Upcoming Student Presentations

Date: Tuesday, January 28, 2025

Presentation 1:

Title of Presentation: Novel optimizers for multi-objective machine learning

Presenter Name: Farzaneh Nikbakhtsarvestani (PhD Candidate, Computer Science, Ontario Tech University)

Supervisor: Dr. Mehran Ebrahimi (Faculty of Science, Ontario Tech University) and Dr. Shahryar Rahnamayn (Faculty of Engineering and Applied Science, Ontario Tech University)

Abstract: Multi-objective optimization is a prevalent challenge in the area of deep learning. There is a lack of robust multi-objective optimization methods applicable in deep learning capable of training networks by simultaneously optimizing conflicting multiple loss functions. Its applications include a wide range of deep neural network branches such as multiloss, multi-task, multi-modal, and crossmodal learning. In this thesis, we develop two novel optimizers: Multi-objective Adam Optimizer (MAdam) and Opposition-base Multi-objective Adam Optimizer (OMAdam). MAdam is a multi-objective extension of the well-known Adam optimization algorithm. MAdam is a classical population-based approach that uses the gradient information of multiple objectives to accelerate population convergence toward an optimal minimum. The method applied a non-dominated sorting algorithm to keep selective population and improve the diversity across the landscape. In OMAdam, opposition based scheme was introduced into MAdam framework as global search is necessary for escaping local optima in gradient-based multi-objective optimization approaches. The Opposition-based MAdam, explores multiple directions over the landscape, that leads to independence from specific initialization. These two optimizers are scalable in every deep learning tasks. In a series of experiments for MAdam and OMAdam applications, we investigate multi objective binary and multi-class classification tasks using novel Madam and OMAdam optimizers. The aim of the study is to provide a range of solutions on the Pareto front to balance trade-offs between conflicting loss functions and to select corresponding model parameters that reflect the

clinical priorities of practitioners (classes and sub-classes). In this regard, we employed these two novel optimizers as AI diagnostic techniques based on the 10-class MNIST and 2-class, 8-class, and 11-class MedMNIST datasets. In each experiment, the Pareto front of model parameters allows for variations in performance metrics and enable the prioritization of specific classes and sub-classes targeted by the end-users. MAdam and OMAdam optimizers achieve F1-scores comparable to those of the single objective Adam optimizer among their many solutions on the Pareto front. Also, we investigate the functionality of Madam and OMAdam optimizers in another deep learning task: Multi-objective regression task. In this study, we modify the optimization process to include conflicts losses for fitting error, model complexity, and Huber loss, which represent the model’s performance on data consistency, regularization, and resistance to outliers, respectively. we incorporate these conflicting loss terms to MAdam and OMAdam framework. These schemes, address the challenges of noise sensitivity and overfitting in scenarios dominated by collinearity or high-dimensional feature spaces. We demonstrate how these two novel optimizers can efficiently train artificial neural networks by balancing these conflicting objectives without the need for multiple retraining cycles. Our approach constructs a Pareto front that provides a spectrum of solutions, each offering a different trade-off between model’s accuracy , complexity and robustness. These methods broadens the applicability of MAdam and OMAdam optimizers in deep learning, particularly for critical applications like financial forecasting and healthcare predictions, where decision-making relies heavily on the robustness and accuracy of predictive models.

Presentation 2:

Title of Presentation: Balancing Security and Usability: A PII-Driven Framework for Password and Honeyword Generation in Modern Authentication Systems

Presenter Name: Hannah Adjei (PhD Candidate, Computer Science, Ontario Tech University

Supervisor: Dr. Miguel Vargas Martin

Abstract: The growing demand for strong yet easy-to-use authentication systems is pushing forward the development of innovative password-generation techniques to respond to the increasing threat of cyber-attacks and the critical need for user-friendly security measures.

This research investigates a novel approach to creating strong yet memorable and usable passwords that utilize Personally Identifiable Information (PII) extracted from user data. A systematic methodology was employed to evaluate PII-based passwords alongside randomly generated PII passwords (RaPII) and purely random passwords (Ra-Pass). Password strength was assessed using zxcvbn scores, cosine similarity measured honeywords’ closeness to actual passwords, and Levenshtein distance quantified memorability. Our comprehensive analysis, including ANOVA followed by Tukey HSD tests, underscored significant distinctions among the different password categories. Building on the results, which highlight the enhanced usability and robustness of passwords from PII, a comprehensive user study utilizing the System Usability Scale (SUS) provided more profound insights into users’ interactions and perceptions of these passwords. This research contributes a solid framework for developing PII-driven password and honeyword systems, integrating statistical analysis with a focus on user experience. We address current cybersecurity challenges and aim to set the foundation for future improvements in authentication technologies. By effectively blending the familiarity of PII with rigorous security protocols, it seeks to provide secure, user-focused solutions.

Presentation 3:

Title of Presentation: Gazed-enabled Implicit Interaction for Scatterplots

Presenter Name: Tania Shimabukuro (MSc Candidate, Computer Science, Ontario Tech University)

Supervisor: Dr. Christopher Collins (Faculty of Science, Ontario Tech University)

Abstract: How we look at the world can tell much about our focus and interests. Yet, this information is not used when displaying information visualizations. This research investigates novel display and interaction designs for create gaze-responsive visualizations optimized to support analysis of the items of interest. Specifically, we investigate four opportunities for implicit gaze-driven interaction for subtle visualization optimizations using eye tracking to improve standard scatter plot tasks: rendering order changes, hover speed modifications, local reference axes, and click accuracy improvements. A user study compared and evaluated the effectiveness of these techniques. Preliminary results show that the technique "hover speed," which consists of changing speed to show a tooltip based on the gaze-driven area of interest, was the favourite among the participants. Opinions on the "reference axis" technique were divided. Overall, the interest model improved their performance in both types of tasks, indicating that gaze-driven techniques can support data analytics tasks positively.

Date: Tuesday, December 3, 2024

Presentation 1:

Title of Presentation: Hybrid architecture for human action recognition using skeleton data

Presenter Name: Muhammad Salik Nadeem (MSc Candidate, Computer Science, Ontario Tech University)

Supervisor: Dr. Faisal Qureshi

Abstract: In this work, we propose a deep learning architecture, incorporating a Graph Convolutional Network (GCN) backbone combined with a partitioning transformer, that achieves results comparable to the state-of-the-art methods in skeleton based multi-person, multi-view human action recognition. By leveraging attention-based GCN, the model captures context-dependent intrinsic topology while enhancing discriminative information. Furthermore, utilizing transformers, we harness their ability to aggregate long-range temporal information, allowing us to learn complex actions by attending to both short-term and long-term temporal windows. This is achieved through our partitioning strategy, which efficiently captures the relationships between neighboring and distant joints, enabling a comprehensive understanding of human movement dynamics. In this work, we also introduce a Cosine-based noise as a new data augmentation strategy for joints across time. This helps improve our model, Hybrid-Graformer, achieve accuracy comparable to the state-of-the-art across various skeleton-based action recognition benchmarks: 94.46% on NTU RGB+D 60 cross-subject split, 97.87% on NTU RGB+D 60 cross-view split. 90.68% on NTU RGB+D 120 cross-subject split, 93.52% on NTU RGB+D 60 cross-view split and 97.91% on NW-UCLA dataset.

Presentation 2:

Title of Presentation: A tokenized mechanism for federated applications in vehicular networks

Presenter Name: Mohammad Hajyghasem (MSc Candidate, Computer Science, Ontario Tech University)

Supervisor: Dr. Shahram Heydari

Abstract: This thesis proposes a tokenized mechanism for federated applications in vehicular networks to address the challenges posed by dynamic and decentralized environments. Vehicular networks require secure, flexible, and scalable communication systems, but traditional approaches struggle to adapt to rapid changes in such environments. A blockchain-based tokenization model is introduced to solve this, enhancing data integrity, security, and incentivization in federated applications. The study employs blockchain smart contracts to manage vehicle-to-vehicle (V2V) interactions and ensure secure, decentralized data management and resource allocation. A detailed system architecture is designed and implemented, followed by performance testing. Results demonstrate that the proposed tokenized mechanism encourages better client performance. By linking resource allocation to token earnings, it will create a competitive yet efficient network where vehicles are rewarded based on their contribution. The system's ability to dynamically manage vehicular communications suggests it could have significant implications for future intelligent transportation systems.

Presentation 3:

Title of Presentation: Hybrid ConVIRT - Enhancing medical image-text representation learning of vision language models

Presenter Name: Abhinav Sharma (MSc Candidate, Computer Science, Ontario Tech University)

Supervisor: Dr. Mehran Ebrahimi

Abstract: Abstract: Challenges in medical imaging are addressed by advancing image-text representation learning, as exemplified by Hybrid-ConVIRT, which builds on contrastive learning frameworks such as ConVIRT and MedCLIP. Traditional medical imaging models rely on costly, expert-annotated datasets, limiting scalability. Transfer learning from general datasets, like ImageNet, often fails to capture the domain-specific nuances critical in clinical applications. The Hybrid model improves upon this by incorporating domain-specific transformations and using CXR-BERT-specialized as the text encoder, tailored specifically for chest X-ray reports. To optimize feature alignment, it introduces a ‘Fused Similarity Matrix (FSM)’ which potentially serves as a better target, it leverages the learned capabilities of MedCLIP and COnVIRT. Evaluation across four datasets (COVID, Tuberculosis, Pneumonia/TB Mix, and CheXpert) using linear probe and zero-shot classification demonstrates Hybrid-ConVIRT’s enhanced discriminative power and feature robustness, particularly on COVID and CheXpert. While MedCLIP shows a slight advantage on the Tuberculosis dataset, it is likely due to the larger training set. The Hybrid model’ s performance supports the potential of combining learnings of both the prior models.

Date: Tuesday, November 12, 2024

Presentation 1:

Presenter Name: Ghazal Bangash (MSc Candidate, Computer Science, Ontario Tech University)

Supervisor: Dr. Loutfouz Zaman

Title of Presentation: HoloPaws: Run – A Comparative Study of Exergame User Interfaces in Mixed Reality and Mobile Devices

Abstract: Using Microsoft’s HoloLens 2, this work explores how diegetic and non-diegetic elements impact player motivation, engagement, and performance in fitness games. This

study compares three setups: an Android fitness app, a non-diegetic HoloLens game, and a diegetic HoloLens game. The Android app provides real-time step data and cadence

feedback through a mobile interface. The non-diegetic HoloLens game displays feedback on the user interface via a cadence bar, step counter, and pop-up messages. In contrast,

the diegetic HoloLens game uses a virtual dog companion to communicate feedback by running alongside the player or sitting when the cadence drops. With 24 participants, the mixed-method study combined gameplay, questionnaire, and interview data. The analysis provided insights into how different feedback styles affect performance, motivation, and user experience. The findings highlight the potential of diegetic elements to enhance engagement in mixed-reality fitness applications, offering valuable design recommendations for outdoor exercise games

Presentation 2:

Presenter Name: Andrei Bosco Bezerra Torres (PhD Candidate, Computer Science, Ontario Tech University)

Supervisors: Dr. Bill Kapralos and Dr. Adam Dubrowski

Title of Presentation: Interdisciplinary Collaboration Towards Developing Platforms for Simulation-Based Learning: Moirai - No-code Virtual Serious Game Authoring Platform - and the GEN - Gamified Educational Network

Abstract: The COVID-19 pandemic lockdowns and stay-at-home orders resulted in an abrupt move from traditional face-to-face instruction to remote learning, particularly affecting healthcare education as laboratories shut down and healthcare placements and clerkships were unavailable. To mitigate, technology, including virtual simulation and serious gaming, has facilitated experiential ("hands-on") learning. Although the pandemic has ended, there is a consensus among healthcare professionals and educators to retain some of the aspects of remote simulation. However, some limitations must be overcome, including the one-size-fits-all approach in most virtual simulations and serious games, which assume fixed scenarios that the educator cannot easily modify. An additional challenge is facilitating the integration of a learning management system with virtual simulations and game-based learning into the curriculum while encouraging peer-based feedback. Following an educational design-based research methodology, I developed two platforms: Moirai, a no-code authoring platform for creating dialogue-based serious games, and the Gamified Educational Network (GEN), an experimental learning management system (LMS). Moirai focuses on allowing educators with limited, if any, programming knowledge to independently create and modify serious games that fit their curriculum using a user-friendly visual programming interface. The GEN LMS focuses on allowing educators to implement courses based on experiential learning (including simplifying the ability to integrate virtual simulations and serious games into a course) and peer-based feedback concepts, with automatized connection between user-generated content and feedback elements. Ten research projects were co-developed in partnership with various laboratories and institutions, adopting Moirai and/or the GEN LMS, performing user studies, and resulting in an iterative evolution of both platforms. Usability studies with subject-matter experts and learners indicate that the GEN LMS and the Moirai Authoring Platform (including the editor functionality and the resulting games) have achieved higher than average usability scores. Although effectiveness has not been tested, usability is the first step in an effective educational application as it will help facilitate interaction, accessibility and possibly engagement for learners, potentially enhancing the learning experience.

Presentation 3:

Presenter Name: Jakob Ethan Anderson (MSc Candidate, Computer Science, Ontario Tech University)

Supervisor:  Dr. Andrew Hogue

Title of Presentation: Exploring Temporal Volumetric Video Compression using Signed Distance Fields

Abstract: Volumetric Video is an exciting medium that enables the visualization of 3D time-varying data. It is often experimented with in either a time-varying Point Cloud or Textured Mesh format, although newer formats are also making their way in. Despite over a decade of focus, there still exist several problems that make it disadvantageous to use for researchers and consumers alike. This research chooses to focus on reducing the size of Volumetric Video consisting of Textured Meshes, as its file sizes often become very large for sequences consisting of a mere few seconds of footage. In particular, this research looks at constructing a Signed Distance Field (SDF) from these Textured Meshes, such that temporal variations in the data may be more easily explored and exploited due to the structured nature of SDFs. The goal of this work is to produce/experiment with a novel compression method for these Time-Varying Textured Meshes that is able to maximize a file size to error compression ratio, while maintaining a decompression speed that is able to run in a real time of 30 Frames Per Second on high-end, consumer-available hardware. The results reflect that as file sizes decrease, the proposed method's error degrades at a slower rate than the chosen state-of-the-art. The conclusion reveals several improvements/possible tests that could be made as an extension of those provided in this work.

Presentation 4:

Presenter Name: Mansour Alqarni (PhD Candidate, Computer Science, Ontario Tech University)

Supervisor:  Dr. Akramul Azim

Title of Presentation: SecureLLAMA: Secure FPGAs Using Llama Large Language Models

Abstract: Field-Programmable Gate Arrays (FPGAs) are increasingly utilized in critical applications across sectors such as infrastructure, defense, and autonomous systems. However, the inherent flexibility of FPGAs introduces significant security vulnerabilities, particularly in the hardware description languages (HDLs) used to program them. This paper introduces SecureLLAMA, an enhanced version of the LLAMA2 model, specifically designed to detect and mitigate FPGA vulnerabilities. Leveraging a novel dataset "FPGAvul" which includes both real-world examples and synthetically generated vulnerabilities.Our dataset FPGAvul addresses vulnerabilities such as initialization errors, clock domain crossing issues, insecure state machines, resource sharing conflicts, and buffer overflows. SecureLLAMA demonstrates superior accuracy in identifying and addressing security flaws in FPGA configurations. Comprehensive evaluation shows that SecureLLAMA significantly improves the detection of vulnerabilities, providing a robust solution for securing FPGAs in embedded systems. The findings of this research have the potential to advance FPGA security practices, ensuring their safe integration in critical environments where reliability is essential.