https://www.um.edu.mt/library/oar/handle/123456789/141711 2026-10-17T08:01:39Z https://www.um.edu.mt/library/oar/handle/123456789/150176 Title: Real-time object detection using lightweight deep learning models on mobile devices for assisting the visually impaired Abstract: This project develops FindME, a real-time object detection application forAndroid designed to assist visually impaired users in locating three commonly used items being sunglasses, keys, and a wallet. The user interacts with the application through voice commands. The app then activates the device camera and provides audio directional guidance towards the object requested by the user based on the detected object’s position within the camera frame, without requiring any additional hardware or internet connectivity. The dataset used for training and testing was constructed using a combination of publicly available annotated images and real-world images captured in domestic environments. A fourth class that represents prescription glasses was included to reduce false positive detections caused by visual confusion between sunglasses and prescription glasses. Background images were also incorporated into the dataset to help reduce false positives. Three object detection architectures were trained and evaluated. These include: You Only Look Once 26 (YOLO26), Real-Time DetectionTransformer large (RT-DETR-l), and RoboflowDetection Transformer Nano (RF-DETR Nano). These deep learning models were compared to determine which model best balances detection accuracy and real-time inference performance on a mobile Android device with limited resources, specifically the Samsung Galaxy S23 Ultra. YOLO26n iter5 was selected as the deployed model since it provided the best balance between accuracy and real-time performance on the target device. The final 8-bit integer (INT8) TensorFlow Lite (TFLite) model achieved a test mean Average Precision at 0.50 Intersection over Union (mAP@50) of 0.9536, a deployed model size of 2.83 megabytes (MB), and approximately 12 frames per second (FPS) with an average latency of 80 ms per frame using Neural Networks API (NNAPI) acceleration. YOLO26s achieved slightly higher validation accuracy but only reached approximately 5 FPS, while RT-DETR-l and RF-DETR Nano achieved approximately 0.7 FPS and 1.0 FPS respectively. These results show that YOLO26n was the only evaluated architecture suitable for real-time on-device deployment in FindME. Description: B.Sc. (Hons)(Melit.) 2026-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/150175 Title: Attaining fault tolerance via a substructural type system Abstract: Distributed and concurrent systems are highly susceptible to partial failures. However, building fault tolerant implementations that can mask these failures is not immediate. This work bridges the gap between definitions of fault tolerance for up to n faults and the use of correct-by-construction methods such as substructural type systems. Such type systems, like behavioural type systems, provide strong correctness guarantees. Weformalise a division between fallible and infallible channels by utilising the semi-axiomatic sequent calculus (SAX) framework for asynchronous concurrency. We introduce an extended behavioural type system and operational semantics that verify and enforce the requirement for sufficient redundancy within a linear framework, using explicit process spawning and channel synchronisation constructs. To validate this approach, the proposed extensions are integrated into an existing Go-based tool built on SAX, providing a practical framework for developing concurrent systems that are correct-by-construction with respect to a specified bounded fault tolerance promise. Description: B.Sc. (Hons)(Melit.) 2026-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/150155 Title: Tree-based ensemble re-benchmarking of EMBER2024 : detecting evasive and emerging threats Abstract: Static malware detection involves analysing a file’s code or binaries without executing them to identify malicious intent. As the field evolves, machine learning has become increasingly central to this process yet evaluating models against evasive and harmful threats requires rigorous, large-scale datasets. Addressing this need, EMBER2024 is a recently released dataset offering over 3.2 million files spanning six file formats, including a challenge set of samples that initially evaded antimalware detection. While the benchmark provided an untuned Light Gradient Boosting Machine (LightGBM) baseline, the authors acknowledge that it is unoptimised and invite further work. Therefore, this dissertation presents a re-benchmarking of EMBER2024, with the aim of evaluating the performance of three tree-based ensemble algorithms, namely LightGBM, eXtreme Gradient Boosting (XGBoost) and Random Forest (RF) across its full task suite using both default and hyperparameter optimised configurations. The results indicate a consistent performance advantage for sequential gradient boosting models (LightGBM and XGBoost) over the bagging-based RF, notably when identifying evasive malware. LightGBM maintained the highest overall performance across both configurations, demonstrating a structural advantage in isolating difficult and unseen samples and positioning itself as a strong candidate architecture for large-scale static malware detection. Notably, although XGBoost consistently placed second in all other evaluation tasks, it produced a surprising result on multi-label attribute prediction tasks, outperforming both LightGBM and RF across all five tasks in distinguishing positive from negative tags. Description: B.Sc. (Hons)(Melit.) 2026-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/150153 Title: Runtime verification and tamper evident logging for regulatory compliance Abstract: As software regulation in the European Union (EU) continues to develop, deployed systems are increasingly expected to provide not only correct behaviour, but also transparency, auditability, and accountability. EU regulatory frameworks such as the AI Act and the Cyber Resilience Act introduce obligations concerning cybersecurity, incident handling, traceability, and evidence preservation, especially for high‐risk and autonomous systems. These requirements extend beyond traditional static verification and testing, creating a need for runtime mechanisms that can observe system behaviour, detect non‐compliance, and retain evidence for later forensic analysis. This work presents an assessment of the feasibility of using both runtime verification and tamper‐evident logging to comply with such regulatory requirements. Wefocus on the MDIA’s ITA guidelines, and develop a tool, FLARE, that realises an ITA harness, which checks live interactions between the system and its environment, and a forensic node, which records verifiable runtime logs. We apply FLARE to a waste‐identification and localisation drone, a cyber‐physical system operating under legal, safety, and data‐related constraints. The case study demonstrates that FLARE can help address regulatory and operational requirements while adding a feasible overhead, offering a practical approach for more compliant software instrumentation. Description: B.Sc. (Hons)(Melit.) 2026-01-01T00:00:00Z