Department of Computer Engineering and Informatics

Permanent URI for this collection

https://bdears.busitema.ac.ug/handle/123456789/761

Browse

Now showing 1 - 20 of 42

Machine learning-based intrusion detection framework for digital forensic readiness.
(Busitema University, 2026-09) Isabirye, Aggrey
University networks face a structural security challenge: machine learning-based intrusion detection systems apply strict data minimisation, discarding non-alert data to preserve throughput, whilst digital forensic readiness demands comprehensive retention of all digital artefacts, enriched with metadata, secured cryptographically, and managed under a documented chain of custody. No existing framework has resolved this incompatibility within a single architecture. This study designed and validated a machine learning-based intrusion detection framework for digital forensic readiness that resolves this tension through an asynchronous forensic artefact capture layer operating in parallel with the detection pipeline. The framework employs two well-established machine learning classifiers that have been extensively applied to benchmark intrusion detection datasets by multiple independent research teams prior to this study: Random Forest, which has been widely used on the UNSW-NB15 dataset and achieved 99.67% accuracy and F1-score in this study, and Long Short-Term Memory (LSTM), which has been widely used on the NSL-KDD dataset and achieved 97.43% accuracy and F1-score in this study. No novelty is claimed in the choice of algorithm or dataset. The contribution of this study is the architectural framework that integrates these established classifiers with embedded forensic readiness mechanisms, and the empirical demonstration that detection performance is not compromised when forensic logging is introduced as an embedded architectural layer. Design Science Research methodology guided the study. Validation employed empirical machine learning evaluation, architectural consistency analysis against twelve formally specified requirements, and scenario-based evaluation across three operational scenarios. All twelve requirements were satisfied. Key quantitative metrics include a detection-to-logging latency of 8.3 milliseconds, storage overhead of 4.7 kilobytes per event, evidence retrieval time of 1.2 seconds, and 100 per cent completeness and integrity rates across all scenarios. To the knowledge of the authors, based on the literature reviewed in this study, the framework is the first formally specified architecture to simultaneously address machine learning-based intrusion detection with concretely specified and empirically eval machine learning-based intrusion detection with concretely specified and empirically evaluated classifiers, pre-incident forensic readiness, and evidence integrity management within a single unified architectural specification. Keywords: intrusion detection systems; digital forensic readiness; machine learning; Random Forest; LSTM; UNSW-NB15; NSL-KDD; Design Science Research; evidence integrity; chain of custody; university networks
Development of a machine learning model for prediction of uroflowmetry parameters and early diagnosis of lower urinary tract disorders.
(Busitema University, 2026) Balayo, Ignatius
This thesis develops and validates a machine learning approach for a non-invasive acoustic uroflowmetry system. The study serves as a proof of concept and lays the groundwork for a larger research project that will collect acoustic uroflowmetry data in Uganda. The research focused on three main objectives: (i) developing methods to process urinary sound recordings and comparing traditional machine learning with deep learning models for predicting urinary flow rate, (ii) creating a sequence-based model to reconstruct urinary flow curves and a classifier to support the early detection of lower urinary tract symptoms (LUTS), and (iii) evaluating all models using the same performance measures to ensure a fair comparison. The study used two clinically labeled datasets provided through a research collaboration with the University of Valladolid. The first dataset contained recordings from 50 patients collected simultaneously using a professional microphone, a smartphone, and a smartwatch. It was used to develop and compare flow-rate prediction models. The second dataset included 110 patients with confirmed clinical diagnoses and was used to build and evaluate the complete diagnostic model. For urinary flow-rate prediction, traditional machine learning models, including Random Forest, Support Vector Regression (SVR), and XGBoost, were compared with deep learning models, including a multilayer perceptron (MLP) and a convolutional neural network (CNN). The traditional models achieved the highest coefficients of determination, with 𝑅2 values ranging from 0.681 to 0.795, demonstrating strong agreement with the clinical measurements. However, the CNN achieved the lowest mean absolute errors, ranging from 2.43 to 2.69 mL/s. These findings suggest that neither approach consistently outperformed the other, as each offered different strengths depending on the evaluation metric. An important finding was that recordings captured using smartphones and smartwatches performed as well as, and in some cases better than, those recorded using the research-grade microphone. This demonstrates that reliable acoustic uroflowmetry can be achieved with affordable, widely available mobile devices and wearables. For diagnostic prediction, a deep learning model combining a convolutional encoder with a bidirectional long short-term memory (BiLSTM) decoder was developed to reconstruct urinary flow curves directly from acoustic recordings. Clinically relevant uroflowmetry parameters, including maximum flow rate (𝑄max), average flow rate (𝑄avg), voided volume (𝑉𝑉), and flow duration (𝑇flow), were extracted from the reconstructed curves and used to train a Random Forest classifier. Using five-fold stratified cross-validation, the classifier achieved an overall accuracy of 76.4% in distinguishing pathological from healthy voiding patterns. The model showed stronger performance in identifying pathological cases, achieving a recall of 0.809 compared with 0.620 for healthy cases. This behavior is appropriate for an initial screening tool, where reducing missed diagnoses is generally more important than maximizing specificity. Overall, this study demonstrates that it is technically feasible to build a complete, lowcost acoustic uroflowmetry pipeline, from audio recording and urinary flow estimation to diagnostic classification, using clinically labeled data. The findings show that consumer devices such as smartphones and smartwatches can provide reliable performance, making the technology more accessible for routine clinical use. The results also highlight that both 𝑅2 and mean absolute error should be considered when selecting predictive models, since each metric reflects a different aspect of performance. Although further clinical validation is required before the system can be deployed in routine healthcare, this work provides a strong foundation for the next phase of the research. Future work will focus on collecting a representative acoustic uroflowmetry dataset in Uganda, adapting the models to the local clinical environment, and improving access to affordable screening tools for lower urinary tract disorders in resource-constrained healthcare settings.
A structured framework for integrating cyber physical systems into smart building management in Uganda.
(Busitema University, 2026) Kajjumba, Kyakulaga Mubarak
This study developed and validated a Cyber-Physical Systems (CPS) framework for Smart Building Management Systems (SBMS) integration in Uganda. The study used a Design Science Research approach and mixed methods to design and test a prototype that integrates sensors, communication networks, edge computing, intelligent control, and security mechanisms. The results revealed that the framework enabled effective real-time monitoring, automated operations, energy management, and improved system responsiveness. In addition, edge computing increased system reliability over network outages. Expert assessment showed high technical feasibility (mean = 4.6) and practical applicability (mean = 4.5). The study concludes that a context-specific CPS framework can provide a pragmatic basis for addressing interoperability, connectivity, cybersecurity, and automation challenges in Uganda’s smart building sector.
Development of an agentic AI framework for power infrastructure monitoring and resilience.
(Busitema University, 2026) Bwire, John Bosco
Monitoring critical infrastructure (CI), particularly national power distribution networks in developing countries such as Uganda, remains a significant challenge due to the limited deployment, high operational costs, and maintenance requirements of conventional monitoring technologies, including Supervisory Control and Data Acquisition (SCADA) systems, field sensors, and roadside surveillance infrastructure. These limitations often result in inadequate realtime situational awareness during service disruptions, delaying incident detection and response. Meanwhile, social media platforms such as X (formerly Twitter) have emerged as valuable sources of real time, and user generated information that can provide rapid insights into infrastructurerelated events. However, the unstructured, noisy, multilingual, and context-dependent nature of social media data presents significant challenges for traditional artificial intelligence and data analytics approaches. This study proposes a fully automated situational awareness framework that leverages Generative Artificial Intelligence, specifically Large Language Models (LLMs), to extract, analyze, and synthesize actionable information from social media streams. The proposed pipeline integrates Twitter API-based data collection, LLM driven relevance filtering and event classification, geolocation inference for identifying affected areas, and visualization tools for presenting incident information in an accessible format. By combining natural language understanding with automated information extraction, the system transforms dispersed social media posts into meaningful intelligence that can support infrastructure operators and emergency response teams. A case study focusing on power disruption monitoring in Uganda is used to evaluate the effectiveness of the proposed approach. Results demonstrate the potential of LLMs to identify and characterize outage related events from social media data with minimal human intervention. The study also addresses key challenges, including ambiguous location references, local language expressions, informal communication styles, and variations in outage-related terminology. The findings suggest that the proposed framework offers a scalable, cost effective, and adaptable solution for enhancing real time monitoring and situational awareness of critical infrastructure in resource constrained environments
An explainable AI-based digital forensics model for image tampering detection in Uganda’s judicial system.
(Busitema University, 2026) Bageya, Ivan
The growing dependence on digital image evidence in Ugandan legal procedures has revealed a key flaw for example the lack of standardized mechanisms for ensuring image integrity and authenticity. Without cryptographic hashing, chain-of-custody procedures, or forensic tools, digital evidence is vulnerable to alteration, jeopardizing judicial outcomes and public trust. This problem is particularly serious in low-resource countries like Uganda, where technological infrastructure lags behind the increasing use of digital evidence in court. This study sought to close this gap by creating and testing an AI-based tamper detection model specific to Uganda's forensic environment. A mixed-methods research strategy was used, which included quantitative surveys and qualitative interviews with legal professionals, forensic experts, and law enforcement officials (n=10), who confirmed the lack of formal verification procedures and voiced conditional support for AI deployment. As a result, a convolutional neural network (CNN) model was designed and tested on a dataset of 12,617 image pairs containing legitimate and tampered images, yielding 90.3% accuracy, 83.5% precision, 94.8% recall, and an AUC of 0.969. This performance exhibits remarkable reliability in distinguishing between modified and legitimate images. Furthermore, the model included Explainable AI (XAI) via Grad-CAM visualization to improve transparency and legal defensibility, addressing the black-box issue widespread in forensic AI applications. These findings provide a practical, evidence-based approach that connects technical feasibility and judicial reality, improving the dependability of digital evidence, promoting legal reforms, and sustaining institutional trust in Uganda's expanding digital forensics environment.
Deep learning-based surgical phase recognition and interpretable workflow anomaly detection in laparoscopic cholecystectomy.
(Busitema University, 2026) Opio, Jonathan
This thesis presents an interpretable deep learning approach for analysing surgical workflows and identifying statistically unusual patterns in laparoscopic cholecystectomy videos. The proposed approach uses a two-stage architecture that combines deep learning-based surgical phase recognition with probabilistic normative modelling. In the first stage, a multimodal temporal architecture is developed by integrating ResNet-50 visual features, instrument information derived from the CholecT50 annotations, a Bidirectional Gated Recurrent Unit (BiGRU) encoder, and a temporal attention mechanism. The model recognises surgical phases from short video windows together with instrument signals. Using a 16-frame sliding window and class-weighted cross-entropy loss, the selected attention-based model achieved a validation accuracy of 74.16%. The selected model was subsequently evaluated on the held-out test set and achieved an accuracy of 72.43%, which provides the principal measure of model generalisation. In the second stage, two normative models are learned from the training data to represent statistically common workflow patterns. The first captures phase-transition probabilities to characterise the expected temporal ordering of surgical phases, while the second captures phase-instrument relationships to characterise commonly observed instrument usage patterns. Deviation thresholds are calibrated from the corresponding training distributions using percentile-based procedures. The transition and semantic components use lower-tail thresholds, while the combined window-level workflow-deviation score uses a 95th-percentile threshold. On the Stage 2 validation data, the approach flagged 4.86% of validation records when ground-truth phase labels were used and 7.44% of validation records when predicted phases were used. These results describe the behaviour of the statistical flagging procedure under the calibrated thresholds and provide empirical evidence that phase-recognition errors can propagate into subsequent workflow-deviation flagging. For each flagged workflow event, the approach provides structured information including the video identifier, frame number, deviation score, component-wise flags, and an interpretable explanation such as a rare phase transition or an expected instrument being absent. These explanations decompose the statistical deviation into temporal and semantic components, supporting transparent inspection of the model’s findings and subsequent expert review. Importantly, the proposed approach identifies statistically unusual workflow patterns relative to patterns learned from the training data. Flagged events are candidates for expert review and should not be interpreted as clinically verified surgical errors, complications, negligence, misconduct, or other forms of wrongdoing. Clinical and domainexpert validation is required before a flagged event can be considered a genuine clinical or procedural anomaly. The approach is therefore intended to support post-hoc workflow analysis, structured auditability, and forensic review rather than to independently establish clinical or legal conclusions
A machine learning model for vertical total electron content forecasting and global navigation satellite system ionospheric delay estimation across selected African stations.
(Busitema University, 2026) Nantale, Tracy Cynthia
Global Navigation Satellite System (GNSS) users in equatorial and low-latitude Africa experience substantial ionospheric delay, yet few regional tools link forecasting to practical delay estimation. This study developed and evaluated a machine-learning-based tool for 24-hour Vertical Total Electron Content (VTEC) forecasting and GNSS ionospheric delay estimation across a network of 31 selected African GNSS stations, using a regional multi-station data set covering the years from 2000 to 2025. Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) architectures were compared across full-feature, exogenous-only, and VTEC-only input configurations. Recent VTEC history was far more informative than the space-weather inputs alone (validation RMSE of approximately 1.91 TECU for theVTEC-onlyGRUversus2.33TECUfor the exogenous-only GRU), and the lighter VTEC-only models matched the full-feature models while needing fewer inputs. A 48-hour lookback window was retained as the preferred operational compromise, and the VTEC-only GRU was selected over the tuned LSTM under a predefined validation criterion, the two performing similarly rather than one being clearly superior. On 446,738 chronological test sequences (10,721,712 forecast-target pairs across the 24 output horizons), the final GRU achieved a Root Mean Square Error (RMSE) of 3.9108 TEC Units (TECU), a Mean Absolute Error (MAE) of 2.5523 TECU, an 𝑅2 of 0.9456, and a correlation coefficient of 0.9724, improving on previous-day persistence by 17.0% and on the International Reference Ionosphere 2020 (IRI-2020) by 48.0%. This advantage held across all forecast horizons and ionospheric conditions, though errors rose when the ionosphere was disturbed, corresponding to a GPS L1 vertical delay-error equivalent of 0.6350 m. The selected GRU was integrated into a prototype that uses the latest 48 hours of station specific VTEC to generate 24-hour VTEC and delay forecasts, providing a foundation for future regional GNSS ionospheric monitoring and delay-estimation services in Africa.
Digital forensics framework for ensuring integrity and traceability of AI-based medical imaging : case study, breast cancer imaging data.
(Busitema University, 2026) Omiel, Lukas
Breast cancer remains a serious worldwide health challenge and one of the leading causes of death in females. While Artificial Intelligence (AI) based medical imaging supports early disease detection, its reliability is currently compromised by lack of mechanisms to validate the integrity and traceability of images processed within the AI pipelines. This study addresses these issues by integrating cryptographic watermarking and hashing mechanism in a forensic by-design framework that performs AI-based medical image classification. The proposed framework performs content and context tamper detection as well as image classification through transfer learning using a pretrained Resnet50 Convolutional Neural Network (CNN). With a Peak Signal-to-Noise Ratio (PSNR) of 52.71 dB and a Structural Similarity Index (SSIM) of 0.9997, forensic evaluation shows that the framework guarantees tamper detection without affecting diagnostic quality. The maximum Integrity Detection Rate (IDR) of 100% was attained by the security layer alert, which also had the lowest False Acceptance Rate (FAR) and False Rejection Rate (FRR) of 0%. Additionally, the findings of the experiment show a classification performance with an accuracy of 91%, a high Area Under the Curve (AUC) of 96.4%, and a benign recall of 94%, providing sensitivity in recognizing tissue that is not malignant. These findings show resilience to integrity violations during image classification. In contexts with limited resources, our hybrid method ensures that diagnostic data is authentic from the time of capture to the final clinical decision by providing the required foundation for AI forensics.
Forensic framework for insider threat detection and mitigation in corporate networks.
(Busitema University, 2026) Kakaire, Godfrey
Insider threats present severe risks to corporate networks due to the authorized access privileges and contextual knowledge of internal users. This study designed and evaluated an integrated, proactive forensic framework using Design Science Research Methodology (DSRM) to enhance early threat detection and mitigation in Ugandan corporate networks focusing on banking, fintech and telecommunications sectors. The proposed framework integrates User and Entity Behavior Analytics (UEBA), real-time monitoring and automated response orchestration using open-source technologies including Wazuh, Elastic Stack, Apache Kafka and custom machine learning models (Isolation Forest, Autoencoder, and Long Short-Term Memory). The framework was evaluated using a multimethod strategy including expert Delphi review, controlled laboratory simulations and a quasiexperimental field study across twelve corporate environments. Expert review achieved strong consensus on design feasibility and forensic readiness (Cronbach’s alpha = 0.89). In controlled simulations, the proposed framework achieved a 91.5% detection rate compared with 67.3% for the rule-based SIEM baseline, reducing Mean Time to Detect (MTTD) from 47.2 to 8.3 minutes and lowering the false positive rate from 34.2% to 12.8%. The field observations during a three-month post-intervention period (April–June 2026) showed shorter containment times with Mean Time to Contain (MTTC) recorded at 18.7 hours compared to a literature-derived baseline of 96.3 hours. These findings demonstrate that integrating behavioral analytics and automated containment can improve threat detection, incident response and forensic readiness in resource-constrained corporate networks without the high costs of proprietary solutions.
Classification of criminal activities in surveillance footage using human detection and temporal sequence modeling.
(Busitema University, 2026) Namuyomba, Angella Zerida
Manual reviewing of surveillance footage to retrieve the evidence is a tiresome and time consuming activity due to the scarcity of frames containing these evidences. This study explores a framework for modeling spatial-temporal features to classify criminal activity in surveillance footage. This study proposes an approach that investigates the use of spatial features derived from YOLOv8 person detections as sequential representations of human behavior for LSTM-based crime classification in surveillance footage. We utilize deep learning object detection algorithms and Recurrent Neural Networks (RNNs) to learn the spatial-temporal features extracted from the UCF crime dataset in order to classify criminal activity in surveillance footage. Object detection allows us to focus our attention on frames that contain humans and human activity to avoid searching through empty frames. We believe that since most, if not all crimes involve human activity, it’s imperative to accurately detect people in the challenging surveillance footage. A comparative study was carried out to assess the performance of the common deep learning object detection tools i.e single shot mult-ibox detector (SSD) and you only look once version 8 (YOLOv8) in detection of humans in challenging surveillance footage. The dataset used for this study was compiled from the HR crime dataset which consists of 13 categories of crimes that are Abuse, Arrest, Arson, Assault, Road Accident, Burglary, Explosion, Fighting, Robbery,Shooting, Stealing, Shoplifting and Vandalism. The study indicated that the YOLOv8 object detection tool outperforms the SSD detector with a mean average precision (mAP) of 0.97 and intersection over union(IOU) of 0.89. The obtained bounding box information together with the confidence scores is converted into sequences which are then used to train and evaluate an LSTM model. The model attained an accuracy of 0.8091, F1 score of 0.8099 and MCC of 0.7919.. This implies that the model ably classfies the criminal activity in the surveillance footage and can facilitate users in faster acquisition of criminal evidence with minimal human supervision
A mobile-based digital notice board system for real-time dissemination of university (BU digital notice board).
(Busitema University, 2026) Moko, Julius; Cheptegei, Laban
This project report presents the design, development, and evaluation of the BU Digital Notice Board, a mobile-based system developed to improve the dissemination of official university notices at Busitema University. The project was motivated by the limitations of physical notice boards and informal communication channels such as WhatsApp groups, which often result in delayed updates, message loss, misinformation, and limited access to official notices. The system was developed using Flutter for the mobile application interface and Firebase services, including Firebase Authentication, Cloud Firestore, Firebase Storage, and Firebase Cloud Messaging, for backend support. The application provides user authentication, role-based access control, categorized notice posting, multimedia notice support, offline access to cached notices, push notifications, and dedicated portals for guild, cultural, and religious communities. The system was evaluated through unit testing, integration testing, and user acceptance testing with selected users from the Faculty of Engineering and Technology. Results showed that the application improved the timeliness, accessibility, and reliability of notice dissemination compared to physical notice boards and informal social media groups. The project demonstrates the potential of mobile-based systems in enhancing institutional communication within university environments. Keywords: Mobile application, Digital notice board, Real-time communication, Firebase, Flutter, Offline-first, Push notifications, Busitema University.
An Iot-driven patient vital monitoring and doctor alerting system for early health intervention.
(Busitema University, 2026) Turinawe, Richard; Nabirye, Barbra Bayuule
Access to timely health monitoring is limited in rural areas like Busia District, where patients with chronic conditions often face delays in diagnosis and treatment due to limited medical personnel, lack of diagnostic equipment, and unreliable internet connectivity. This project aims to develop a low-cost, IoT-driven wearable device capable of continuously monitoring essential vital signs, including heart rate, oxygen saturation (SpO₂), blood pressure and the body temperature. The system incorporates a real-time alert mechanism that notifies doctors or caregivers when abnormal readings occur, ensuring prompt intervention. Requirements for the system were identified through stakeholder consultations and literature review. A prototype wearable device was designed and integrated with Wi-Fi connectivity to support both online and offline operation, addressing rural infrastructure challenges. Field validation was conducted in local healthcare settings to evaluate accuracy, usability, and reliability. Results indicate that the device can effectively monitor patient vitals, provide timely alerts, and operate sustainably in low-resource environments. This system offers a practical solution to improve patient care, reduce preventable complications, and enhance health outcomes in rural communities.
Automated hydroponic fodder growth chamber with live monitoring and AI mold detection.
(Busitema University, 2026-07) Kabwangu, Muhamed; Talemwa, Dickson
Livestock production heavily depends on the consistent availability of high quality fodder and hydroponic fodder systems have become a sustainable solution for efficient animal feed production. However, these systems are often affected by challenges such as mold contamination and the lack of continuous monitoring, which can reduce fodder quality and result in losses. Traditional manual inspection methods are time consuming, inconsistent and unable to provide real-time monitoring of system conditions. [1] This project presents the design and implementation of an Automated Hydroponic Fodder Grow Chamber with AI-Based Mold Detection and Real-Time Web Monitoring System to address these challenges. The system automates the monitoring of key environmental parameters such as temperature, humidity and pH level using sensors connected to a microcontroller. A camera-based Artificial Intelligence model was integrated to detect mold growth on hydroponically produced rice fodder. The AI model processes images frames using computer vision techniques and classifies the fodder as either healthy or mold-infected. The results are displayed in real time on a web-based dashboard together with sensor data for easy monitoring. The integration of automation, artificial intelligence and web technologies enables continuous monitoring, reduces manual inspection efforts and improves the accuracy and speed of detecting mold contamination. This leads to improved fodder quality, reduced losses and more efficient livestock feed production.
ICT policy for the disabled in offing.
(New Vision, 2017-11-01) Kasanga, Kyetume
Information and Communication Technologies (ICTs) have become the leading medium for communicating, transacting, informing, education and entertaining all over the world. Usage of technologies such as television, radio, fixed and mobile telephony,
ICT graduands warned against misuse of the media.
(New Vision, 2017-09-11) Odeke, Faustine
Over 200 participants who attended a two-week information communication technology (ICT) training have been cautioned against the misuse of media platforms. The training, which was organised by the ICT ministry and held at Asinge Senior Secondary School in Tororo district, attracted a large number of participants and some were turned away due to limited computers and space.
Technology linking Karamoja to services
(New Vision, 2018-06-01) Mulondo, Lawrence
Pastoralists are trained in the use of social media platforms, such as WhatsApp and Facebook. They are embracing technology to connect with veterinary doctors, district leaders and security officers in case of emergencies.
Biotechnology law a welcome devt.
(Daily Monitor, 2017-10-30) Nyanzi, Timothy Joseph
After several years of activism, consultative meetings and nationwide sensitisation of both policy makers and the public, the National Biotechnology and Biosafety Bill 2012 has finally been passed into law- the National Biosafety Act 2017. This is a great milestone for Uganda as a nation,
Why cyber labs can never replace traditional practicals
(New Vision, 2017-08-02) Nakajubi, Gloria
Gloria Nakajubi explores the effectiveness of cyber labs as alternative to hazardous chemicals used in science experiments.
Smart jackets detects pneumonia faster
(Sunday Vision, 2017-07-30) Affedda, Emmanuel
A group led by Brian Turyabagye has created a jacket-like device named 'mama ope'. The jacket helps detect pneumonia early in order to facilitate quick life-saving treatment.
Government to buy local ICT innovations- Rugunda
(New Vision, 2017-08-02) Kayiwa, Edward
Prime Minister Dr Ruhakana Rugunda has said the Government will purchase ICT applications that are created locally in its bid to improve service delivery ahead of the vision 2020. "We should develop and adopt ICTs for smarter cities, education, environment, agriculture, healthcare and commerce." Rugunda.

Browse

Recent Submissions