Malaysian Journal of Computer Science

FINGER KNUCKLES PATTERNS AND FINGERNAILS RECOGNITION FOR PERSONAL IDENTIFICATION BASED ON MULTI-MODEL DEEP LEARNING FEATURES

Haitham Salman Chyad — 2025-01-30

Because biometric recognition systems are reliable and distinctive, they are widely used in many different applications. Hand-based person recognition has gained a lot of attention in recent years because of its stability, feature richness, dependability, and increased user acceptance. Although the dorsal side of the hand can be quite helpful in personal identification, it does not receive much attention. Finger knuckle and finger nail biometric traits can be obtained from a single dorsal hand scan. This research paper presents an approach for person identification using the dorsal finger knuckle and fingernails. It provides a structure for automatic person identification, which includes the segmentation of the detected components with hand images using the Hands Module (Media Pipe Module). The research paper focuses on the key points that hand components consist of the base knuckle, the major knuckle, the minor knuckle, the thumb knuckle, and the fingernails, which are one of the important biometric features. Specifically, a multi-model deep learning neural network (DLNN) is used to extract distinct features from each model using the DenseNet201 and Inception V3 models. The dorsal finger knuckle and fingernails of ten concatenated fingers are used to recognize all features extracted by these models. Different similarity metrics are used to compute the matching procedure for every model individually. An evaluation of the proposed approach was performed using datasets consisting of 11,076K hands with left and right hands dorsal, for 190 persons, and 4,650 Poly U, often known as Hong Kong Polytechnic University, Contactless with right hand dorsal for 502 persons. The proposed structure was achieved with results indicating that the Inception V3 models are better than the DenseNet201 model on the 11,076K Hands dataset and the ’Poly U HD’ dataset. The left-hand results are better than the right results on the 11,076K Hands dataset and the fingernails produce consistently higher identification results than other hand components, with a rank-1 scores of (99.96% and 96.28%) for inceptionV3 model, (98.11% and 93.42%) for denseNet201 model in the 11,076K Hands dataset and with a rank-1 scores of (97.07%) for inceptionV3 model, (94.83%) for denseNet201 model in the ’Poly U HD’ dataset. According to the multi-model deep learning approach proposed in the work, this approach achieved a significant improvement, achieving rank-1 scores of 99.96% compared to previous studies, which plays an important role in knuckle recognition.

CYBERBULLYING DETECTION IN SOCIAL MEDIA USING PRE-TRAINED LANGUAGE MODELS

Jasmeen Kah Ying Bong — 2025-01-30

The rapid integration of Information and Communication Technologies (ICT) has revolutionized online communication, yet it has also led to the emergence of cyberbullying, a harmful digital behaviour. This study addresses the urgency of combating cyberbullying and its negative impacts by using advanced pre-trained language models (PLMs) through transfer learning in detecting cyberbullying in social media. The goal is to enhance cyberbullying detection's effectiveness to create safer online spaces. Cyberbullying detection model using transfer learning, DistilBERT, DistilELECTRA, and MiniLM PLMs were explored. The PLMs' evaluation using the AMiCA dataset, MiniLM achieves the highest performance in detecting cyberbullying, with an accuracy of 97.84% in cross-validation and 98.57% in hold-out testing, while DistilBERT and DistilELECTRA also perform well, achieving accuracies of 97.34% and 98.03%, and 97.58% and 92.97%, respectively. MiniLM consistently maintains competitive F-measures, addressing class imbalance. Overall, MiniLM stands out with high accuracy and micro F1-scores, outperforming other models. Comparative analysis reaffirms MiniLM's excellence in binary classes and overall evaluation showcasing the effectiveness of transfer learning compared to previous studies. In conclusion, this study demonstrates the capabilities of PLMs for cyberbullying detection and suggests future research directions.

PERSONALIZED EXPLAINABILITY REQUIREMENTS ANALYSIS FRAMEWORK FOR AI-ENABLED SYSTEMS

Jia Kai Quah — 2025-01-30

Artificial Intelligence (AI) has evolved into an indispensable technology that assists humans in making better decisions through predictive analysis and personalized recommendations in numerous sectors. However, complex machine learning (ML) models become less transparent and may recommend incorrect decisions, which leads to a loss of confidence and trust. Consequently, explainability is considered a key requirement of AI-enabled systems. Recent studies focus on implementing explainable AI (XAI) techniques to improve the transparency and trustworthiness of ML models. However, analyzing the explainability requirements of different stakeholders, especially non-technical stakeholders for AI-enabled systems, remains challenging. It lacks a comprehensive and personalized requirements analysis process that investigates the risk impact of outcomes produced by ML models and analyzes diverse stakeholder needs for explanations. This research proposes a framework with a requirement analysis that includes four key stages: (1) domain analysis, (2) stakeholder analysis, (3) explainability analysis, and (4) translation and prioritization, to analyze the personalized explainability needs of four types of stakeholders (i.e., development team, subject matter experts, decision makers and affected users) for AI-enabled systems. As demonstrated by the case study, it is feasible to apply the proposed framework to analyze diverse stakeholders' needs and define personalized explainability requirements for AI-enabled systems effectively.

MOTEC: The Malay Offensive Text Classification using Extra Tree and Language Standardization

Fairuz Amalina — 2025-01-30

Cyberbullying has increased globally, with offensive text contributing significantly. Detecting offensive text in Malay is challenging due to non-standard Malay text, unique social media writing styles, a lack of standardization, and limited resources. This study proposes the Malay Offensive Text Classification (MOTEC) framework to address these challenges. The MOTEC framework incorporates a Malay standardization preprocessing task, utilizing three specialized dictionaries: (a) abbreviations, (b) noisy text, and (c) Malaysian dialects. This approach enhances data quality by converting non-standard text into standardized Malay sentences before classification. For feature extraction, the framework employs Term Frequency-Inverse Document Frequency (TF-IDF). This statistical method evaluates the importance of words in a document relative to a collection of documents, coupled with an Extra Tree classifier for the classification process. Evaluating the MOTEC framework using a private dataset collected from Twitter, this study achieved a classification accuracy of 94%, significantly outperforming other studies, which reported an accuracy of 84%. The MOTEC framework substantially improves the classification of offensive Malay text by enhancing accuracy, reducing execution time, and improving data quality through effective language standardization.

DEEP NEURAL NETWORK APPROACHES FOR AUTISM DETECTION IN CHILDREN USING VOCAL BIOMARKERS: A SURVEY

Qurratul Ain — 2025-01-30

Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by diverse social and communication challenges, with symptoms varying widely in severity. Research indicates that these challenges are often evident in the speech patterns of children with autism, making vocal biomarkers a promising avenue for early detection. Advances in machine learning, particularly deep neural networks (DNNs), offer powerful tools for analyzing these biomarkers. Despite their potential, the application of DNNs in this area remains underexplored. This survey provides a comprehensive review of the current state of DNN-based approaches, with a focus on Siamese Neural Networks, for detecting ASD through vocal biomarkers. The paper systematically examines existing speech assessment methods, evaluates the effectiveness of these neural network models, and highlights the key challenges in voice-based ASD detection. It concludes by identifying critical gaps in the research and proposing future directions to enhance the development of robust, real-world applications for early autism diagnosis.