Department of Computer and Electrical Engineering and Computer Science

Deep learning strategies combined with wearable sensors have advanced the capabilities of monitoring systems in biomedical applications, offering precise and efficient solutions for diagnosing and managing diseases. However, applying these systems faces several challenges. One of the challenges is the diminishing performance when these systems encounter new data with more complex patterns than those seen before. Another challenge is the limited availability of labeled data, on which deep learning-based systems depend highly. Additionally, obtaining high-quality labeled data to train deep learning models is often expensive, requiring significant time and resources. Another significant challenge is ensuring the practicality, accessibility, and convenience of the monitoring systems.
This dissertation proposes an innovative deep learning framework to overcome these challenges and improve system generalization performance in classification and regression tasks, specifically monitoring patients with neurological disorders like Parkinson’s.

The Cyber-Physical Systems (CPSs) and Internet of Things (IoT) have become emerging and essential technologies of the past few decades that connect various heterogeneous systems and devices. Sensors and actuators are fundamental units in most CPS and IoT systems, they are used extensively in vehicle systems, smart health care systems, smart buildings and cities, and many other types of applications. The extensive use of sensors and actuators, coupled with their increasing connectivity, exposes them to a wide range of threats. Given their integration into various systems and the use of multiple technologies, it is very useful to characterize their functions abstractly. For concreteness, we study them here in the context of autonomous cars. An autonomous car is an example of a CPS, which includes IoT applications. For instance, IoT units allow an autonomous car to be connected wirelessly to roadside units, other vehicles, and fog and cloud systems. Also, the IoT allows them to collect and share information on traffic, navigation, roads, and other aspects. An autonomous car is a complex system, not only due to its intricate design but also because it operates in a dynamic environment, interacting with other vehicles and the surrounding infrastructure. To manage these functions, it must integrate various technologies from different sources. Specifically, a diverse array of sensors and actuators is essential for the functionality of autonomous vehicles.

Human Activity Recognition (HAR) plays a crucial role in various applications, including healthcare, fitness tracking, security, and smart environments, by enabling the automatic classification of human actions based on sensor and visual data. This dissertation presents a comprehensive exploration of HAR utilizing machine learning, sensor-based data, and Fusion approaches. HAR involves classifying human activities over time by analyzing data from sensors such as accelerometers and gyroscopes. Recent advancements in computational technology and sensor availability have driven significant progress in this field, enabling the integration of these sensors into smartphones and other devices. The first study outlines the foundational aspects of HAR and reviews existing literature, highlighting the importance of machine learning applications in healthcare, athletics, and personal use. In the second study, the focus shifts to addressing challenges in handling large-scale, variable, and noisy sensor data for HAR systems. The research applies machine learning algorithms to the KU-HAR dataset, revealing that the LightGBM classifier outperforms others in key performance metrics such as accuracy, precision, recall, and F1 score. This study underscores the continued relevance of optimizing machine learning techniques for improved HAR systems. The study highlights the potential for future research to explore more advanced fusion techniques to fully leverage different data modalities for HAR. The third study focuses on overcoming common challenges in HAR research, such as varying smartphone models and sensor configurations, by employing data fusion techniques.

The Internet of Things (IoT) has undergone remarkable expansion in recent years, leading to a proliferation of devices capable of connecting to the internet, collecting data, and sharing information. However, this rapid growth has also introduced a myriad of security challenges, resulting in an uptick in cyber-attacks targeting IoT infrastructures. To mitigate these threats and ensure the integrity of data, researchers have been actively engaged in the development of robust Intrusion Detection Systems (IDS) utilizing various machine learning (ML) techniques. This dissertation presents a comprehensive overview of three distinct approaches toward IoT intrusion detection, each leveraging ML methodologies to enhance security measures. The first approach focuses on a multi-class classification algorithm, integrating models such as random forest, logistic regression (LR), decision tree (DT), and Xgboost. Through meticulous evaluation utilizing evaluation metrics including F1 score, recall, and precision under the Receiver Operating Characteristics (ROC) curve, this approach demonstrates a remarkable 99 % accuracy in detecting IoT attacks. In the second approach, a deep ensemble model comprising Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) architectures is proposed for intrusion detection in IoT environments. Evaluation on the UNSW 2018 IoT Botnet dataset showcases the proficiency of this approach, achieving an accuracy of 98.4 % in identifying malicious activities. Lastly, the dissertation explores a real-time Intrusion Detection System (IDS) framework deployed within the Pyspark architecture, aimed at efficiently detecting IoT attacks while minimizing detection time.

Telemetry data has become a crucial resource for detecting abnormal driving behaviors, especially for elderly drivers with Mild Cognitive Impairment (MCI) or dementia. This thesis proposes a novel spatial deep learning method that combines traditional telematics features with Grid-Index Resolution (GIR) to enhance the detection of abnormal driving behavior. By utilizing grid-indexed spatial-temporal analysis, the approach aims to capture more intricate driving patterns, which are often missed by traditional methods that rely only on basic telematics data such as speed, direction, and distance.
The methodology integrates Simple Neural Networks (SNN) to process traditional telematics features and Convolutional Neural Networks (CNN) to handle spatial relationships through grid-based data. The fusion of these two feature sets into a combined model improves the model's ability to accurately classify normal and abnormal driving behaviors.
This thesis evaluates the proposed approach using a dataset collected over 3.5 years from elderly drivers, including those with MCI. Experimental results demonstrate that the combined model achieves a classification accuracy of 97%, outperforming existing methods. The findings suggest that integrating grid-based spatial-temporal analysis into deep learning models offers significant potential for improving road safety, insurance risk assessment, and targeted interventions for at-risk drivers.

Short-circuit faults can cause significant damage to power grid infrastructure, resulting in costly maintenance for utility providers. Rapid identification of fault locations can help mitigate these damages and associated expenses. Recent studies have demonstrated that graph neural network (GNN) models, using phasor data from various points in a power grid, can accurately locate fault events by accounting for the grid’s topology—a feature not typically leveraged by other machine learning methods. However, despite their high performance, GNN models are often viewed as ”black-box” systems, making their decision logic difficult to interpret. This thesis demonstrates that explanation methods can be applied to GNN models to enhance their transparency by clarifying the reasoning behind fault location predictions. By systematically benchmarking several explanation techniques for a GNN model trained for fault location detection, we assess and recommend the most effective methods for elucidating fault detection predictions in power grid systems.

Promoting diversity in STEM fields is essential to fostering innovation and addressing global challenges. Despite extensive efforts, the representation of minority groups, including women, in undergraduate computer science and engineering programs remains low, posing significant barriers to equity and inclusivity in STEM education (Nicole & DeBoer, 2020).
This systematic review explores the socio-economic and cultural challenges discouraging minority students from pursuing degrees, specifically computer science and engineering disciplines. A comprehensive literature search was conducted across databases such as IEEE Xplore, Google Scholar, and Scopus using specific search terms. Studies were chosen based on clear inclusion and exclusion criteria.
Data was carefully extracted and analyzed, focusing on primary obstacles such as the scarcity of role models, biases, and educational barriers. To evaluate the quality of the studies included in the review, Covidence’s quality assessment tools were used, ensuring methodological rigor and consistency across the studies.

Businesses are the driving force behind economic systems and are the lifeline of the community as they help in the prosperity and growth of the nation. Hence it is important for the business to succeed in the market. The business’s success provides economic stability and sustainability that helps preserve resources for future generations. The success of a business is not only important to the owners but is also critical to the regional/domestic economic system, or even the global economy. Recent years have witnessed many new emerging businesses with tremendous success, such as Google, Apple, Facebook etc.. Yet, millions of businesses also fail or fade out within a rather short period of time. Finding patterns/factors connected to the business rise and fall remains a long-lasting question that puzzles many economists, entrepreneurs, and government officials. Recent advancements in artificial intelligence, especially machine learning, has lent researchers the powers to use data to model and predict business success. However, due to the data-driven nature of all machine learning methods, existing approaches are rather domain-driven and ad-hoc in their design and validations, particularly in the field of business prediction. The main challenge of business success prediction is twofold: (1) Identifying variables for defining business success; (2) Feature selection and feature engineering based on three main categories Investment, Business, and Market, each of which is focused on modeling a business from a particular perspective, such as sales, management, innovation etc.

In the modern data landscape, vast amounts of unlabeled data are continuously generated, necessitating development of robust unsupervised techniques for handling unlabeled data. This is the case for fraud detection and healthcare sectors analyses, where data is often significantly imbalanced. This dissertation focuses on novel techniques for handling imbalanced data, with specific emphasis on a novel unsupervised class labeling technique for unlabeled fraud detection datasets and unlabeled cognitive datasets. Traditional supervised machine learning relies on labeled data, which is often expensive and difficult to create, particularly in domains requiring expert input. Additionally, such datasets suffer from challenges associated with class imbalance, where one class has significantly fewer examples than another, complicating model training and significantly reducing performance. The primary objectives of this dissertation include developing a novel unsupervised cleaning method, and an innovative unsupervised class labeling method. We validate and evaluate our methods across various datasets, which include two Medicare fraud detection datasets, a credit card fraud detection dataset, and three datasets used for detecting cognitive decline.
Our unique approach involves using an unsupervised autoencoder to learn from dataset features and synthesize labels. Primarily targeting imbalanced datasets, but still effective for balanced datasets, our method calculates an error metric for each instance. This metric is used to distinguish between fraudulent and legitimate cases, allowing us to assign a binary class label. To further improve label generation, we integrate an unsupervised feature selection method that ranks and identifies the most important features without using class labels.

Mobility monitoring in urban environments can provide valuable insights into pedestrian and vehicle movement – where people want to go, how they get there, and the challenges they face along the way. Today, local governments can automate the acquisition of such data using video surveillance to understand the potential impact of investment and policy decisions. However, public disapproval of computer vision due to privacy concerns opens opportunities for research into alternative tools built with privacy constraints at the core of the design. WiFi sensing emerges as a promising solution. Modern mobile devices ubiquitously support the 802.11 standard and regularly emit WiFi probe requests for network discovery. We can passively monitor this traffic to estimate the levels of congestion in public spaces.
In this dissertation, we address three fundamental research problems pertaining to developing streetscape-scale mobility intelligence: scalable infrastructure for WiFi signal capture, passive device localization, and device re-identification.