Pusat Pengajian Sains Komputer - Tesis
Browse
Recent Submissions
- PublicationEnhanced Usability Framework With Need-Based And Push-Based Factors For Malaysia Elderly’s Mobile Health Application(2025-09)Lim, Kah HaoDigital health services are increasingly in demand in Malaysia, but the adoption of mobile health (mHealth) applications among the elderly remains low. Usability challenges are a key barrier, yet limited research has focused on designing health service apps specifically for elderly users. This thesis reviews existing usability models and proposes an enhanced usability framework that includes seven key factors: user-friendliness, efficiency, learnability, help and documentation, memorability, need-based, and push-based. A mobile app prototype was developed to demonstrate this framework, and a usability study was conducted with 151 elderly participants in Malaysia.
- PublicationPosition Update And Search Strategy Based Binary Particle Swarm Optimization For Feature Selection(2025-09)Sani, TijjaniThe rapid expansion of data due to the advances in science and technology causes an increase in both the number of instances and features (dimensions). When the dimensionality of data increases, the computational cost also increases, mainly exponentially. Consequently, designing a feature selection method that can select relevant features is very important. The binary particle swarm optimization (bpso) algorithm often converges to a local optimum quickly (local optima stagnation), missing better opportunities that lead to premature convergence. In addition, the original position update formula cannot effectively balance exploitation (local search) and exploration (global search) in the search process. Hence, this research work proposes a modified binary particle swarm optimization method for feature selection. Its main feature first, is the application of a computational reduction which lowers the number of redundant features and computation costs. Secondly, determines the probability of the particles switching positions using position values rather than velocity values which speed up convergence and overcome the premature convergence simultaneously. Two variant methods for determining the probability of changing the position of a particle element were introduced. These result in the two variants of enhanced binary particle swarm optimization (ebpso) for feature selection, called ebpso1 and ebpso2
- PublicationKnowledge-Enhanced Deep Neural Network For Legal Judgment Prediction And Explanation(2025-09)He, CongqingThis study aims to bridge the gap by developing the JuriSim framework to enhance the performance and explainability of LJP. Firstly, we propose a rationale generation in the JuriSim framework by introducing event chains as auxiliary knowledge. This enhances the model’s ability to focus on important legal events when generating rationales, thereby improving the effectiveness of legal judgment explanations. Secondly, we propose a dual residual cross-attention mechanism that integrates knowledge of rationales and legal events with the fact description.
- PublicationCollaborative-Based Approach Utilizing Ensemble Feature Selection For Detecting Http-Get Ddos Attacks In Cloud Computing Environments(2025-05)Ashhab, Ziyad Reefat Hamzeh AlCloud computing environment (cce)-based services present a novel paradigm for remote business management. One of the primary advantages of utilizing cce is the availability of on-demand services, thereby facilitating a pay-per-use model. This makes cce technology a convenient means of facilitating services over the internet. However, security vulnerabilities, such as distributed denial of service (ddos) attacks, particularly http-get ddos attacks at the application layer, pose a significant threat to service availability in cces. This thesis proposes a collaborative approach utilizing ensemble feature selection to detect http-get ddos attacks in cces. The proposed approach comprises six phases. The first phase entails data gathering and pre-processing, responsible for collecting and processing data from multiple sources. The second phase involves dataset generation, comprising the creation of a synthetic cce-specific dataset. The third phase focuses on feature enrichment, aiming to augment the avws access log extracted from vm activity and resource logs to enhance the detection of http-get ddos attacks. The fourth phase entails dataset validation, aimed at validating the dataset to ensure its validity and readiness, and confirming that it meets the requirements of a benchmark dataset. The fifth phase involves ensemble feature selection, aimed at selecting the most crucial and minimal feature set that contributes to detecting http-get ddos attacks. The sixth phase aims to develop a deep learning detection model based on long short-term memory (lstm) to detect http-get ddos attacks on cce accurately.
- PublicationEnhanced Feature Selection With Stacking Method For Ddos Attack Detection In Software Defined Networking Environment(2025-05)Alasfour, Tareq I. A. AlasfourSoftware-defined networking (sdn) is a networking approach that separates the control plane from the data plane. However, the dynamic and programmable nature of sdns introduces new security challenges, particularly in detecting distributed denial of service (ddos) attacks. The proliferation of ddos attacks significantly threatens network accessibility and performance. Traditional feature selection methods struggle with the complexity of network traffic data, resulting in poor detection performance. To address this, we propose a genetic algorithm wrapper feature selection (gawfs) method. This approach integrates chi-squared (chi2) and genetic algorithm (ga) techniques with the kendall rank correlation method to select the most relevant features. Gawfs effectively reduces feature dimensions, eliminates redundancy, and identifies crucial correlated features for classification. To further enhance detection accuracy, we employ a stacking ensemble model. This model combines multi-layer perceptron (mlp) and support vector machine (svm) as base classifiers, with a random forest (rf) as the meta-classifier. Our proposed classifier achieves impressive accuracy rates of 99.86% for seen data and 98.89% for unseen data, representing improvements of approximately 5% and 40%, respectively, over previous studies. Additionally, the training time is reduced to 2,593 seconds, an improvement of approximately 29.92%.