School of Computing and Informatics Technology (CIT) Collection

Permanent URI for this collection

http://10.29.1.53:4000/handle/10570/88

Browse

Recent Submissions

Now showing 1 - 5 of 628
  • Item
    Machine learning models for short-term rainfall prediction using Uganda’s Lake Victoria Basin Weather Dataset
    (Makerere University, 2025) Tumusiime, Andrew Gahwera
    As climate change intensifies, accurate short-term rainfall forecasting has become an urgent research priority. Numerical Weather Prediction models often struggle with precipitation due to high computational requirements and large error margins. This dissertation addresses these challenges by introducing a curated multi-station dataset for the Lake Victoria Basin (LVB) and systematically evaluating Machine Learning and Deep Learning approaches for rainfall forecasting in a data-scarce African environment. Regression experiments benchmarked Random Forest, Support Vector Regression, Neural Network Regression, Least Absolute Shrinkage and Selection Operator, Gradient Boosting, and Extreme Gradient Boosting (XGBoost) Regression. Among these, XGBoost consistently achieved the lowest error, with Mean Absolute Error values as low as 0.006, 0.018, and 0.005 mm h−1 for Uganda, Kenya, and Tanzania, respectively. To complement these continuous forecasts, rainfall classification was implemented using Multi-Layer Perceptrons, Convolutional Neural Networks, and Long Short-Term Memory networks (LSTM). LSTM outperformed alternative architectures, achieving weighted F1-scores above 90% at multiple stations. To overcome data scarcity, a transfer-learning strategy was developed by fine-tuning pre-trained LSTM models from data-rich stations and applying them to the data-limited station of Kisumu, yielding performance improvements of up to 3%. An ensemble of these transfer-learned models using an Exponential Weighting Algorithm further enhanced robustness, delivering gains of up to 5% in F1-score. Overall, the dissertation demonstrates that an ensemble-based transfer-learning framework, grounded in a regionally curated dataset, can substantially improve rainfall forecasting in East Africa. The integration of regression, DL classification, and transfer learning provides methodological advancement and operational potential, contributing to more reliable weather services in data-scarce regions such as the LVB.
  • Item
    Fraud detection in mobile money transactions using differentially private machine learning techniques
    (Makerere University, 2025) Kagaba, Dennis
    Mobile money has become a vital financial tool in regions with limited banking infrastructure, fostering financial inclusion for underserved populations. However, its rapid growth has led to rising fraud threats such as refund scams, account takeovers, and identity misuse. Machine learning (ML) models show promise in detecting fraud through transaction patterns, but they rely heavily on sensitive personal data, raising concerns about privacy, security, and regulatory compliance. These risks may reduce user trust and hinder adoption. This thesis explores the use of privacy-preserving ML, specifically differential privacy to enable effective fraud detection while safeguarding user data. This study adapts privacy preserving ML models by integrating differential privacy into supervised algorithms, including Gaussian Naive Bayes, Logistic Regression, Decision Trees, and Random Forest. A labeled synthetic Mobile money transaction dataset is used to train these models, leveraging the IBM Diffpriv library to enforce differential privacy constraints. Model performance was further assessed across varying privacy levels using the differential privacy parameter e to evaluate the privacy utility trade off. The results showed that Logistic Regression experienced substantial degradation under strict privacy € = 1, with accuracy dropping to 0.496 and recall to 0.047, indicating high sensitivity to privacy noise. Meanwhile, Gaussian Naive Bayes, Decision Trees, and Random Forest remained robust, maintaining high accuracy and recall even at lower epsilon values. Across the full range of privacy budgets, Random Forest and Gaussian Naive Bayes demonstrated the most stable performance, highlighting their suitability for privacy preserving fraud detection. These results indicate that differentially private machine learning models can effectively detect frandulent transactions while preserving user privacy. The findings contribute to the development of secure, ethical and privacy sensitive fraud detection systems for mobile financial services.
  • Item
    A knowledge management portal to support the sharing of best practices in community-based obesity prevention and advocacy in Uganda
    (Makerere University, 2025) Kamukama, Diana
    Obesity has gradually become a silent epidemic in Uganda, contributing to rising rates of diet-related non-communicable diseases (NCDs) such as hypertension, type 2 diabetes, cardiovascular disease, and certain cancers leading to increased mortality. This alarming trend has placed considerable strain on Uganda‘s health system and highlights critical gaps in community-level knowledge about effective practices for preventing and managing obesity. With limited focus by the government on actions that specifically target the prevention and management of obesity, there is a growing need for increased advocacy, particularly through knowledge sharing to engage communities in physical activity and promote diet modifications for healthier lifestyles. Knowledge Management (KM) has been proven to be a valuable approach to addressing health knowledge gaps by enabling access to, sharing of, and application of evidence-based practices within communities. However, the effective implementation of KM in the obesity care continuum remains largely untapped in Uganda. This study addresses that gap by developing a Knowledge management portal for Obesity Prevention and Advocacy (KOPA), a digital platform intended to serve as a single point of access to evidence-based best practices for obesity prevention and advocacy, targeting both healthcare professionals and community members. Using an Action Research approach, this study identified key challenges and user requirements that informed the portal‘s design and implementation. KOPA combines a searchable library of best practices on healthy eating and physical activity, collaboration tools, content for different user groups, simple obesity-tracking, and clear Uganda-specific guidance on healthy lifestyle change. Evaluation with healthcare professionals and community members found it relevant, easy to use, and helpful for finding, organising, and sharing practical information that supports behaviour change. KOPA shows how a centralized, knowledge-driven platform can strengthen obesity prevention and advocacy in resource-constrained settings. Future work includes expanding multilingual content, introducing lightweight mobile access, connecting with Ministry of Health resources, and running a multi-site evaluation to confirm gains in knowledge, activity, and diet quality.
  • Item
    A hybrid deep learning model for detection and mitigation of distributed denial of service attacks in software-defined networks
    (Makerere University, 2026) Emukuny, Martin
    The increasing adoption of Software-Defined Networking (SDN) has introduced flexibility and programmability in modern networks but has also exposed the SDN controller to Distributed Denial of Service (DDoS) attacks that exploit its centralized architecture. Existing intrusion detection approaches largely focus on detection without providing effective mitigation, and many rely on outdated or non-SDN datasets, limiting their real-world applicability. Furthermore, models based on single deep learning architectures often fail to capture both the spatial and temporal characteristics of DDoS traffic, leading to high false positive and false negative rates. This study proposes a hybrid deep learning model that integrates Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks for the detection and mitigation of DDoS attacks in SDN environments. The CNN component extracts spatial correlations among flow features, while the LSTM component captures temporal dependencies, thereby enhancing spatio-temporal learning. The model was trained and evaluated using the LR-HR DDoS 2024 dataset, an SDN specific dataset containing both low-rate and high-rate attacks. Data preprocessing included exploratory analysis, feature normalization, and class balancing using Synthetic Minority Oversampling Technique (SMOTE). The model was implemented in TensorFlow and deployed in a simulated SDN environment using Mininet and the OpenDaylight controller, where detected attacks were mitigated through dynamic flow rule enforcement. Experimental results show that the hybrid CNN–LSTM model achieved an accuracy of 98.7%, a precision of 0.987, a recall of 0.987, and an F1-score of 0.987, outperforming standalone CNN and LSTM models. The model further attained ROC-AUC and PR-AUC values of 0.995 and 0.993 respectively, with confusion matrix analysis confirming reduced misclassification rates. Although the hybrid model required slightly higher computational resources, 3.2 million parameters, an average training time of 47 minutes per epoch, and an inference latency of 3.1 milliseconds, it remained practical for real-time SDN deployment. Limitations include reliance on a single dataset, evaluation within a simulated testbed, and the exclusion of additional contextual features such as topology data. Future work should focus on validating the model with multiple real-world datasets, integrating attention mechanisms or transformer architectures, optimizing for lightweight deployment, and testing resilience against adversarial attack strategies.
  • Item
    Predicting PM2.5 along road paths using deep neural networks and spatio-temporal models
    (Makerere University, 2025) Kalyesubula, Stephen
    Outdoor air pollution remains a major public-health challenge in rapidly urbanizing regions such as Uganda, where road tra!c and industrial activity generate elevated levels of fine particulate matter PM2.5. Road users—drivers, pedestrians, cyclists, and nearby communities—face heightened exposure risks, especially along congested and industrial corridors. However, the sparse and uneven distribution of sensors along these road networks makes it di!cult to understand how pollution varies along frequently used routes. The purpose of this study was therefore to develop a high-resolution approach for predicting near-road PM. in areas with limited sensor coverage by combining deep learning and geostatistical modelling. Analysis of AirQo sensor data from the Jinja (5 stations) and Kibuli (4 stations) industrial air clouds, covering April 2023 to April 2024, revealed clear di”erences in pollution levels, with Jinja consistently recording higher PM2.5 concentrations and more frequent Unhealthy–Hazardous episodes than Kibuli. To model these temporal and spatial dynamics, this study employed a Long Short-Term Memory (LSTM) neural network to predict hourly PM2.5 at sensor locations, achieving low errors across devices. The LSTM outputs were then integrated with several spatial interpolation techniques; Ordinary Kriging, Regression Kriging, Classification Kriging, Universal Kriging, and Radial Basis Function, evaluated using k-fold cross-validation across multiple variogram models. Results showed that Universal Kriging with the Exponential variogram achieved the best performance (RMSE = 1.88), outperforming all other variogram–model combinations. The final road-level maps highlighted severe pollution hotspots along industrial corridors in Jinja and mixed but elevated concentrations in Kibuli. Overall, the study demonstrates the e”ectiveness of combining deep learning and geostatistical modelling for high-resolution near-road PM2.5 prediction, providing actionable insights for targeted air-quality management in Ugandan urban environments.