IRis Bishop's University
Not a member yet
114 research outputs found
Sort by
Detecting Interprocedural Software Vulnerabilities Using Causal Deep Learning
Full text linkModern software systems often exhibit long interprocedural execution paths that span multiple functions and modules. The resulting interactions can introduce subtle vulnerabilities that adversaries may exploit. Many existing static, dynamic, and learning-based vulnerability detectors operate as coarse function, file or graphlevel classifiers and provide limited insight into how a vulnerability originates and propagates through the program. This dissertation introduces a unified approach that integrates vulnerability detection with explicit reconstruction of executable root-to-sink vulnerability traces, leveraging taint-propagation analysis to produce structured, traceable explanations that support exploitability assessment and remediation in real-world software systems. The approach represents programs as augmented graphs that integrate syntactic and semantic relations within and across functions. Interprocedural program representations are initialized via a pretrained structure-aware encoder and subsequently refined by a relation-aware graph neural network that integrates adaptive causal reasoning guided by Causal Knowledge Graph priors. Grounded in this representation, a constrained decoder constructs the most probable executable root-to-sink chain for each detected vulnerability and validates semantic consistency and aliasing constraints against program dependencies. Experiments evaluate detection performance and explanation quality using standard detection metrics combined with causal and propagation-path criteria. These criteria assess whether predicted chains satisfy feasibility constraints, recover the relevant interprocedural structure, and remain stable under targeted code edits. Experimental results demonstrate that the approach consistently generates accurate, executable root→propagation→sink chains that reflect underlying interprocedural behaviour and provide structured explanations for effective vulnerability localization and root-cause analysis in complex software systems.© Md Iqbal Hossain Shuvo, 202
Distinct Spatial and Functional Dynamics for Alpha and Gamma Oscillations in Human Visual Cortex
No full textVisual processing in the human brain relies on the functional organization of the visual cortex, characterized by retinotopic mapping and orientation tuning. Retinotopic mapping preserves the spatial layout of the visual field within cortical regions such as V1, V2, and V3, while orientation tuning enables neurons to respond selectively to specific stimulus angles, supporting edge detection and shape perception. Neural oscillations, particularly in the alpha (8–12 Hz) and gamma (30–80 Hz) frequency bands, play a central role in regulating these processes. Alpha rhythms are associated with inhibitory control, gating irrelevant information and facilitating selective attention, whereas gamma rhythms promote coherent communication between cortical areas, integrating sensory information for precise spatial and feature-specific processing. Electroencephalography (EEG) offers a high temporal-resolution method for investigating these dynamics, revealing how alpha suppression and gamma enhancement are modulated by visual stimuli. Experimental designs employing retinotopic and orientation-specific stimuli—such as rotating wedges, expanding rings, and oriented gratings—have shown that gamma responses are strongly spatially tuned and linearly summate across visual subfields, while alpha responses are broader, less spatially specific, and can occur even without direct visual input. Furthermore, gamma activity exhibits robust orientation selectivity, with oblique gratings often evoking stronger responses than horizontal or vertical ones, whereas alpha modulation is comparatively weaker and more variable. These findings indicate distinct but interacting circuit mechanisms underlying alpha and gamma rhythms, reflecting differences in spatial specificity, feature tuning, and their roles in visual attention. Understanding these relationships provides a mechanistic framework for linking large-scale neural dynamics to perception, with implications for both basic neuroscience and applications such as brain–computer interfaces and clinical interventions targeting visual and attentional disorders.© Sanaz Ghaffari Sarvarmaleki, 202
Network Inference And Graph Learning in Characterizing Seizure Dynamics from EEG Signals
Full text linkThis research aims to enhance seizure detection from EEG recordings using Wavelet Tools and Graph Neural Networks (GNNs). EEG signals are inherently nonstationary and complex, rendering traditional analysis methods less effective. Wavelet transforms provide a multi-resolution analysis, capturing both time and frequency information, which is crucial for identifying transient events like seizures. By utilizing cross-spectrum and bispectrum analysis, this work extracts both linear and nonlinear connectivity features to reveal dynamic interactions across brain regions and frequency bands. While Within Frequency Coupling (WFC) and Cross Frequency Coupling (CFC) have proven to be promising features for seizure detection in conventional machine learning approaches, these methods heavily rely on manual feature extraction. In contrast, Deep Learning methods can automatically learn representations from raw data. However, traditional architectures like CNNs and RNNs often struggle to capture the complex, multi-contextual relationships inherent in EEG data. Building upon state-of-the-art GNN models, we propose a unified network design that integrates spatial, temporal, and semantic information. We explore various message-passing and aggregation architectures within spectral GNNs to develop an optimal structure for seizure detection. The proposed framework is evaluated on the Temple University Hospital Seizure (TUSZ) Corpus, comparing its performance against existing methods using metrics such as AUROC and recall. Results demonstrate the superior performance of the Graph Wavelet Neural Network and suggest the importance of proper scale parameter selection.© Wei Cao, 202
On the Equivalence between Must-Test Specifications and the Linear Temporal Logic
Full text linkThis paper explores the connections between two formal verification techniques: must testing and Linear Temporal Logic (LTL). Both are vital for validating system behaviors and ensuring proper operation. Must testing captures dynamic behaviors and ensures system requirements are met under all execution scenarios, while LTL specifies and verifies temporal properties within systems. Concretely, we explore the equivalence between must testing and LTL. We develop a practical (algorithmic) framework to translate must-test specifications into equivalent LTL formulas. While we thus go only half-way toward establishing an equivalence, we believe that the conversion the other way around is possible as well. On a practical note, we also note that model checking (the algorithmic LTL verification framework) is a very mature technology widely uses in practice, while must testing is deployed to a much lesser extent. We therefore argue that our conversion is much more useful for practical applications. This being said, a conversion of LTL formulae into equivalent must tests is a necessary future expansion of this study to establish formally the equivalence of the two frameworks.@ Atefeh Farhadi, 202
A Novel Approach to Time Series Forecasting: BKA Optimization of XGBoost
Full text linkTime-series forecasting is critical for decision-making in domains like finance, meteorology, and energy management, where suboptimal predictions can lead to significant economic losses. While machine learning models such as XGBoost offer promising accuracy, their performance heavily depends on hyperparameter tuning— a process challenged by high-dimensional, interdependent parameters and
the risk of local optima. This study based on BKA-XGBoost, a novel framework that integrates the Black Kite Algorithm (BKA), a metaheuristic optimizer inspired by avian foraging behavior, with dynamic constraint projection to enhance XGBoost’s hyperparameter optimization.
BKA’s dual-phase exploration-exploitation mechanism addresses limitations of conventional methods (e.g., PSO’s premature convergence and GA’s fixed mutation rates), while the dynamic constraint mechanism ensures feasible hyperparameters throughout optimization. Experimental results on eight real-world energy datasets demonstrate that BKA-XGBoost reduces forecasting errors by 12–18% in MAE and 19% in MAPE compared to standard XGBoost, translating to potential annual savings of $600K for medium-scale power grids—a critical improvement given that even a 1% MAPE reduction mitigates substantial operational costs. The framework also achieves 22% faster convergence and 62% fewer invalid hyperparameters than unconstrained optimization, validated through rigorous time-series cross-validation and statistical testing (p < 0.01).© Yang Ming, 202
Human Action Decoding Analysis from Neurophysiological Time Series
Full text linkThis thesis investigates the hypothesis that electroencephalography (EEG) signals can be used to reliably decode and classify distinct motor actions during gameplay with high accuracy and cross-subject generalizability. Using machine learning methods and frequency-based feature extraction, particularly focusing on Beta band activity, we developed a classifier capable of recognizing in-game actions from EEG data. The results demonstrate successful classification of motor actions and consistent cross-subject generalizability, validating our core hypothesis. The developed methodology showed robust performance in identifying task-relevant neural markers, with Beta band features emerging as the most reliable for action classification. While the study achieved its primary objective, the distinction between motor intentions and executed actions could not be directly assessed due to experimental design limitations. This work establishes a framework for real-time EEG-based action decoding, paving the way for more intuitive human-computer interaction systems in gaming and assistive technology.© Raana Nouri, 202
The Final Act: Navigating Death, Ritual, and Caregiving through Playwriting
Full text link@ Benjamin Tabah, 202
FOR THE LOVE OF LEARNING! Exploring Teacher Perceptions of Student Motivation.
Full text linkThis study explores the perceptions of 13 secondary teachers, with experience ranging from 3 to 30 years, regarding their role in fostering student motivation and a love of learning in English-language minority schools in rural Quebec. Motivation is widely recognized as a key factor in student success. Frameworks such as Self-Determination Theory, which emphasizes the role of autonomy, relatedness, and competence, and Growth Mindset, which highlights the belief in the ability to develop through effort, provide a foundation for understanding how students engage with learning. Fostering a love of learning is a newly constructed teacher competency that was an addition to the 2021 Reference Framework of Professional Competencies for Teachers. The competency aims to contribute to improved graduation rates in the province, as outlined in the Quebec Ministry of Education's Policy on Educational Success: A Love of Learning, A Chance to Succeed. This framework aims to increase success outcomes by 2030.
Guided by the research question, "How do teachers perceive their role in motivating students?", this study employed a qualitative approach, using open-ended interviews to explore teachers’ viewpoints. Coding and analysis of the interviews led to the emergence of three themes: the significance of student-teacher relationships, challenges that must be faced for teachers to support students with motivation, and the strategies the teachers believe contribute to engagement. Social-Emotional Learning (SEL), building empathy, and fostering democratic values in the classroom are offered as essential components for cultivating motivation and supporting student success.@ Todd Smith, 202
“It just kind of defines you”: Expectations and Reactions to Injury and Athletic Identity in Post-Secondary Student-Athletes
Full text linkStudent-athletes are distinct within the wider athletic system as well as the general student population. Their unique athletic identity and experience with injury may impact the overall well-being and success of this group. Using a mixed methodology with a quantitative survey and semi-structured qualitative interview, participants provided data on injury history, athletic identity, coping strategies, well-being, and personal experiences. The final sample included 16 student-athletes (n = 9 cisgender women; n = 7 cisgender men) from a wide breadth of sports such as basketball, figure skating, karate, lacrosse, and others. The type of injuries experienced ranged from sprains and tears to concussions and herniations. Four qualitative themes were identified using reflexive thematic analyses. Results revealed how the background dynamics (e.g., coach, teammates, sport’s culture) played a role in the student-athlete’s injury experience; the interaction between the student and athlete identity; the various facets of the student-athletes’ injury experience; and which recovery resources and treatments were utilized and the student-athletes evaluation of the resources’ success. This study can share how student-athletes view their own identity and allow athletic and academic communities to have a greater understanding of the obstacles, challenges, and effective injury coping skills faced by this group.© Rebecca Benyk, 202
An Empirical Evaluation of Model Design and Code Representation for Deep Learning-Based Source Code Vulnerability Detection
Full text linkAs modern software systems continue to grow in scale and complexity, ensuring their security has become an increasingly demanding challenge, with hidden vulnerabilities threatening both reliability and safety. Deep learning offers powerful approaches to automating vulnerability detection by capturing structural and semantic code patterns that are often missed by traditional rule-based techniques. Research in this area has largely advanced along two main paradigms: sequence-based, which treat code as token streams, and graph-based, which leverage structural representations of code. The rise of pre-trained transformers has further accelerated progress by modeling semantic and contextual nuances with greater fidelity. Yet key questions remain unresolved, including the specific contribution of pre-training to vulnerability detection, the relative merits of sequence-based and graph-based approaches, the influence of structural design choices within graph models, and the ability of these methods to generalize across different programming languages. This work undertakes a comprehensive empirical study to address these gaps. The results demonstrate that pre-training substantially improves detection performance, underscoring the value of contextualized embeddings for capturing code semantics. Graph-based architectures, particularly those built on code property graphs, consistently deliver more accurate results by integrating syntactic structures with semantic dependencies. The findings further reveal that fusing syntactic and semantic representations produces the most reliable detection of complex, real-world vulnerabilities. At the same time, current methods exhibit significant performance degradation when applied in multilingual settings, highlighting the urgent need for language-independent strategies. By examining the interplay of pre-training, model architecture, and code representation, this study establishes a rigorous evaluation framework that clarifies their individual and collective contributions to reliable vulnerability detection. Beyond advancing accuracy, it emphasizes the importance of explainability in helping developers understand detected risks, laying the groundwork for scalable, generalizable systems capable of securing today’s increasingly diverse and heterogeneous software environments.© Abdullah Khan, 202