Saudi Cultural Missions Theses & Dissertations
Permanent URI for this communityhttps://drepo.sdl.edu.sa/handle/20.500.14154/10
Browse
82 results
Search Results
Item Restricted Mitigating Distribution Shift and Label Noise in Deep Neural Network Based Network Intrusion Detection Systems(Saudi Digital Library, 2026) Alotaibi, Fahad Mastor T; Maffeis, SergioDeep Neural Network (DNN)-based Network Intrusion Detection Systems (NIDS) have shown strong performance on offline benchmarks under controlled conditions. However, real-world deployment remains challenging. This thesis focuses on two key challenges: (i) distribution shift, in which benign traffic evolves and attackers continuously devise new strategies; and (ii) label noise, arising from imperfect automated labelling of large volumes of network traffic. To address these challenges, the thesis first provides a critical literature review and develops a structured taxonomy of techniques for handling distribution shift in NIDS and label noise in both NIDS and malware datasets. Building on these insights, it introduces three frameworks. Mateen adapts one-class anomaly detection to evolving benign traffic by combining selective labelling with an ensemble-based mechanism that identifies and responds to shifts with minimal manual effort. Rasd extends multi-class NIDS to detect and integrate newly emerging attack classes, substantially reducing labelling costs through strategic selection of a small, informative, and diverse subset. SLB mitigates label noise by partitioning the dataset into clean and noisy sets and iteratively refining both the model and the labels. Each framework is evaluated extensively across multiple datasets and compared with state-of-the-art baselines. Mateen improves the anomaly-detection F1 score by 4.13% under a light-shift scenario (CICIDS2017) and by 72.6% under an extreme-shift scenario (Kitsune). Rasd increases the novel class detection F1 score by 6.83% on CICIDS2017 and by 19.21% on CSE-CIC-IDS2018. SLB reduces the noise rate in CICIDS2017 (with 30% injected random noise) to below 1.2% and outperforms the vanilla baseline by 11.83% in macro F1. This thesis serves as a reference for researchers and practitioners in cyber security and artificial intelligence. Beyond its literature review and taxonomy, it contributes three frameworks that collectively enhance the robustness of DNN-based NIDS, achieving state-of-the-art results on the evaluated benchmarks.8 0Item Restricted Predicting Carbon Credit Prices Using Advanced Machine Learning Techniques(Saudi Digital Library, 2026) Rayan, Najdi; Wang, HaiAccurate forecasting of carbon credit prices supports risk management, investment decisions, and policy assessment in the context of climate action. EU ETS carbon prices exhibit volatility, non-linearity, and non-stationarity, which reduces the effectiveness of traditional forecasting models. This dissertation proposes and evaluates a three-stage hybrid machine learning model for one-day-ahead forecasting of EU Emissions Trading System (EU ETS) carbon prices. The architecture follows a divide-and-conquer strategy. First, Wavelet Packet Decomposition (WPD) decomposes the carbon price signal into multiple frequency components. Second, a Gated Recurrent Unit (GRU) network models temporal dependencies and forecasts the trend component. Third, an Extreme Gradient Boosting (XGBoost) model predicts and corrects the GRU residual errors using wavelet-derived detail components as input features. The model was trained and tested on a dataset covering January 2018 to December 2024. The dataset includes EU ETS carbon prices, Brent crude oil prices, and electricity prices, while the forecasting model is univariate and uses the carbon price series only. On an unseen test set of 510 days, the model achieved a Mean Absolute Percentage Error (MAPE) of 1.66%, a Root Mean Squared Error (RMSE) of 4.86 EUR/ton, and a Mean Absolute Error (MAE) of 4.41 EUR/ton. The results indicate that combining signal decomposition, deep learning, and gradient boosting provides stable forecasting performance for EU ETS carbon prices under realistic evaluation conditions.6 0Item Restricted DEEP LEARNING-BASED SEGMENTATION FOR PRECISION RADIATION THERAPY IN BREAST CANCER(Saudi Digital Library, 2025) Alanazi, Hamdah; Wazir Muhammad, PhDBreast cancer is a major health burden, and clinicians need accurate tumor segmentation to deliver radiation therapy precisely and efficiently. This thesis bench- marks two three-dimensional (3D) deep learning architectures U-Net and SegResNet for automated segmentation of breast tumors on dynamic contrast-enhanced MRI. This work uses the MAMA-MIA benchmark, a (large-scale multicenter dataset for developing and evaluating artificial intelligence (AI) models for breast cancer imag- ing). MAMA-MIA consist of 1,506 breat cancer subjects. We applied a standardized Medical Open Network for AI (MONAI) preprocessing and training pipeline to build and evaluate deep-learning models for medical imaging. Models were assessed with the Dice Similarity Coefficient (DSC), Intersection over Union (IoU), overall accuracy, and the 95th-percentile Hausdorff distance (HD95), alongside qualitative visualiza- tions and Bland–Altman analyses. U-Net achieved DSC 0.7334, IoU 0.5791, accuracy 0.9984, HD95 33.13 mm, loss 0.0836, and 333.6 s/epoch over 60 epochs. SegResNet achieved DSC 0.7132, IoU 0.5542, accuracy 0.9981, HD95 37.58 mm, loss 0.0915, and 546.1 s/epoch over 60 epochs. Our results show that, U-Net achieved higher overlap and boundary metrics than SegResNet. These findings are preliminary and limited to tumor masks on this dataset; no external validation, user study, or clinical deployment was performed.11 0Item Restricted Analysing Large-Scale Attacks in IoT Environments using ML/DL(Saudi Digital Library, 2025) Bokhari, Mohammed Ibrahim K; Neetesh, SexenaThe fine-grained classification of malicious network traffic presents a significant and persistent challenge in cybersecurity, primarily due to the extreme class imbalance inherent in real-world network data. Conventional machine learning approaches, which apply a single, unitary model to the problem, have demonstrated limited success, often failing to effectively identify rare but critical minority attack classes. This dissertation argues that the conventional model paradigm is fundamentally flawed for this problem space and proposes a hierarchical, multi-stage classification framework as a more robust alternative. This research presents a comprehensive, multi-faceted investigation into this problem, using the 34-class CICIoT2023 dataset as a benchmark. The study was conducted across four distinct experimental paths, comparing two ensemble methods (XGBoost and Random Forest) and two class-handling strategies (a "Grouped" approach that manually merges similar classes and an "Ungrouped" approach that tackles all 34 classes directly). Within this structure, we designed and implemented a 4-tier hierarchical framework that employs a "divide and conquer" strategy, using an initial classifier to handle majority traffic and a class-level routing mechanism to delegate ambiguous samples to specialised recovery tiers. An adaptive resampling strategy was deployed within these tiers, concentrating aggressive SMOTE only where required. The empirical results provide a holistic validation of the proposed architecture. The optimal configuration—an Ungrouped, XGBoost-led hierarchical framework—achieved a final accuracy of 0.9228 and Macro-F1 score of 0.7948, a substantial improvement over all other experimental paths and conventional baselines. More significantly, this approach demonstrated a more than 800% increase in the F1-score for some of the under-represented minority classes. The analysis also revealed a key architectural principle: classifier performance is role-dependent, with different ensemble methods excelling in different roles within the hierarchy, highlighting the importance of managing the bias-variance trade-off at a systemic level. Finally, this work provides a rigorous, data-centric analysis that distinguishes between model limitations and the inherent limitations of the dataset, identifying a "dataset-induced ceiling" on performance for 5 of the 34 classes. The primary contribution of this dissertation is, therefore, a methodologically robust and architecturally novel framework, validated through a comprehensive, multi-path experimental design. The principles of hierarchical decomposition and adaptive resource allocation are domain-agnostic and offer a promising direction for future research into extreme imbalance problems.10 0Item Restricted Context-Aware Fake News Detection Using Deep Learning(Saudi Digital Library, 2026) Alghamdi, Jawaher; Suhuai, Luo; Yuqing, LinThe pervasive spread of fake news (i.e., intentionally misleading information presented as legitimate news) on social media poses a pressing global challenge, distorting public discourse and undermining societal trust. Fake news is often crafted to imitate credible sources and exploit human judgment while leveraging the ways social media prioritizes and disseminates content. This thesis investigates fake news detection (FND) through deep learning (DL) approaches that extend beyond content-only analysis to incorporate contextual and behavioral signals. The primary objective is to develop scalable, accurate, and interpretable (i.e., capable of providing transparent and human-understandable reasoning for predictions) FND systems that can operate effectively across diverse domains and languages. Early detection methods relied heavily on surface-level or manually engineered content features, which proved insufficient for identifying sophisticated forms of fake news. To address these limitations, this thesis introduces hybrid architectures that combine transformer-based encoders with convolutional, recurrent, and attention mechanisms, integrating semantic, linguistic, stylistic, and context-aware features. These models capture both fine-grained textual cues and broader semantic relationships spanning entire articles, enhancing the ability to differentiate factual content from misleading narratives. Beyond textual signals, the work incorporates user metadata, posting histories, temporal dynamics, and relationships across headlines, article bodies, and comments, strengthening detection in situations where textual evidence alone is ambiguous. Scalability is approached through parameter-efficient fine-tuning, where lightweight adapters are integrated into transformer models. A novel fusion mechanism balances general pre-trained knowledge with task-specific information, while selective layer-wise adaptation focuses on semantically critical layers to maximize efficiency. To improve multilingual coverage, the thesis proposes a hybrid summarization technique that distills salient information and reduces redundancy, enabling robust performance without dependence on translation pipelines. This framework operates effectively in both high-resource and low-resource languages, supporting inclusive and scalable deployment. The thesis also tackles cross-domain generalization, introducing domain-adaptive models that extend FND to diverse topics and formats. One approach reframes FND as a prompt-based zero-shot learning (ZSL) task, employing cloze-style prompts and domain-aware augmentation to improve adaptability without fine-tuning. A second approach employs domain-specific expert subnetworks enriched with topic and entity information, combined with adversarial learning to mitigate domain bias. Experiments on multiple benchmark datasets confirm the effectiveness of these methods, achieving state-of-the-art (SOTA) performance. Results demonstrate that the integration of content, context, and propagation-aware signals substantially enhances FND. In summary, this thesis delivers a suite of DL solutions that advance efficiency, robustness, and adaptability in FND. The work addresses scalability, multilinguality, and cross-domain resilience, offering practical and deployable tools for countering the global threat of fake news across platforms, domains, and languages.6 0Item Restricted Explainable AI for Biometric Brain Modeling: 3D Anatomical Analysis Across Gender and Aging(Saudi Digital Library, 2025) Alghamdi, Ryan; Jiang, RichardThis thesis investigates the use of deep learning models for predicting sex and age from structural MRI data, with a focus on interpretability. Three architectures were trained: a ResNet, a 3D CNN, and an Ensemble model. The Automated Anatomical Labeling (AAL) atlas was used to parcellate the brain into 116 regions, enabling a region based occlusion framework. Two complementary approaches were applied: inverse occlusion, where predictions rely on a single active region, and normal occlusion, where one region is masked while the remainder of the brain is preserved. All models achieved high training accuracies (95–99%), but their performance dropped notably on held out 70/15/15 splits, reflecting overfitting and emphasizing the importance of split evaluation. Inverse occlusion consistently identified plausible neuroanatomical markers, including the cerebellar Crus I, calcarine cortex, precentral and postcentral gyri, precuneus, and inferior temporal lobe. In contrast, normal occlusion produced flat or inconsistent results, suggesting reliance on global artifacts or scanner fingerprints rather than region specific features. These findings show that region only occlusion provides more reliable insights into localized brain structure differences than conventional occlusion. Key limitations include dependence on preprocessing pipelines, restricted dataset size, computational demands, and reliance on atlas based parcellation. Despite these, the framework demonstrates a reproducible method for evaluating the regional basis of deep learning predictions in neuroimaging. Future work should expand dataset coverage, refine preprocessing, and extend occlusion analysis to combinations of regions to capture network level effects. This work contributes a regional explainability framework to improve the interpretability and reliability of deep learning in brain imaging.26 0Item Restricted Advances in Artificial Intelligence for Energy Forecasting and Performance Management in Buildings(Saudi Digital Library, 2026) Alkhatani, Nasser; Petri, IoanAccurate energy forecasting is essential for intelligent building management, supporting operational optimisation, strategic planning, and demand-side flexibility. However, existing forecasting methods often struggle to remain accurate across multiple time horizons and to generalise across different building types with limited data. This thesis addresses these challenges by developing a modular modelling framework that advances both multi-horizon forecasting and cross-building adaptability. The first contribution is a hybrid forecasting model (SVR → XGBoost → LSTM) designed to deliver stable prediction performance across four horizons: 24 hours, one week, one month, and one year. The hybrid design leverages the complementary strengths of its components SVR for noise reduction, XGBoost for nonlinear feature learning, and LSTM for long-range temporal modelling resulting in improved robustness and generalisation compared with single-model approaches. The second contribution introduces a deep hybrid model (CNN → GRU → LSTM) within a transfer learning framework. Pretrained on multi-building datasets and fine-tuned using limited data from new buildings, this approach enhances cross-domain adaptability while reducing training time and data requirements, demonstrating the practical value of transfer learning for scalable energy forecasting. A third contribution integrates statistical peak detection to support the identification of high- demand events, enabling forecasting outputs to inform grid-interactive building operations. Rigorous evaluation including multi-metric assessment, residual diagnostics, ablation testing, and statistical significance analysis confirms the reliability and robustness of the proposed models. Overall, the thesis provides methodological and empirical advances that strengthen data-driven building energy management. The results show that hybridisation and transfer learning, when carefully designed, can enhance accuracy, stability, and generalisation, offering a scalable pathway toward more efficient and sustainable smart building operations.16 0Item Restricted GAN-Enhanced Super-Resolution Pipeline for Robust Word Recognition in Low-Quality Non-English Handwriting.(Saudi Digital Library, 2025) Shahbi, Bilqis; Xia, PanqiuExecutive summary The dissertation tackles a critical issue where current optical character recognition (OCR) technologies fall short: correctly identifying handwritten non-English scripts in poor quality and deteriorated settings. While OCR technologies have matured for printed English and other Latin-based languages, scripts such as Arabic, Devanagari, and Telugu remain underrepresented due to structural complexities, cursive connections, diacritics, ligatures, and the limited availability of annotated datasets. These challenges are amplified by real-world factors such as low-resolution scans, noisy archival documents, and mobile phone captures, where fine details necessary for recognition are lost. The study proposes a two-stage deep learning pipeline that integrates super-resolution with recognition, explicitly designed to address these shortcomings. The first stage of the pipeline utilises Real-ESRGAN, a generative adversarial network specifically optimised for real-world image degradation. Unlike earlier models such as SRCNN, VDSR, and ESRGAN, which often prioritize aesthetics or hallucinate features, Real-ESRGAN reconstructs high-resolution images with sharper strokes, preserved ligatures, and clear diacritics. Its Residual-in-Residual Dense Block (RRDB) architecture combines residual learning and dense connections, enabling robust recovery of fine-grained textual details. By preserving structural fidelity rather than merely visual appeal, Real-ESRGAN ensures that enhanced images retain the critical features necessary for recognition. The second stage utilises a Convolutional Recurrent Neural Network (CRNN) with Connectionist Temporal Classification (CTC) loss function. The CRNN combines convolutional layers for feature extraction, bidirectional LSTM layers for capturing sequential dependencies, and CTC decoding for alignment-free sequence prediction. This integration eliminates the need for explicit segmentation, a complicated task in cursive or densely written scripts. Together, the two stages form a cohesive system in which image enhancement directly supports recognition accuracy. To ensure robustness, the research incorporated extensive dataset preparation and preprocessing. Handwritten datasets for Arabic, Devanagari, and Telugu scripts were selected to reflect structural diversity. Preprocessing included resizing, artificial noise simulation (Gaussian noise, blur, and compression artefacts), and augmentation (rotations, elastic distortions, and brightness adjustments). These techniques increased dataset variability and improved the model's ability to generalize to real-world handwriting scenarios. Evaluation was conducted at both image and recognition levels. Image quality was assessed using the Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity Index (SSIM). At the same time, recognition performance was measured using the Character Error Rate (CER) and the Word Error Rate (WER). This dual evaluation ensured that improvements in image clarity translated into tangible recognition gains. The results confirm the effectiveness of the proposed pipeline. Real-ESRGAN showed improvements compared to SRCNN, VDSR, and ESRGAN, with higher PSNR and SSIM values across scripts. These gains reflect superior structural fidelity, particularly in preserving Arabic cursive flows, Devanagari's horizontal headstrokes, and Telugu's stacked ligatures. Recognition accuracy also improved compared to baseline low-resolution inputs. Script-specific analysis showed more precise word boundaries in Arabic, sharper conjuncts and diacritics in Devanagari, and more distinct glyph separations in Telugu. When benchmarked against traditional OCR systems, such as Tesseract, the pipeline demonstrated clearer recognition outcomes, indicating the critical role of task-specific super-resolution in OCR, proving that enhancing input quality directly strengthens recognition performance. The dissertation makes contributions across methodological, empirical, theoretical, and practical domains. Methodologically, it demonstrates the value of integrating enhancement and recognition stages in a fine-tuned pipeline, ensuring that improvements in image clarity yield measurable gains in recognition. Empirically, it validates the effectiveness of Real-ESRGAN for handwritten text, showing consistent improvements across structurally diverse scripts. Theoretically, it advances the understanding of script-sensitive OCR, emphasizing the preservation of structural features such as diacritics and ligatures. Practically, the work highlights applications in archival preservation, e-governance, and education. By enabling more accurate digitisation of handwritten records, the system supports inclusive access to information and the preservation of linguistic heritage. The study acknowledges several limitations. The scarcity of diverse annotated datasets constrains the model's generalizability to other scripts such as Amharic or Khmer. The computational expense of training Real-ESRGAN limits its feasibility in low-resource settings. Occasional GAN artefacts, where spurious strokes or distortions appear, pose risks in sensitive applications such as legal documents. Moreover, the pipeline has not been extensively tested on mixed-script texts, common in multilingual societies. These limitations suggest avenues for future work, including developing larger and more diverse datasets, designing lightweight models for real-time and mobile deployment, integrating script identification for mixed-language documents, and incorporating explainable AI for greater transparency in recognition decisions. In conclusion, the dissertation demonstrates that GAN-enhanced super-resolution is not merely a cosmetic tool but an essential step toward robust OCR in non-English handwritten texts. By aligning image enhancement with recognition objectives, the proposed pipeline reduces error rates and generalizes across diverse scripts. Its implications extend beyond technical achievement to cultural preservation, digital inclusion, and the democratization of access to information. At the same time, the identified limitations provide a roadmap for future research, ensuring that multilingual OCR evolves into a truly global and inclusive technology.22 0Item Restricted CADM: Creative Accounting Detection Model in Saudi-Listed Companies(Saudi Digital Library, 2025) Bineid, Maysoon Mohamed; Beloff, NataliaIn business, financial statements are the primary source of information for investors and other stakeholders. Despite extensive regulatory efforts, the quality of financial reporting in Saudi Arabia still requires improvement, as prior studies have documented evidence of creative accounting. This practice occurs when managers manipulate accounting figures within the boundaries of the International Financial Reporting Standards to present a more favourable image of the company. Although various fraud detection methods exist, identifying manipulations that are legal yet misleading remains a significant challenge. This research introduces the Creative Accounting Detection Model (CADM), a deep learning (DL)-based approach that employs Long Short-Term Memory (LSTM) networks to identify Saudi-listed companies engaging in creative accounting. Two versions of the model were developed: CADM1, trained on a simulated dataset based on established accounting measures from the literature, and CADM2, trained on a dataset tailored to reflect financial patterns observed in the Saudi market. Both datasets incorporated financial and non-financial features derived from a preliminary survey of Saudi business experts. The models achieved training accuracies of 100% (CADM1) and 95% (CADM2). Both models were then tested on real-world data from the Saudi energy sector (2019–2023). CADM1 classified one company as engaging in creative accounting, whereas CADM2 classified all companies as non-creative but demonstrated greater stability in prediction confidence. To interpret these results, a follow-up qualitative study involving expert interviews confirmed CADM as a promising supplementary tool for auditors, enhancing analytical and oversight capabilities. These findings highlight CADM’s potential to support regulatory oversight, strengthen auditing procedures, and improve investor trust in the transparency of financial statements.17 0Item Restricted Artificial Intelligence, Deep Learning, and the Black Box Opacity: International Law and Modern Governance Framework for Legal Compliance and Individual Responsibility(Saudi Digital Library, 2025) Aloqayli, Muhannad Khalid; Linarelli, JohnThis dissertation examines the unprecedented challenges that deep learning models in artificial intelligence pose to international humanitarian law frameworks governing armed conflict, addressing critical questions about international humanitarian law compliance capabilities, legal personality under the framework of international law and international humanitarian law, and international individual criminal responsibility when autonomous weapons systems employ deep learning models in decision-making processes. Chapter Two provides a comprehensive technical analysis of deep learning architectures, including convolutional neural networks, recurrent neural networks, generative adversarial networks, and transformer networks, and their military applications in target recognition, threat assessment, and autonomous operations. The analysis demonstrates that properly trained deep learning systems can achieve exceptional accuracy in tasks relevant to the principles of distinction and proportionality. However, this technical capability exists alongside a fundamental limitation: the “black box challenge,” whereby decision-making processes emerge from statistical pattern recognition across billions of parameters in ways that remain incomprehensible to human operators, creating unprecedented challenges for legal compliance and individual responsibility. Chapter Three evaluates whether granting legal personality to advanced artificial intelligence could address emerging responsibility gaps. Applying the analytical pragmatic approach through dual criteria of “value context” and “legitimacy context,” the analysis reaches definitive negative conclusions. Granting artificial intelligence legal personality would contradict international humanitarian law’s human-centered foundations, fail to fill responsibility gaps, and potentially shield humans from liability while introducing conceptual incoherence into established normative structures. Chapter Four demonstrates that deep learning, as a black box model in statistical learning, fundamentally challenges traditional international frameworks for individual criminal responsibility. The analysis reveals structural incompatibilities between algorithmic opacity and the requirements of the Rome Statute for mens rea and actus reus. Similarly, command responsibility doctrines face parallel challenges when commanders possess formal control over systems whose decision-making processes transcend human comprehension. The dissertation proposes a modified command responsibility framework recognizing commanders as “AI enablers” rather than traditional superiors, establishing reasonable governance standards for controlled environments while imposing strict liability for high-risk deployments. This framework preserves meaningful accountability while acknowledging technological constraints, shifting focus from comprehending opaque statistical processes to governing deployment decisions and operational contexts within commanders’ control.62 0