Recent Submissions
Data-Driven Injury-Risk Prediction in Competitive Swimming using Convolutional Neural Networks
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Wickramagedara S.D.N; Subasinghe S.A.S.S; Ramanayake I.U; Wijesinghe S.D.R.N; Sumathipala, P; Kumarasinghe, S
Recent advancements in computer vision and machine learning have improved real-time human motion analysis surpassing conventional biomechanical studies by focusing on sports accessibility and movement improvement. This paper presents SwimSight, a mobile application designed to identify injury-prone movements in swimming techniques and provide personalized prevention recommendations. The system consists of three primary stages, which are video processing and pose estimation, joint-angle analysis, and injury-risk prediction. The preprocessing stage enhances video quality through frame stabilization utilizing OpenCV based filters, conventional computer vision technique. Pose estimation is executed through a marker less pipeline built on modern 2D pose estimators, that extract stable 2D keypoints to convert them into joint angles. An ensemble of 1D-CNN and 2D-CNN deep learning models is applied to predict identify risky shoulder and knee (the most common injury areas) movements with improved generalization. The system further integrates movement clustering to tailor recommendations. Through this pipeline, SwimSight delivers a robust and user-friendly solution that advances performance by injury prevention in aquatic sports.
Analysis of ML Models in Detecting and Mitigating Anomalies in Healthcare Networks
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Naisar, F; Yapa, K
The increasing integration of digital technologies in healthcare has introduced significant cybersecurity challenges. The sensitivity of patient data, combined with the complexity and interconnectedness of healthcare networks, makes them prime targets for sophisticated cyberattacks. Traditional security mechanisms, including rule-based intrusion detection systems, exhibit notable limitations in detecting evolving and advanced threats. Machine learning and deep learning techniques have emerged as powerful solutions to enhance IDS capabilities, enabling real-time anomaly detection and automated threat mitigation. This study utilizes the CICIoMT2024 and IoMT Traffic data dataset to evaluate the effectiveness of ML-based IDS methodologies specifically within healthcare network environments. Initially, various models such as Random Forest (RF), Extreme Gradient Boosting (XGB), Logistic Regression (LR), and Deep Neural Networks (DNN) were assessed for their efficacy in detecting cyber threats. However, a hybrid CNN-LSTM architecture was ultimately finalized due to its superior performance in handling imbalanced datasets, modeling complex temporal feature interactions, and delivering high detection accuracy with minimal false positives. The research analyzes challenges faced by healthcare-oriented IDSs, including high false positiveness, data imbalance, computational resource constraints, and limitations inherent in available datasets. Future research directions are proposed, emphasizing the integration of federated learnings, optimized feature selection techniques, real-time IDS deployment strategies, and development of privacy-preserving cybersecurity frameworks tailored for healthcare environments.
Evaluating Performance and User Perceptions of Multi-Cloud Database-as-a-Service Solutions: A Mixed-Methods Study
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Daundasekara D.M.P.D; Malsha K.K.S; Ilangasinghe R.Y; Ranasinghe R.A.K.D; Haddela, P. S.
Multi-cloud adoption has accelerated the use of managed Database-as-a-Service (DBaaS) platforms, yet limited research provides provider-agnostic, empirical insight into how performance, scalability, and cost differ across vendors or how these objective outcomes align with practitioner perceptions. This paper presents a mixed-methods evaluation of three major cloud database platforms: Amazon RDS, Azure SQL Database, and Google Cloud SQL. Quantitative benchmarking experiments were conducted using identical workloads on the benchANT framework to measure latency, throughput, and scalability, while a parallel survey collected perception data from academic lecturers, IT employees, and undergraduates. The study's novelty lies in triangulating benchmark results with user perceptions to reveal alignment gaps. Empirical findings show that AWSRDS and Google Cloud SQL deliver strong horizontal scalability, while Azure SQL favors vertical scaling but exhibits higher latency. A moderate correlation (r = 0.62) was found between perceived performance and measured throughput, but perceived reliability showed a weaker relationship to actual latency (r = 0.34), indicating notable perception divergences. Cost analysis further highlights trade-offs involving egress fees, storage tiers, and replica placement. The results provide actionable guidance for workload placement, elasticity planning, and multi-cloud governance, and demonstrate the value of a mixed-methods approach for evaluating cloud database services.
Robust DoS Attack Detection for V2X Networks Using RNN Variants with Attention
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Meegammana, N.W; Fernando, H
V2X communication is a foundational technology for connected and autonomous transportation, enabling vehicles, infrastructure, and pedestrians to exchange timely and reliable information essential for collision avoidance, traffic optimisation, and road safety. Denial-of-Service (DoS) attacks pose critical threats to Vehicle-to-Everything (V2X) communication by disrupting real-time safety and mobility functions. This study presents a comparative evaluation of SimpleRNN, LSTM, and GRU models for DoS detection using a balanced, normalized dataset of 120,000 samples derived from the VeReMi Extension. All models employ a unified architecture integrating attention mechanisms and dense layers, and are trained across two temporal window sizes (5 and 10) to capture varying time dependencies. Hyperparameter optimization is conducted using random search to enhance performance and training efficiency. Models are benchmarked on classification metrics (accuracy, precision, recall, F1-score), inference latency, memory and CPU usage, and model footprint (Keras/TFLite formats). Results show that GRU achieves the highest detection performance (F1-score up to 0.972, accuracy up to 97.2%), RNN offers the smallest model size (0.09 MB) and fastest inference (¡ 1 s), while LSTM provides high precision (up to 0.970) but requires tuning to improve recall. These insights support the development of secure, lightweight, and real-time intrusion detection systems for next-generation V2X environments.
A Sysmon Event-based Machine Learning Approach to Ransomware Detection
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Peiris, A; Kumara, S; Peiris, T; Ramanayake, H; Abeywardena, K,Y; Wijesooriya, A
The rapidly increasing cyber threat landscape has caused a huge impact on a large number of sectors worldwide, and among them ransomware is one of the most critical threats that should be controlled effectively. This paper presents a solution for detecting ransomware by leveraging features extracted from Windows Sysmon event logs, combined with machine learning based classification. A minimal set of key Sysmon events are collected and a comprehensive set of features are generated from it by a lightweight endpoint agent. These feature vectors are then transmitted to a cloud-based inference model for real time analysis, which ensures low overhead on the endpoints while allowing for a high scalability. Four machine learning models were evaluated on a publicly available dataset and XGBoost was selected as the suitable model for this solution due to its high accuracy while maintaining the fastest inference speed. Overall, this solution is capable of detecting ransomware with high precision and a low false negative rate, while maintaining fast inference times and a minimal processing overhead on the endpoints. The architecture enables scalable, real-time ransomware detection suitable for both enterprise and individual environments.
The SLIIT Research Document Archive (RDA) is the institutional repository of SLIIT, managed by the SLIIT Library. The primary purpose of SLIIT RDA is to manage, store, and disseminate SLIIT research output with its community and beyond, reaching the wider public. This plays a pivotal role in preserving the academic legacy of the institute.
The collection comprises the research output of SLIIT staff and postgraduate research students, including research publications, conference and symposium papers, books, book chapters, theses, and other scholarly materials. Access to full texts may be restricted depending on the access and licensing terms.