Presented by: Damanpreet Singh

Date: Tuesday, December 2, 2025
Time: 12:30 pm
Place: Zoom - see below.

Join Zoom Meeting

https://uvic.zoom.us/j/88492710303?pwd=363d5QLjoowCZRoHXadtHtgxiSpyyy.1

Meeting ID: 884 9271 0303

Password: 349942

One tap mobile

+17789072071,,88492710303# Canada

+16475580588,,88492710303# Canada

Abstract: The rapid growth of malware attacks has created an urgent need for automated systems capable of accurately detecting and understanding malicious behavior. This project presents a comprehensive work for Malware Detection and Categorization using Machine Learning and Large Language Models (LLMs). The system's goal is to improve cybersecurity by not just detecting malware but also producing concise, intelligible descriptions of every threat it finds. The Microsoft Malware Classification dataset, which comprises approximately 21,000 malware samples grouped into nine primary families with corresponding .byte and.asm files, was adopted for the project. Since only malicious samples were present in the original dataset, roughly 15,000 benign files were added to enable binary categorization of malicious and non-malicious programs. XGBoost, LightGBM, SVM (RBF), and KNN were among the machine learning models that were trained and tested independently on both datasets. By applying the SMOTE technique, the dataset imbalance was reduced, thereby improving classification accuracy and mitigating bias toward the majority ofmalware families. Using the SMOTE technique, class imbalance was addressed. The models were evaluated for both binary classification (malicious vs. benign) and multi-class family prediction, achieving high detection performance. To enhance interpretability, an LLM-based explanation module was integrated. Following classification, the anticipated malware family is sent to an LLM (through Ollama), which produces a natural-language synopsis outlining the traits, actions, and defenses of the malware. Users can upload files, view predictions, and read the generated explanations in real time thanks to an intuitive Gradio interface. In order to provide both technical accuracy and human interpretability, the developed system successfully blends large language models for explainable analysis with machine learning for precise detection. By assisting researchers and security analysts in proactive malware defense, this method advances the field of intelligent cybersecurity by bridging the gap between detection and comprehension.

Presented by: Bhan Singh

Date: Thursday, December 4, 2025
Time: 10:00 am
Place: Zoom - see below.

Join Zoom Meeting
https://uvic.zoom.us/j/84799606469?pwd=8XjFd2YFJsAnaKexGWPZFOywCBtgBe.1

Meeting ID: 847 9960 6469
Password: 857861
One tap mobile
+17789072071,,84799606469# Canada
+16475580588,,84799606469# Canada

Dial by your location
        +1 778 907 2071 Canada
        +1 647 558 0588 Canada
Meeting ID: 847 9960 6469
Find your local number: https://uvic.zoom.us/u/kCAnGlowK

Abstract: The ocean plays a dual role in the Earth’s climate system, as a stabilizing reservoir of heat and carbon, and as an early indicator of ecological stress. Its changing state emerges from complex interactions among physical, chemical, and biological drivers, making traditional region-based assessments insufficient for capturing true environmental variability. This thesis develops a machine-learning framework that allows global in-situ observations from the World Ocean Database (WOD) to self-define the ocean’s natural biogeochemical regimes. After rigorous preprocessing and hierarchical spatio-temporal imputation, multiple clustering algorithms were used to uncover coherent patterns in the data, followed by supervised classification models to evaluate regime separability and interpret key drivers. The analysis identifies five distinct and ecologically meaningful regimes: productive coastal upwellings, oligotrophic gyres, polar waters, oxygen-minimum zones, and transitional open-ocean systems, reflecting the ocean’s intrinsic organization. By integrating unsupervised discovery with supervised validation, this work demonstrates how global ocean observations can be transformed into quantitative, interpretable indicators of ocean health, contributing to the broader vision of a data-driven digital twin ocean.

Presented by: Amirreza Balouchi

Date: Tuesday, December 9, 2025
Time: 12:00 pm
Place: Zoom - see below.

Join Zoom Meeting
https://uvic.zoom.us/j/87392272186?pwd=2ssHYuttIMQ69kX7UKCkSTFRYPkbKB.1

Meeting ID: 873 9227 2186
Password: 294335
One tap mobile
+16475580588,,87392272186# Canada
+17789072071,,87392272186# Canada

Dial by your location
        +1 647 558 0588 Canada
        +1 778 907 2071 Canada
Meeting ID: 873 9227 2186
Find your local number: https://uvic.zoom.us/u/kb5m1YrfpF

Note: Please log in to Zoom via SSO and your UVic Netlink ID

Abstract: 

Wangiri fraud is a pervasive telecommunications scam that exploits missed calls to lure victims into returning calls to premium-rate numbers, causing substantial financial losses for network operators and consumers. This study presents a machine learning framework for detecting Wangiri fraud in highly imbalanced and unlabeled Call Detail Record (CDR) datasets. The framework employs an unsupervised labeling method based on domain-driven heuristics and advanced feature engineering to capture temporal, geographic, and behavioral patterns indicative of fraudulent activity. To address class imbalance, resampling techniques, including the Synthetic Minority Oversampling Technique (SMOTE), Random Undersampling (RUS), and their hybrid variant, are systematically evaluated. Five classifier families (Logistic Regression, Decision Trees, Random Forests, XGBoost, and Multi-Layer Perceptrons) are benchmarked with and without isotonic and sigmoid probability calibration. Results show that ensemble methods, particularly Random Forest and XGBoost, achieve near-perfect performance, with accuracy exceeding 0.99 on balanced datasets while maintaining interpretability. The proposed pipeline offers a scalable and practical solution for Wangiri fraud detection, enabling operators to mitigate financial risks and enhance network resilience.

Presented by: Pratik Sinha

Date: Thursday, December 11, 2025
Time: 10:00 am
Place: Zoom - see below.

Meeting Details: 

Time: Dec 11, 2025 10:00 AM Vancouver

Join Zoom Meeting

https://uvic.zoom.us/j/81399376372

Meeting ID: 813 9937 6372

One tap mobile

+16475580588,,81399376372# Canada

+17789072071,,81399376372# Canada

Dial by your location

        +1 647 558 0588 Canada

        +1 778 907 2071 Canada

Meeting ID: 813 9937 6372

Find your local number: https://uvic.zoom.us/u/kcba6SxA1L

Abstract: 

Access control is a cornerstone of information security, defining how entities interact with protected digital resources. Traditional rule-based frameworks, though effective in static environments, struggle to adapt to modern, data-intensive ecosystems where roles, attributes, and contextual conditions evolve continuously. Recent advances in machine learning have introduced new opportunities to automate access control through predictive and adaptive modeling yet progress remains constrained by the scarcity of real-world datasets, inconsistent benchmarking methodologies, and limited evaluation under controlled data conditions. This thesis presents a reproducible framework for evaluating machine-learning based access control models using synthetic, configurable datasets. The proposed data generation process emulates healthcare authorization structures, incorporating tunable role hierarchies, permission ratios, and anomaly patterns to simulate varying entropy and data noise. A suite of ML architectures, including decision-tree ensembles, feed-forward networks, residual networks, and transformer-based tabular models are systematically benchmarked using standardized preprocessing and evaluation metrics. Experimental results reveal that decision-tree ensembles achieve strong baseline performance on small, structured datasets, while neural and transformer-based models exhibit greater robustness, scalability, and generalization as dataset volume and entropy increase. These findings validate the effectiveness of synthetic datasets for reproducible access-control research and demonstrate the impact of data scale on model elasticity and stability.

Presented by: Rudra Pratap Singh

Date: Thursday, December 11, 2025
Time: 11:00 am
Place: Zoom - see below.

Join Zoom Meeting

https://uvic.zoom.us/j/82178502011?pwd=Smd6DcSGwcl5ldtEngVLc3LaF6zjxa.1

Meeting ID: 821 7850 2011

Password: 277114

One tap mobile

+17789072071,,82178502011# Canada

+16475580588,,82178502011# Canada

Dial by your location

        +1 778 907 2071 Canada

        +1 647 558 0588 Canada

Meeting ID: 821 7850 2011

Find your local number: https://uvic.zoom.us/u/ks2F3GsAN

Abstract: The health of Canada’s blue-carbon ecosystems—kelp forests, seagrass meadows, and salt marshes—plays a vital role in marine biodiversity and long-term carbon sequestration. Yet these ecosystems are increasingly vulnerable to anthropogenic and natural stressors such as temperature variation, pH fluctuations, heavy-metal pollution, and hydrocarbon extraction. Traditional monitoring methods, relying on sporadic field sampling and manual analysis, fail to capture the temporal and spatial complexity of these changes. This thesis, Development of Machine Learning-Based Techniques for Monitoring and Analyzing the Effects of Natural and Manmade Stressors on Canada’s Blue Carbon Ecosystem Using a Secure Underwater Communication Suite, presents a comprehensive hardware-driven approach to address these gaps. The research involves the design, fabrication, and laboratory validation of a modular underwater sensor suite deployed via a Blue Robotics ROV platform to collect high-resolution oceanographic data. The integrated system measures temperature, salinity, dissolved oxygen, pH, turbidity, and chlorophyll concentrations through a network of calibrated probes, ensuring precise and repeatable environmental sensing. To support continuous operation, a secure underwater communication and data-handling framework was developed using a hybrid Ethernet-acoustic link and lightweight encryption protocols to preserve data integrity and mitigate cyber vulnerabilities within the Internet of Underwater Things (IoUT). Extensive laboratory testing in controlled aquatic environments demonstrated stable sensor calibration, minimal noise drift (< 0.05% FS), and consistent data throughput at depths up to 1 m. Complementary studies explored intrusion detection and federated-learning frameworks for distributed underwater nodes, strengthening the resilience of the proposed communication network

Presented by: Zev Isert

Date: Tuesday, December 23, 2025
Time: 11:00 am
Place: Zoom - see below.

Zoom Meeting:

https://uvic.zoom.us/j/7196958429?pwd=JGFamqjv7uO6LwsedPdQvxI5GdLIqU.1&omn=89185059543

Meeting ID: 719 695 8429

Password: 799800

One tap mobile

+17789072071,,7196958429# Canada

+16475580588,,7196958429# Canada

Dial by your location

        +1 778 907 2071 Canada

        +1 647 558 0588 Canada

Meeting ID: 719 695 8429

Find your local number: https://uvic.zoom.us/u/kEch4EAsc

Abstract:

This seminar presents methods for proactively monitoring data drift in classification models, aiming to provide early warnings before significant changes occur in the prediction distribution. These "proactive alerts" allow model operators to initiate retraining processes before models enter persistent failure modes, contrasting with traditional “reactive” alerts that only trigger after observing sufficient evidence of data misclassification. The research was conducted in collaboration with Revela Systems, addressing a common industry need for robust machine learning model monitoring.

The motivation for proactive monitoring stems from significant costs and time savings associated with preventing model failures. Existing monitoring approaches often fall short, either due to test data not being representative of live data or the inherent unpredictability of live data with unknown drift. The proposed approach aims to be lightweight, scalable, and industry-deployable by making pragmatic trade-offs against academic alternatives. The core contribution is a novel algorithm for detecting data drift in black-box classification models, focusing on understanding and tracking the model’s decision boundary. This algorithm is designed for production environments, capable of making predictions on both live and synthetic data. An incremental version of the algorithm, which does not require access to training data, can detect concept drift as it happens by querying the model. Furthermore, the thesis presents a method for synthesizing examples in high-dimensional feature spaces, crucial for mapping these decision boundaries.

The methodology centers on extracting meaning from a classifier’s decision boundary by learning how its output changes through data synthesis and live data monitoring. A classifier divides its input space into regions, with decision boundaries existing where the corresponding discriminant functions are equal. Two main approaches are explored: an eager mapping of the decision boundary in a reduced dimension and a lazy probing technique in the original feature space. Both techniques utilize a form of raycasting to determine if input data trends toward these boundaries, which implements the proactive component of the approach.

Experimental results were obtained using a custom “incident generator” tool that synthesizes infinite data streams based on the statistics of fixed-size datasets. The tool allows for real-time

mutation and injection of various data drift types, including incremental, sudden, and gradual drifts. The metric for evaluating the system’s performance was the percentage of probing rays that collided with a decision boundary. Experiments showed the monitoring system’s ability to detect incremental drifts, providing early indications as data trends toward decision boundaries. While the system could react to sudden drifts, forecasting them proved challenging due to the abrupt nature of the change. Gradual drifts offered an opportunity for forecasting, though as changes are non-positional, reaction could be delayed. The experiments also highlighted frailty in ray collision percentages as a measurement and a remaining significant processing time for certain high-dimensional scenarios.