Pusat Pengajian Sains Komputer - Tesis

Browse

Now showing 1 - 20 of 55

A Denoising Generative Adversarial Network Based On Enhanced Feature Mapping Of Data Augmentation For Image Synthesis
(2024-05)
Chen, Li
Generative adversarial networks (GANs) have become a significant research topic in deep learning for image synthesis. GANs can produce diverse and high-quality results through the collaboration between the generator and discriminator. However, building a robust and stable GANs model remains a significant challenge. Previous research has attempted to enhance the original GANs by utilizing various algorithms to measure divergence between data distributions, implementing different network structures, or combining them with other structures to achieve better results. But these improvements were often limited to a single perspective. This research introduces the Denoising Feature Mapping GAN (DNFM-GAN), a GAN variant that enhances the stability of the model's training by improving both the generator and discriminator components. Specifically, the generator's ability is enhanced by adding data with noise as an extra input. This requires the generator to learn how to generate images from partially damaged data, leading to better representations learned from the data. To ensure the generator's stability and robustness, it is important to minimize the volatility caused by generator loss. Additionally, generating two types of data, 𝐺𝐺(𝑧𝑧) and 𝐺𝐺(𝑧𝑧+𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛) can increase the difficulty of discrimination for the discriminator when combined with real data. Moreover, traditional GANs often encounter the issue that Jensen-Shannon Divergence inaccurately measures changes in distribution distances, making it difficult to optimize the generator.
A Framework For Automatic Code Switching Speech Recognition With Multilingual Acoustic And Pronunciation Models Adaptation
(2014-05)
Ahmed, Basem H. A.
Recognition of code-switching speech is a challenging problem because of three issues. Code-switching is not a simple mixing of two languages, but each has its own phonological, lexical, and grammatical variations. Second, code-switching resources, such as speech and text corpora, are limited and difficult to collect. Therefore, creating codeswitching speech recognition models may require a different strategy from that typically used for monolingual automatic speech recognition (ASR). Third, a segment of language switching in an utterance can be as short as a word or as long as an utterance itself. This variation may make language identification difficult. In this thesis, we propose a novel approach to achieve automatic recognition of code-switching speech. The proposed method consists of two phases, namely, ASR and rescoring. The framework uses parallel automatic speech recognizers for speech recognition. We also put forward the usage of an acoustic model adaptation approach known as hybrid approach of interpolation and merging to cross-adapt acoustic models of different languages to recognize code-switching speech better. In pronunciation modeling, we propose an approach to model the pronunciation of non-native accented speech for an ASR system. Our approach is tested on two code-switching corpora: Malay-English and Mandarin-English.
A Machine Learning Classification Approach To Detect Tls-Based Malware Using Entropy-Based Flow Set Features
(2022-11)
Kinan
As internet encryption has grown to safeguard users’ privacy, malware has evolved to leverage encryption protocols such as Transport Layer Security (TLS) to conceal its hazardous connections. The difficulty and impracticality of decrypting TLS network traffic before it reaches the Intrusion Detection System (IDS) has driven numerous research studies to focus on anomaly-based malware detection without decryption employing various features and Machine Learning (ML) algorithms. Nonetheless, several of these studies used flow features with low feature importance value and poor capability to distinguish malicious flows, such as the number of packets sent and received in a flow or its duration. Furthermore, the outliers and frequency-based flow feature transformations (FTT) applied to mitigate the poor flow feature have several flaws. This thesis proposes a TLS-based malware detection (TLSMalDetect) approach based on ML classification to address flow feature utilization limitations in related work. TLSMalDetect includes periodicity-independent entropy-based flow set (EFS) features produced by an FFT technique. The efficiency of EFS features is assessed in two ways: (1) by comparing them to the relevant related work’s features of outliers and flow using four feature importance methods, and (2) by analyzing the classification performance in the scenarios with and without EFS features. This study also investigates TLSMalDetect detection performance using seven ML classification algorithms and identifies the one with the highest accuracy.
A Multi-objective Evolutionary Algorithm Based On Decomposition For Continuous Optimization Using A Step-function Technique
(2022-06)
Chuah, How Siang
Multi-objective optimization is an area of study which solves complex real-world problem that involves two or three objectives. Multi-objective Evolutionary Algorithm based on Decomposition (MOEA/D) is one of the algorithms that utilize the concepts of decomposition and neighbourhood to solve multi-objective problems. One of the recent MOEA/D algorithms, i.e., Constant-distance based Neighbours for MOEA/D with Dynamic Weight Vector Adjustment (MOEA/D-AWACD), integrates the concept of a constant-distance neighbourhood and a dynamic weight vector design. This combination creates a flexible neighbourhood that can adapt to the weight vectors changes. However, MOEA/D-AWACD’s performance is dependent on a constant-distance parameter,
A Simulated Annealing-Based Hyper-Heuristic For The Flexible Job Shop Scheduling Problem
(2023-03)
Kelvin, Lim Ching Wei
Flexible job shop scheduling problem (FJSP) is a common optimisation problem in the industry. The use of parallel machines allows an operation to be executed on a machine assigned from a set of alternative machines, raising a combination of machine assignment and job sequencing sub-problems. A straightforward technique to solve the FJSP is to apply a pair of machine assignment rule (MAR) and job sequencing rule (JSR), i.e. a MAR-JSR pair. However, the performance of each MAR-JSR pair is problem-dependent. In addition, within an algorithm execution, the MAR-JSR pair performs differently at different problem states. Given a wide range of MAR-JSR permutations, selecting a suitable MAR-JSR pair for a FJSP becomes a challenge. Positive outcomes on the application of simulated annealing-based hyper-heuristic (SA-HH) in addressing similar scheduling problem has been reported in the literature. Hence, this research proposes the SA-HH to produce a heuristic scheme (HS) made up of MAR-JSR pairs to solve the FJSP. The proposed SA-HH also incorporates a set of problem state features to facilitate the application of MAR-JSR pairs in the HS according to the state of the FJSP. This research investigates two variants of SA-HH, i.e. SA-HH based on the HS with problem state features (SA-HHPSF) and without problem state features (SA-HHNO-PSF).
A Visual Approach For Requirement Traceability
(2022-12)
Madaki, Abdulkadir Ahmad
Requirement traceability is a significant method of tracing and identifying the life of requirement in forward and backward directions during software development lifecycle. It is used to support impact analysis, requirement change, maintenance, verification, and validation of a software system. Visualization is one of the most suitable visual representations of requirement traceability data as it has so many aspects that can be explored. It offers detail and visible demographic symbols to show the traceability of requirements artefacts relationships. However, displaying many traceability links effectively and efficiently is a big challenge, because a software system with large numbers of artefacts and traceability links quickly gives rise to scalability and visual clutter issues. Therefore, a visual framework is designed and implemented as a tool to visualize and manage the traceability of requirements artefacts relationships. The framework follows an advance graphical user interface guide for Visual Information-Seeking which focus on overview first, zoom and filter, then details-on-demand. The tool used visualization techniques as colour-coded symbols on a node-link diagram to present data. Users can traverse the graph for an impact analysis method to understand data and make decisions during the software development life cycle. The evaluation results show a positive respond in terms of usefulness and ease of use factors. The average score mean for usefulness are 4.33 (86%), whereas the average score mean for ease of use are 4.25 (85%). This results show that the framework is useful in tracing links between requirements artefacts, easy to use as is highly effective to improve user interaction.
Acltshe-Amts: A New Adaptive Brain Tumour Enhancement And Segmentation Approaches
(2014-05)
Alkhafaji, Ali Fawzi Mohammed Ali
Brain tumor subregion segmentation from multimodal Magnetic Resonance (MR) images is of great interest for better tumor diagnosis. Multilevel thresholding is one of the prominent approaches used for brain image segmentation. Currently, when applying multilevel thresholding for brain tumor segmentation, two important problems must be carefully addressed. First, the MR brain images suffer from sensitivity to intensity inhomogeneity, poor contrast, and hidden details, which corrupt the original MR image during capturing. Second, the conventional multilevel thresholding approaches, including optimization-based thresholding approaches, have several main issues, such as manual adjustment of multilevel thresholds, dedicating single thresholding criteria as objective functions, leading to the bias of the thresholding towards a specific type of MR image, and the requirement to tune a large number of control parameters. In this thesis, a two-stage approach is proposed to address these issues. In the first stage, a new image enhancement approach called Adaptive Clip Limit Tile Size Histogram Equalization (ACLTSHE) is proposed to improve contrast, highlight the hidden details, and achieve homogenized intensity distribution of MR images. The ACLTSHE integrates Contrast-Limited Adaptive Histogram Equalization, Multi-Objective Whale Optimization Algorithm, Discrete Entropy (DE), Peak Signal-to-Noise Ratio (PSNR), and Structure Similarity Index (SSI) to improve the quality of MR images while preserving the original structure of the MR images. In the second stage, a new approach called Adaptive Multilevel Thresholding Segmentation (AMTS) is proposed for unsupervised brain tumor xxi subregion segmentation from normal brain tissue. The AMTS approach segments and extracts the whole tumor, core tumor, and enhanced tumor regions from the brain MR images, integrating the Multi-Objective Grasshopper Optimization algorithm, Kapur Entropy, Cross-Entropy, and Localized active contour.
An Efficient Distributed Slotted Multi-Hop Wireless Mac Protocol For Internet Of Things
(2021-11)
Mohammad Ali Sarvghadi
In Wireless Sensor Networks (WSNs), the Media Access Control (MAC) layer is responsible for coordinating Sensor Nodes (SNs) as their access to the media. Real time data flow needs precise scheduling for sensor nodes to access the medium in order to send the data on time. Time slotted transmission scheduling is needed to access the media without collision. To achieve this, sensor nodes need to be synchronized and know about each other’s notion of time. However, time synchronization requires message exchanges, which incur high overhead. The proposed MAC protocol, which consists of algorithms that minimizes collision and detects hidden terminals, eliminates the message overheads of global time synchronization
Analysable Chaos-based Design Paradigms For Cryptographic Applications
(2024-07)
Abba, Abubakar
Chaos-based cryptography has garnered significant attention, with many designs focusing on obscuring security through complex designs that make them difficult to analyze, improper design structures (ad-hoc designs) with inaccurate keyspace justification. These compromise the standards of well designed, simple, and secure design principles in cryptographical design protocol and do not facilitate future cryptanalytic efforts. Moreover, to date, there have not been any chaos-based cryptosystems being implemented to secure real-world communications. This study proposes simple and analyzable paradigms based on well-established cryptographic principles (SPN and Feistel) to address these issues. First, an in-depth review is conducted on the current state-of-the-art in the field of chaos-based cryptographic algorithms to identify the challenges of various design and evaluation methods that have been developed over the years. A comprehensive analysis into a largely overlooked problem in chaos-based cryptosystems is conducted, highlighting their unusually large keyspaces. Multiple examples demonstrate instances of chaos-based ciphers overestimating keyspaces and reveal practical and theoretical challenges in key generation approaches. The study highlights recommendations and alternative solutions for utilizing secret keys in chaos-based cryptography. Then, a simple chaos-based key schedule that ensures the involvement of every bit of the secret key in the generation of round keys is proposed. The proposed key schedule successfully passed both the NIST and ENT statistical test suites, indicating that highly complex designs are unnecessary to achieve desirable security properties.
Analysis Of Feature Reduction Algorithms To Estimate Human Stress Conditions
(2022-09)
Arasu, Darshan Babu
Stress is a normal reaction of the human organism which triggered in situations that require a certain level of activation. This reaction has both positive and negative effects on everyone’s life. Thermal-based imaging has shown promising results in detecting stress in a non-contact and non-invasive manner. Therefore, this study aimed to present analyse of the performance of feature classify when combining with feature selection algorithm to estimate human stress based on the facial feature of thermal imaging. Three hybrid classifiers, Support Vector Machine (SVM), Decision Tree (DT) and Logistic Regression (LR) combined with feature reduction analysis, Principal Component Analyse (PCA) and Analysis of Variance (ANOVA) was evaluated with 10-fold validation to compute classification accuracy. Four statistical features was extracted; mean, maximum, minimum and standard deviation of the gray scale value from six area regions of interest. Results showing that hybrid classifier DT-ANOVA achieves higher accuracy of 62% compared to others 90 combination classifiers. The findings demonstrated that DT-ANOVA performs well with a small dataset, while SVM and LR can improve the accuracy when fused with ANOVA for a big dataset. The findings also suggested that ANOVA can provides comparable performance as PCA.
Bounded Box-Zoning Integrated Approach For Children Handwriting Recognition
(2024-03)
Qausbee, Nik Nur Adlin Nik
This study focuses on recognizing handwritten character datasets of children through the integration of image processing and machine learning. The bounding box, one of the best structural feature extraction methods, has demonstrated high performance in the optical character recognition (OCR) pipeline. However, its implementation in children’s handwriting has shown a decreasing trend in alphabet detection. Similarly, zoning, a powerful technique under statistical feature extraction demonstrated good classification but is limited by having unlimited feature values and is not applicable for characters with high variations, such as children’s handwriting. The objectives of this study are to identify significant English alphabets based on their features, propose a bounded box-zoning integrated approach to improve the OCR pipeline and identify the accuracy of the proposed method. The Minnesota Handwriting Assessment (MHA) was utilized for data collection, involving handwriting samples collected from 90 children aged between 6 to 9 years old. The study then proceeded with image processing steps, including alphabet grouping into ‘small’ and ‘tail’ groups, feature extraction using the proposed hybrid method (bounded box-zoning), and classification using the multi-input neural network method.
Classification Of Cpg Island And Promoter Regions Using Rare K-Mer Motifs
(2015-07)
Mohamed Hashim,Ezzeddin Kamil
Empirical analysis on DNA k-mers is proven to be an effective means to discover functional elements in the human genomes. Among the empirical works, “rare k-mer" (RKM) is a very interesting subject to be studied due to their unique sequence properties. RKMs were referred as DNA k-mers (of k=7 to 11) that have a low frequency in k-mer frequency distributions of mammalian genomes, yet there are large variations of their mass at the lower k-mer spectra. Our first objective is to discover RKM motifs in the human genome; to identify potential RKM computational applications in biology; and to infer representations for the identified applications. In short, the first goal was achieved by using comparative strategy and several bioinformatic tools (of UCSC browsers, EpiGRAPH, and Galaxy) which correlated RKMs with several genomic features of CpG Islands (CGIs), promoter, 5’ Un- Translated Regions (5’UTRs), and open chromatin regions; and by using intrinsic string mining approaches which identified several unique RKM topological, compositional, and clustering properties in the correlated genomic features.
Clustering Ensemble And Hybrid Of Deep Learning For Spatio-Temporal Crime Predictions
(2024-06)
The increase in the urban population poses challenges in managing services and safety from criminal activities. The concerned stakeholders intend to predict the time, location, number, and types of crimes to take suitable preventive measures. Accurate identification and prediction of crime hotspots can significantly benefit the concerned stakeholders in preventing crime by creating accurate threat visualizations and allocating police resources efficiently. Several techniques have been proposed for crime prediction, but they are limited in accuracy and predicting crime according to crime type on an hourly, monthly, and seasonal basis. Crime hotspot detection approaches are primarily sensitive to initial parameter selection and finding clusters of varying shapes and densities. Similarly, existing Crime prediction approaches are limited in capturing non-stationary data and long-term dependencies by focusing on crime types. Thus, the crime detection and prediction mechanisms need improvement in the number of crimes, crime span, accuracy, and dense crime region and prediction. The core objective of this study is twofold. First, it proposes a crime hotspot detection model to improve accuracy using Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) and its clustering ensemble to capture varying shapes and densities clusters and improve accuracy. HDBSCAN is used with varying parameter initialization in the generation mechanism under the cluster ensemble paradigm. Moreover, six different distance measures are used to ensure diversity. In addition, an evaluation function is proposed parameterized by silhouette score to select the stable clustering among a pool of clustering solutions to ensure quality.
Communication Efficient Decentralized Collaborative Learning Of Heterogeneous Deep Learning Models Using Distillation And Incentive-Based Techniques
(2023-07)
Zahid, Iqbal
Smart devices, collectively, have very valuable and real-time data which can be used to train very efficient deep learning models for AI applications. However, due to the sensitive nature of this data, people are more concerned about the privacy of their data and not willing to share it. Therefore, there is a need to learn from this valuable data in a decentralized fashion by withholding data localized on these intended devices and efficiently performing necessary computation on these devices by exploiting their computational resources. Statistical heterogeneity and fully model heterogeneity are among the key challenges in applying the Decentralized Learning (DL) approaches in real scenarios. Typically, all existing DL techniques assume that all devices would have homogeneous model architecture. However, in real applications of DL, due to different computational resources and distinct business needs of devices, it is intuitive that they may have completely different model architectures. Very limited work has been performed to address fully model heterogeneity problem. In the same way, some work has been performed to address the statistical heterogeneity however mostly is hard to apply in real scenarios or is only for limited use cases.
Compute Language Interface: A Transparent Wrapper Library For Multi Cpu-Gpu
(2013-03)
Ooi, Keng Siang
The Graphics Processing Unit (GPU) processing capability is getting more powerful than before. Compute intensive and data parallelism applications are proven to perform better on the GPU than on the Central Processing Unit (CPU). However, available General-Purpose Computing on Graphics Processing Unit (GPGPU) programming frameworks which are available publicly are unable to reach beyond the single computer limitation to utilize multiple CPUs and GPUs at different computers in a distributed computing system easily. This study presents the Compute Language Interface (CLI) which is a wrapper library that enables the existing OpenCL applications access to all available CPUs and GPUs in a distributed computing system through Message Passing Interface (MPI) transparently. It is designed to improve the scalability of the OpenCL applications on a distributed computing system while maintaining the same set of application programming interface (API) in the original OpenCL library. The applications can access all available CPUs and GPUs in different computers in a distributed computing system as ifall the CPUs and GPUs are in the same computer.
Concept And Relation Extraction Framework For Ontology Learning
(2023-08)
Al-Aswadi, Fatima Nadeem Salem
Extracting valuable knowledge and representing it in a machine-understandable form is considered one of the main challenges of semantic web and knowledge engineering fields. The explosive growth of textual data is coupled with the increasing demand for ontologies. Ontology Learning (OL) from text is a process that aims to automatically or semi-automatically extract and represent the knowledge from text into the machine-understandable form. Ontology is a core scheme representing knowledge as a set of concepts and their relationships within a domain. Extracting the concepts and their relations is the backbone for an OL system. The existing OL systems have many limitations and drawbacks, such as not efficient for extracting relevant concepts especially, for large-length dataset; depending on a large amount of predefined patterns to extract relations, and extracting very few types of relations. In this thesis, a framework called Concept and Relation Extraction Framework (CREF) is proposed. It consists of four main stages: enhancing pre-processing method, developing methodology for the concept extraction task, proposing a new text representation approach for relations, and improving relation extraction method. The first stage involves proposing a new Concept Extraction stopwords (CE-stopwords) for scientific publications while the second stage involves introducing a new Domain Time Relevance (DTR) metric and proposing a Developed Concept Extraction Method based on DTR called (DTR-DCEM).
Covid-19 Misinformation Classification On Twitter In Malaysia Using A Hybrid Adaptive Neuro-Fuzzy Inferences System (Anfis) And Deep Neural Network (Dnn)
(2023-01)
Ravichandran, Bhavani Devi
The spread of Covid-19 misinformation on social media had significant real-world consequences, raising fears among internet users since the pandemic has begun. Worldwide, researchers have shown an interest in developing deception classification methods to reduce the issue. This study aims to create an accurate model for the classification of Covid-19 misinformation in social media. This research has also conducted a systematic literature review to identify the most efficient method for classification with 35 papers. According to existing studies, the most efficient method for classification with the highest accuracy is the ANFIS and the DNN models. Thus, it was identified that the hybrid model of ANFIS-DNN shows the highest accuracy results. Therefore, the main goal of this study is to classify Covid-19 misinformation using an optimised hybrid model of ANFIS-DNN on social media based on the level of risk. A total of 8,000 Malaysian-based Tweets were extracted from Twitter based on topics related to Covid-19. The dataset is explored, cleaned, pre-processed, and the tweets were grouped into BoW model. Then, the proposed ANFIS-DNN is used to run the pre-processed dataset and the accuracy performance result shows 99%. Evaluation performance indexes such as confusion matrix, and accuracy are implemented in this research. The proposed model is then compared with ANFIS, DNN, Logistic Regression, SVM, Random Forest, and XGBoost. Furthermore, the accuracy is compared with other related works.
Data Integrity For Cloud Computing With Homomorphic Encryption
(2022-08)
Awadallah, Ruba N S
Cloud computing is a new computing model in which resources are provided as general utilities that users can lease through the Internet on-demand fashion. How- ever, this technology has numerous data security concerns. Asymmetric Homomorphic cryptosystem has been acknowledged as one of the potential solutions for achieving secure cloud computing since it can provide data privacy and confidentiality. Homo- morphic Encryption (HE) is a cryptosystem that allows cloud computing to operate computations on encrypted data. However, HE schemes are non-compliance to indis- tinguishability under adaptive chosen-ciphertext attack (IND-CCA2) because of their malleable nature. Moreover, the client (data owner) cannot prove if the data has been manipulated by the Cloud Service Provider (CSP) since CSP has absolute authority over the client’s data. The client is also unable to trace the processes that are being applied to the data once the data is outsourced. The client can also not verify the CSP’s reliability in applying the required operations on the required data. Therefore, implementing HE alone is not sufficient against various data integrity attacks. Two dif- ferent verifiable cloud computing schemes are being proposed in this work to address these questions. The first verifiable cloud computing scheme uses modular arithmetic to produce verified data to verify the CSP’s HE computations over a finite integer field. The performance of the proposed scheme varied based on the underlying cryp- tosystems used. However, based on the tested cryptosystems, the scheme has 1.5% storage overhead and a computational overhead that can be configured to work below 1%.
Enhanced Automated Framework For Cattle Tracking And Classification
(2022-09)
Williams, Bello Rotimi
Employing computer vision-based methods in monitoring individual cows has become what researchers are striving for. Computer vision-based methods could be used to monitor each individual cows. The accuracy of the existing methods and frameworks is below expectation in handling these tasks. Moreover, they can still be improved to achieve better and more accurate results. The goal of this research is to provide a framework for better cattle tracking and classification systems. An enhanced object tracking algorithm (PFtmM) that integrates enhanced particle filter algorithm (PFtm) with mean-shift tracker (M) is proposed and deployed as first step to address the problems arise due to occurrence of occlusion and non-linear movement of cow objects in video. The integration of particle filter with mean-shift tracker considers the following techniques: (1) temporary memory for keeping tracks of occluded cow objects; (2) supplementing each algorithm’s weakness by the strength of the other for tracking non-linear movement. Strength of particle filter (PF) is its non-linearity property which it uses to track object’s non-linear movement but, with high computational time and search range as its weakness. Temporary memory (tm) strength is its ability to track full occlusion with reduced computational time and search range. Mean-shift strength is its sensitivity to object’s movement and colour distribution by using similarity function but, with inability to track object’s non-linear movement and full occlusion as its weakness.
Enhanced Bluetooth Low Energy 5 Aodv-Based Mesh Communication Protocol With Multipath Support
(2023-04)
Muhammad Rizwan Ghori
Wireless Ad-Hoc Networks (WAHN) are growing more widespread due to improvements in Internet of Things (IoT) technologies. Several WAHN technologies, including ZigBee, Z-Wave, Threads, Bluetooth Low Energy (BLE) are available. Despite progress on IoT, network-based routing continues to be a critical problem that has to be tackled. In view of the above, this study focuses on a BLE-based mesh network.

Browse

Browsing Pusat Pengajian Sains Komputer - Tesis by Title

Results Per Page

Sort Options