor such a game to amputee patients.
This capstone introduces Electromyography (EMG) Dance, a prototype Phantom Limb Pain (PLP) rehabilitation game that is designed to alleviate the effects of below-the-knee PLP. The game proposes to use at-home technologies such as portable EMG sensors and Virtual Reality (VR) to create a rhythm game that could be used to motivate patients to complete their PT exercises through daily “dancing” sessions from the comfort of their home. The evaluation of such a software application is limited without access to its target audience (lower limb amputee patients), hence, we have focused on defining and setting up practices metrics that can be used for future development that can benefit this, and other related projects. In addition, preliminary testing was carried out using healthy subjects and showed some encouraging results, particularly those that point to game usability and immersion experience. Finally, this capstone seeks to lay the foundation for future research and developments of this application for amputee patients, as well as providing a framework for other projects in the NeuroRehab effort.
Friday, December 13
REZA RAHIMIDERIMI
Chair: Dr. Hazeline Asuncion
Candidate: Master of Science in Computer Science & Software Engineering
9:00 A.M.; Join Reza Rahimiderimi’s Online Defense
Project: Enhancing Security Vulnerability Prediction in Software Projects Using Large Language Models and Deep Learning Techniques
Security vulnerabilities pose significant risks to software products, leading to financial losses, data breaches, and reputational damage for organizations. Traditional vulnerability detection methods often occur post-production, resulting in increased costs and delayed mitigation. This project aims to predict the severity levels of software vulnerabilities, as defined by the Common Vulnerability Scoring System (CVSS), based on software product specification documents and vulnerability descriptions. By leveraging deep learning models and Large Language Models (LLMs) such as DistilBERT, this research seeks to enable early identification of high-severity vulnerabilities during the development lifecycle, allowing organizations to address security issues proactively.
The methodology involves a dual approach of training models on both software specification documents and Common Vulnerabilities and Exposures (CVE) descriptions to predict CVSS severity levels. This comparative analysis determines which source provides better predictive performance. Baseline machine learning models were implemented for initial assessments, including Support Vector Machines and Random Forests. Advanced deep learning architectures, such as Recurrent Neural Networks (RNNs) and Long Short-Term Memory Networks (LSTMs), were employed to capture complex patterns and sequential dependencies in textual data. Feature extraction techniques utilizing LLMs were explored to enhance model performance by leveraging contextual embeddings and semantic understanding of the text.
Results indicate no noticeable improvements in model performance for deep learning and large language models (LLMs) compared to black box classifiers. After hyperparameter tuning and training on the vulnerability summary data, the Random Forest model achieved the highest accuracy among black box classifiers, reaching 79%. Among deep learning models, the convolutional neural network (CNN) recorded an accuracy of 79%, while DistilBERT, representing the LLMs, achieved 72%. Moreover, when we compared these results against the model performance on document specifications, we observed approximately a 15 to 20% decline in accuracy. Specifically, the results were as follows: Random Forest had 52% accuracy, CNN scored 67%, and DistilBERT attained 60% accuracy.
The analysis illustrated the differences, revealing that training on software specification documents has the potential for early-stage predictions in vulnerability assessment. However, compared to CVE descriptions, it accounts for less than 20% of the capabilities.
Master of Science in Cybersecurity Engineering
AUTUMN 2024
Monday, December 2
JEFFREY MURRAY JR.
Chair: Dr. Brent Lagesse
Candidate: Master of Science in Cybersecurity Engineering
11:00 A.M.; Join Jeffrey Murray Jr.’s Online Defense
Project: GhostPeerShare
Fully Homomorphic Encryption (FHE) schemes allow computations over encrypted data without access to the decryption key. Microsoft’s Simple Encryption Arithmetic Library (SEAL) provides a state-of-the-art C++ library that enables addition, subtraction, and multiplication on encrypted integers or real numbers. SEAL is difficult to implement on Android due to its outdated documentation. Furthermore, it requires expertise in C++, Kotlin, and Gradle. This project presents a natively compiled Dart plugin that abstracts the underlying C/C++ library. The FHEL plugin enables developers to access SEAL’s full functionality within other Dart plugins and Flutter applications. To evaluate the versatility of the plugin, we employ two implementations: 1) the Distance Measure Dart plugin and 2) the GhostPeerShare Flutter application. The Distance Measure plugin implements Kullback-Leibler Divergence, Bhattacharyya Coefficient, and Cramer Distance, along with their FHE counterparts. GhostPeerShare adapts the implementation from Proof of Presence Share (Pop-Share), a mobile application that can detect similar videos using FHE. GhostPeerShare achieved a precision score of 0.9614 and an F1 score of 0.9709 for identifying similar videos. Although video pre-processing was considerably slower, the average FHE computation of each distance measure matched that of Pop-Share. These results demonstrate that GhostPeerShare represents a significant advance in the accessibility of FHE within the mobile community.
Thursday, December 5
CHRISTIAN BERGH
Chair: Dr. Marc Dupuis
Candidate: Master of Science in Cybersecurity Engineering
11:00 A.M.; Join Christian Bergh’s Online Defense
Thesis: Cybersecurity is Stressful: The Impact of Stress on Identifying Phishing Attacks
Stress is an emotion that impacts everyone, it can have profound physical impacts on health and cognitive functions. The responsibilities placed on average workers in many fields can lead to increased workplace stress. As technology continues to be integrated into all facets of life and work, many industries have become de facto technology companies with large or valuable technical datasets. Many of these industries are facing an onslaught of cyber-attacks in an attempt to gain access to those large or valuable datasets. One of the most common cyber-attacks, and often found to be the entry way for many data breaches, remains to be phishing attacks. This combination of factors places a large amount of responsibility for a business’s digital security on the shoulders of every employee, while often security is not the primary responsibility of this employee. Malicious actors know that humans are the most vulnerable portion of any security network and understand that employees who may not understand how their company’s digital network functions and do not understand the value of access credentials to a malicious actor can be the easiest to target. Phishing attacks are designed and created to illicit stress emotions and a sense of urgency from the target in order to trick that target into clicking a link, downloading a file, or entering their credentials somewhere they should not. This study uses the framework and core concepts developed in the Trier Social Stress Test to create an acute instance of increased stress on participants. Participants are then shown several screenshots of emails and asked to determine if the email shown is a phishing attack. By analyzing the participants stress levels before and after the phishing attack identification test along with their performance on the test we are able to determine if there Is a significant impact on a participant’s ability to identify phishing attacks under increased stress.