Computer Science - Doctoral Theses
Permanent URI for this collection
Browse
Recent Submissions
Item Modelling uncertainty in cloud workload forecasting: a Hybrid Bayesian Neural Network approach(University College Cork, 2024) Rossi, Andrea; Brown, Kenneth; Prestwich, Steve; Visentin, Andrea; Science Foundation Ireland; Horizon 2020The exponential growth of cloud services utilisation has rendered robust and accurate workload forecasting paramount. Predicting future computational demands is essential for efficient resource management, cost optimisation, and guaranteeing service level agreement (SLA). At the same time, the increase in cloud resource demand poses new challenges related to data centres' carbon emissions and their environmental impact. However, these predictions are characterised by inherent uncertainty stemming from both fundamental limitations in data and the dynamic nature of cloud usage. This dissertation delves into workload forecasting in cloud computing. Our primary focus is quantifying predictions' inherent uncertainty, an important but often overlooked aspect. Traditionally, forecasting models provide point estimates, leaving cloud providers blind to the potential range of future demands. This lack of uncertainty awareness can lead to suboptimal resource allocation, unnecessary costs, and SLA violations. In particular, our work focuses on quantifying the uncertainty of different natures. The epistemic uncertainty arises from limitations in our knowledge or model, while the aleatoric uncertainty stems from the variability in the data itself, reflecting the inherent randomness of cloud workloads. By capturing both these forms of uncertainty, we gain a deeper understanding of the prediction's confidence interval, empowering informed decision-making. This work introduces a novel Hybrid Bayesian Neural Network (HBNN) model designed to capture both types of uncertainty in workload forecasting by incorporating a Bayesian layer at the network's end. The HBNN surpasses conventional approaches by estimating the future probability distribution to compute the associated confidence intervals. This quantifiable uncertainty empowers the model for more accurate prediction. The thesis expands upon the HBNN model by exploring its application in bivariate forecasting, simultaneously predicting processing units and memory demands. This approach proves good accuracy at the cost of more training data required. Furthermore, this work investigates the impact of various factors on forecast accuracy, including training data size, unseen data generalisation, and the potential of transfer learning across different cloud environments. Furthermore, recognising the potential computational hurdles associated with Bayesian neural networks, this work introduces a clustering-based preprocessing technique. This technique intelligently selects training data points, significantly reducing computational cost while maintaining forecast accuracy. This development enables the HBNN model to be deployed in cloud environments at scale. The thesis comprehensively evaluates the HBNN model through extensive experiments on real-world cloud workload datasets. The results demonstrate the model's superior performance in quantifying uncertainty, achieving higher accuracy than baseline models and showcasing improved generalisation on unseen data. By quantifying inherent uncertainty, HBNN empowers cloud providers to manage cloud resources more efficiently, reliably and cost-effectively.Item Loneliness detection using technology(University College Cork, 2024) Qirtas, Malik Muhammad; Pesch, Dirk H J; Bantry White, EleanorOver the last decade, especially during the COVID-19 pandemic, loneliness has emerged as a global issue. Although loneliness is a common issue that most people experience at some stage in their lives, it can have harmful effects on both physical and mental health if this condition becomes chronic. To avoid the long-term effects, detecting it in the early stages is critical, which would lead to timely intervention and treatment. Modern smartphones and wearables are equipped with a wide range of sensors that can provide a vast amount of information on people's daily lives and behaviours in real time, which can help detect early signs of loneliness. Researchers collect extensive data on a user's daily life and behavioural patterns, including social activities, mobility patterns, communication, and activity data through sensors like accelerometers, heart rate sensors, microphones, and GPS and Bluetooth connectivity acting as proximity sensors. All these capabilities of smartphones and wearables have made them a powerful tool for monitoring users' health and well-being passively. Many studies have used passive sensing for loneliness detection in recent years, but these methods have severe limitations. The main issues are related to relying on generic models, and neglecting the dynamic, multifaceted nature of loneliness. This thesis made several contributions to address these critical gaps in the literature. First, we presented an approach for personalized loneliness detection through the behavioural grouping of passive sensing data. We used unsupervised machine learning for this and then trained customized models for each identified group that outperformed generic models for loneliness detection. Building upon this work, we addressed its limitation of static approach for loneliness detection. So, we proposed an approach for dynamic loneliness detection that can adjust to users' changing behaviours. We also analyzed how users' behaviours changed throughout the study period and how these changes are linked with their loneliness levels. This contribution helps detect loneliness in its early stages and highlights the importance of recognizing individual differences in how loneliness manifests. Previous research on loneliness with passive sensing data often analyzed loneliness as a single concept. This thesis explored the different types of loneliness and their behavioural manifestations. Mainly, we investigated how behavioural patterns can distinguish between social and emotional loneliness and what is the predictive power of these digital biomarkers to detect the specific type of loneliness in order to develop more targeted interventions. Research shows that loneliness and depression often co-occur, but their behavioural overlap wasn't well understood in terms of behavioural patterns extracted through passive sensing. We investigated the complex relationship between these two conditions and identified both overlapping and distinct behavioural markers captured through passive sensing along with the predictive power of these markers. This work helps us understand how they interact with each other, and this could lead to improved detection and intervention strategies for individuals experiencing these conditions.Item Detecting targeted interference in the Internet of Things(University College Cork, 2024) Morillo, Gabriela; Roedig, Utz; Pesch, Dirk H J; Science Foundation IrelandThis thesis investigates targeted jamming interference detection to enhance security in the Internet of Things (IoT) infrastructures. The study starts by assessing the critical role of IoT system monitoring in securing large networks, emphasising the need for automated solutions to detect and mitigate threats, ensuring continuous and reliable operations. This provided insight into how interference monitoring solutions should be implemented. The development of this kind of detector is important as naturally occurring interference requires a different response than targeted interference attacks. A significant portion of the thesis is dedicated to addressing vulnerabilities in the Narrowband-Internet of Things (NB-IoT), a Low Power Wide Area Network (LPWAN) radio technology required for large-scale IoT deployments. Initially, it looks specifically into how interference with NB-IoT synchronisation signals can lead to Denial of Service (DoS) attacks, highlighting the need to prevent and mitigate such vulnerabilities. A novel attack on the initial communication steps is provided in this investigation. To address these challenges, this work introduces a novel method for detecting targeted interference at the User Equipment (UE) level in NB-IoT networks. Our solution utilises network performance data and subframe loss rates to differentiate between targeted attacks and naturally occurring interference, which is critical as they require different responses. The costs associated with designing dedicated detectors for each technology, including established and upcoming ones, are high. Therefore, we propose a technology-independent approach to detect targeted interference across various IoT networks. This solution, designed to function on resource-constrained IoT devices, analyses packet loss rates and patterns to detect the presence of targeted attacks. This detection technique has been proven through comprehensive assessments using several IoT technologies, including NB-IoT and IEEE 802.15.4 GTS, demonstrating its effectiveness in distinguishing targeted interference from natural interference. This work advances the state of the art in detecting malicious interference in IoT environments by introducing a technology-independent targeted interference detection method capable of operating on resource-constrained IoT devices. Unlike prior research, which has primarily focused on machine learning IDS or including additional hardware for their solutions, our approach monitors packet loss rates and patterns across different wireless communication technologies (e.g. Narrowband Internet of Things and IEEE 802.15.4) to perform statistical anomaly detection. This is the first research to propose and validate a comprehensive, technology-independent framework that effectively distinguishes between targeted attacks and natural interference, thereby significantly enhancing the security and resilience of heterogeneous IoT deployments. Overall, our research emphasises the importance of robust monitoring systems and innovative defence mechanisms to safeguard IoT infrastructures against evolving and emerging threats while also contributing valuable insights and tools to enhance the resilience of critical IoT applications.Item Time-Sensitive Networking for industrial IoT: integration, analysis, and performance evaluation(University College Cork, 2024) Seliem, Mohamed; Pesch, Dirk H J; Zahran, Ahmed; Science Foundation IrelandIndustrial automation networks demand precise timing, minimal latency, and negligible packet loss for efficient real-time data exchange. Time-Sensitive Networking (TSN) emerges as a crucial technology for future automation, promis ing enhanced timing accuracy, reduced packet delay, and improved networking determinism. This thesis explores and innovates within TSN functionalities to address key aspects of industrial networking and related technologies. The critical need for reliable real-time data exchange across industries is examined, introducing TSN principles such as time synchronisation, deterministic communication, traffic shaping, and Quality of Service (QoS) assurances. Through simulation, typical industrial use cases and traffic requirements are evaluated, focusing on priority queuing, Time Aware shaping (TAS), and Credit Based Shaping (CBS) to meet latency constraints. The findings demonstrate TSN’s ability to orchestrate network traffic while adhering to strict timing requirements, highlighting its practical relevance in industrial automation. In smart manufacturing environments, the optimisation of industrial networks for Quality Control and Classification After Production (QCAP) is emphasised. By leveraging TSN standards, diverse QoS requirements are addressed to enhance efficiency and reliability. Fault tolerance in Industrial Internet of Things (IIoT) applications is also investigated using network calculus principles to analyse worst-case latency, providing insights into network performance and stability. The research integrates TSN with Software Defined Networks (SDN) to manage network configurations, focusing on traffic scheduling in industrial applications. Network-device contracts are proposed for traffic schedule computation and distribution, demonstrating scalability through Mininet emulation. Additionally, Wi-Fi is explored as a complementary technology for IoT applications, evaluating its potential to reduce latency and enhance industrial automation. This thesis offers a comprehensive analysis of TSN performance across various scenarios and its integration with complementary technologies, providing valuable insights for advancing industrial automation and connectivity within the industry 4.0 paradigm.Item Uncertainty in Recommender Systems(University College Cork, 2024) Coscrato, Victor; Bridge, Derek G.; Science Foundation IrelandRecommender Systems have emerged as a powerful tool in the information era. Due to the overwhelming number of items (products and services) currently offered on digital platforms, it is often necessary to use a system capable of ranking the items and offering those that are most relevant to each user. These systems typically use historical user-item interaction data to build models that can predict the relevance of each item to the user. There has long been a focus on increasing recommendation accuracy through the development of new prediction models. However, this is just one of the ways to improve these systems. It is also possible to equip them with new tools that extend their functionality in different ways. The tools that we focus on in this dissertation are uncertainty estimators. The problem of uncertainty is relevant to Recommender Systems in at least two ways: prediction uncertainty and label uncertainty. Prediction uncertainty is the expected imprecision of the predictions given by the system's model. Label uncertainty is the chance that interactions used to learn the prediction model are mislabeled. This dissertation reports by far the most extensive study of these two types of uncertainty, offering a varied set of methods for their estimation, ranging from heuristic data metrics to novel uncertainty prediction models. In overview, this dissertation is the largest compilation of methods for estimating prediction uncertainty and label uncertainty in Recommender Systems to date. This collection includes already-existing methods -- that we survey, rewrite in a common notation, implement, make available under an open license, and compare in-depth -- and many original methods, some that derive directly from existing work, but others that involve complex modeling. We divide our work into three branches: prediction uncertainty in explicit feedback-based systems, prediction uncertainty in implicit feedback-based systems, and label uncertainty in implicit feedback systems. While this dissertation proposes new uncertainty estimation methods, the novel work in this dissertation is not restricted to new estimation methods. We also propose new techniques for evaluating prediction uncertainty estimators. Furthermore, we present and validate novel ways of using uncertainty estimators to improve the operation of a Recommender System. At the core of our research program, and for each of the three branches cited above, we have rigorous validation of our prediction and label uncertainty estimation methods through large-scale, reproducible empirical studies on publicly available recommendation datasets that unveil important insights into the performance and usefulness of the proposed methods. These studies include both the novel and surveyed uncertainty estimation methods, and make use of the novel uncertainty evaluation techniques that we propose. This work can be an important mechanism for promoting new research on this topic that is still largely unexplored in the world of Recommender Systems. Thus, this dissertation is a contribution to the field of Recommender Systems, not just in terms of an all-encompassing compendium of uncertainty estimation methods for practitioners, but also in guiding future work. Given that the landscape of Recommender Systems continues to evolve, our work is poised to shape the discourse about uncertainty in the field.