Pusat Pengajian Sains Komputer - Tesis

Browse

Now showing 1 - 5 of 499

Enhanced Least Significant Bit-Based Algorithms For Spatial Domain Image Steganography
(2025-06)
Al Enzi, Abdalla Ramah
In digital image security, steganography conceals data within images using least significant bit (lsb) methods, which often compromise image quality. This research introduces three novel algorithms: pixel indices least significant bit (pilsb), semi-adaptive least significant bit (sdlsb), and reversible rewritable least significant bit (rrlsb), which aim to improve stego-image quality by minimizing the number of modified bits during the embedding process. The goal is to address the limitations of traditional lsb steganography, particularly issues of compromised security, reduced imperceptibility, and limited payload capacity. The methodology involves a comparative analysis of the three algorithms, measuring peak signal-to-noise ratio (psnr), mean squared error (mse), and structural similarity index (ssim) across different image and secret message sizes. Pilsb uses red and green lsbs for data embedding and blue lsbs for indexing, improving embedding efficiency. Sdlsb dynamically decomposes bytes based on color intensity to increase payload, while rrlsb employs reverse lsb in a two-stage process for indirect data embedding.
Enhancing 2D Joints Estimation In Markerless Motion Capture For Improved Tracking Of Spinal Movements
(2025-09)
Pauzi, Ainun Syarafana
This research aims to improve the anatomical accuracy of 2D human pose estimation models by enhancing the level of detail in the skeletal representation, particularly for the spine region. The research is guided by two main objectives: (1) to identify which of three widely used deep learning models (OpenPose, MediaPipe BlazePose, or MoveNet) most accurately predicts keypoints by comparing model outputs with Inertial Measurement Unit (IMU) data; and (2) to develop a curve-fitting algorithm using Bezier and B-Spline formulas to create realistic spine curvature based on new spine keypoints.
An Improved Static Analysis Approach For Detecting Input Validation Vulnerabilities In Web Application
(2025-09)
Marashdih, Abdalla Wasef Mohammad
This thesis proposes a novel approach for detecting XSS and SQLi vulnerabilities. First, a static analysis technique is introduced to identify feasible execution paths in the PHP source code, an area currently lacking dedicated tools or methods. Identifying feasible paths significantly reduces false positives in static analysis outcomes. Second, taint analysis is employed to trace the sources of vulnerabilities, confirm their execution, and assess the application of appropriate sanitisation along those feasible paths.
An Interplay Of Bilingualism On Language Skills And Cognitive Functions Among Saudi International School Students
(2025-05)
Ali, Abd Ali Shams Mhmood Abd
Blockchain technology introduces a new decentralized paradigm era avoiding the reliance on trusted third parties. It is a transparent and distributed ledger which is designed fundamentally for digital cryptocurrencies but has since been extended to various industries. However, its immutability obligates significant challenges including storing illicit contents, privacy regulations violations, and restricting data management flexibility. Therefore, redactable blockchain has emerged as a leading solution enabling controlled immutable contents amendment. Transaction-level redaction reinforced by fine-grained access control forms the cornerstone of the current redaction mechanisms. This redaction concept essentially depends on modifying mutable transactions governed by predefined access policies specified by the transaction owner. Modifiers equipped with necessary rewriting privileges and who fulfil the associated access policy are enabled to perform modifications. However, the existing redaction mechanisms infrastructures are inefficient. For instance, the chameleon hash ephemeral trapdoor (chet),
Integration Of Dynamic Loss Function Autoencoder In Boost
(2025-04)
Shamsudin, Haziqah
Highly class imbalance together with high data complexity (feature overlap and poor class separability), presents a significant challenge in machine learning. Traditional classifiers often exhibit bias towards the majority class, resulting in poor performance on the minority class, which is frequently the class of interest. Existing methods address imbalance or complexity, but rarely both effectively, and often lack adaptivity during training. This thesis addresses these challenges through a series of algorithmic enhancements.

Browse

Recent Submissions