Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control

Ambidextrous Blockchain Governance to Strengthen BankCo’s Digital Transformation through COBIT 2019 Traditional and DevOps

2025-07-05T12:26:25+00:00

The adoption of blockchain in banking accelerates digital transformation by enhancing transparency and operational efficiency. However, it also introduces governance challenges related to compliance, accountability, and system integrity in a highly regulated environment. This study addresses these challenges by developing a blockchain governance solution based on the ambidextrous of COBIT 2019 Traditional and DevOps Focus Areas. A governance model was built and evaluated through iterative steps. Data were collected via interviews with key stakeholders and triangulated with internal documents, including annual reports, risk frameworks, and policy records, until data saturation was achieved. The analysis applied the ambidextrous COBIT 2019 framework across seven governance components. Governance and Management Objectives (GMOs) were prioritized using design factors, national regulations (POJK No.11/2022 and SOE Minister No.PER-2/MBU/03/2023), and insights from prior studies. APO12: Managed Risk was identified as the most prioritized GMO. A capability gap analysis revealed missing leadership roles, overlapping security responsibilities, and underdeveloped risk management practices. Recommendations include formalizing key governance roles and strengthening risk handling for blockchain and DevOps environments. These enhancements are expected to increase the maturity level of APO12 from 3.5 to 4.1, thereby improving BankCo’s risk management, compliance, and innovation capabilities. Ultimately, the findings contribute to continuous digital innovation by aligning risk management practices with strategic performance goals and adaptive control mechanisms rooted in emerging technology principles.

Modified U-Net for Leaf Segmentation of Eucalyptus pellita Seedlings in Open Natural Environments

2025-07-05T12:07:26+00:00

Segmenting plant leaves in natural environments was challenging due to fluctuating lighting, complex backgrounds, and heterogeneous leaf morphology. This study was conducted aiming at addressing the above mentioned issues by developing a modified U-Net architecture for segmenting Eucalyptus pellita seedlings in open nursery settings. The proposed solution introduced a ResNet50 encoder pre-trained on ImageNet, enhanced regularization in the decoder, and a combined loss function comprising Binary Cross-Entropy and Dice Loss to optimize pixel-wise accuracy and shape conformity. A total of 2,181 high-resolution RGB images were collected under three distinct lighting conditions: cloudy, sunny, and scorching. All images were manually annotated, stratified, and augmented with geometric and photometric transformations. Model training employed adaptive learning rates and early stopping strategies. The results showed the highest median segmentation score of 0.867 under cloudy conditions, followed by 0.853 under scorching conditions, and 0.838 under sunny conditions. Statistical testing confirmed significant differences across lighting scenarios. Visual inspection further demonstrated the model’s ability to preserve spatial details and mitigate the impact of shadows, reflections, and cluttered backgrounds. Despite the decline in precision under sunny conditions, segmentation consistency remained high. In conclusion, the developed model successfully addressed key challenges in leaf segmentation under variable outdoor lighting. The findings support its use for robust, high-precision segmentation, offering a foundation for real-time plant health monitoring in nursery-scale applications.

Transfer Learning Approaches for Non-Organic Waste Classification: Experiments with MobileNet and VGG-16

2025-05-23T03:12:06+00:00

This research aims to develop machine learning (ML) models for classifying non-organic waste. The goal is to support more effective waste management by increasing recycling rates, reducing landfill use, and minimizing environmental impact. The ML models proposed in this paper classify 20 types of non-organic waste collected from the internet, which consists of 2,552 instances. Experiments on our dataset reveal key findings. First, MobileNet, achieving 86% accuracy, outperforms VGG-16, which reaches 72% accuracy. Second, both models effectively classify distinct objects such as cigarette butts, toothbrushes, and glass bottles, demonstrating strong pattern recognition for these categories. Third, both models struggle with misclassification in visually similar categories, particularly paper-based waste like cardboard, carton packaging, books, and foam packaging. Fourth, MobileNet shows notable confusion in classifying plastic packaging, carton packaging, and books, while VGG-16 exhibits greater misclassification in foam packaging, cardboard, and newspapers. These results highlight the challenge of distinguishing waste types with overlapping textures and shapes. Moreover, it presents the urgency of improving the model to distinguish visually similar waste materials. Looking at the number of labels used in this paper compared with existing studies, the findings demonstrate the competitiveness of our models for non-organic waste classification.

Game development of Banjar Archive for interactive cultural education ultilizing large language models

2025-04-22T01:24:51+00:00

The preservation of Banjar cultural heritage is threatened by globalization andthe fading interest of younger generations. This study addressed thesechallenges by developing an interactive educational game using the GameDevelopment Life Cycle (GDLC) framework and integrating Large LanguageModels (LLMs) for adaptive and immersive player interactions. The six stagesof GDLC namely initiation, pre-production, production, testing, beta, andrelease were systematically applied, resulting in a game that blends dynamicnarratives to engage players while educating them about Banjar culture. BlackBox Testing verified 14 test scenarios that all passed successfully, ensuringsystem stability and reliability. Additionally, user experience evaluation usingthe Game Experience Questionnaire (GEQ) highlighted high levels ofimmersion (4.936), competence (4.448), flow (3.124) and positive affect(4.976) among players, with minimal reported tension (1), challenge (1.744)and negative affect (1.07). These results demonstrated that the gamesuccessfully balances educational goals with engaging gameplay, fosteringmeaningful connections to Banjar heritage. By leveraging LLM technology thegame enhances interactivity offering an innovative approach to Banjar culturalpreservation in the digital era. This research extends the existing body ofknowledge on AI-driven gamification strategies in heritage conservation with aspecific focus on Banjar culture.

Spoon Stabilization for Essential Tremor Patients Using PID Control Optimized by PSO

2025-03-26T10:16:25+00:00

Essential tremor is a neurological disorder that causes involuntary hand tremors, interfering with daily activities such as eating. This study developed a spoon stabilization system controlled by a Proportional-Integral-Derivative (PID) controller, which was tuned using Particle Swarm Optimization (PSO) and the Cohen-Coon method for performance comparison. The system utilized an inertial measurement unit to detect tremors, while a Kalman filter reduced noise before a microcontroller controlled a servo motor to stabilize the spoon. The system was evaluated through simulations and hardware implementation, with performance assessed based on rise time, overshoot, delay time, and settling time. The results showed that the Kalman filter significantly reduced noise, lowering the average pitch angle error deviation from 1.028° to 0.037° and the roll angle error from 0.822° to 0.031°. The PSO-based tuning outperformed the Cohen-Coon method in response speed and system stability, achieving a faster rise time (0.09 s for roll, 0.34 s for pitch), a shorter settling time (0.74 s for roll, 0.59 s for pitch), and a lower delay time (0.1 s for roll, 0.15 s for pitch). However, the Cohen-Coon method resulted in a lower overshoot for the roll angle (6.08%) compared to the PSO-based tuning (11.98%).

XGBoost-Powered Ransomware Detection

2025-07-31T08:28:35+00:00

Ransomware remains a rapidly evolving cyber threat, causing substantial financial and operational disruptions globally. Traditional signature-based detection systems are ineffective against sophisticated, zero-day attacks due to their static nature. Consequently, machine learning-based approaches offer a more effective and adaptive alternative. This study proposes an approach utilizing XGBoost for highly effective ransomware detection. We conducted a rigorous comparative analysis of prominent ensemble learning algorithms—XGBoost, Random Forest, Gradient Boosting, and AdaBoost—on the RISS Ransomware Dataset, comprising 1,524 instances. Our experimental results unequivocally demonstrate XGBoost as the superior ensemble model, achieving an impressive 97.60% accuracy and F1-Score. This performance surpassed Gradient Boosting (97.20%), Random Forest (96.94%), and AdaBoost (96.50%). Furthermore, this study benchmarked XGBoost against established state-of-the-art (SOTA) methods, including Support Vector Machine (SVM) and the SA-CNN-IS deep learning approach. The comprehensive results confirm that XGBoost not only excels among ensemble methods but also outperforms or matches these leading SOTA techniques, solidifying its position as an exceptionally effective and adaptive solution. These findings underscore the limitations of conventional security measures and emphasize the critical need for advanced, data-driven detection methods to combat the dynamic landscape of ransomware threats.

A Metaheuristic wrapper approach to feature selection with genetic algorithm for enhancing XGBoost classification in diabetes prediction

2025-07-05T12:36:08+00:00

This study addressed the problem of selecting the most relevant features for improving the accuracy of diabetes classification using health indicator data. The research focused on a binary classification task based on the Behavioral Risk Factor Surveillance System dataset, which comprised over seventy thousand records and twenty-one predictive features related to individual health behaviors and conditions. A metaheuristic wrapper approach was developed by integrating a Genetic Algorithm for feature selection with an XGBoost classifier to evaluate the predictive quality of each feature subset. The fitness function was defined as the average classification accuracy obtained through cross-validation. In addition to feature selection, hyperparameter optimization of the XGBoost model was carried out using a Bayesian-based search strategy to further enhance performance. The proposed method successfully identified a subset of fourteen optimal features that contributed most significantly to the prediction of diabetes. The final model, combining the selected features and optimized parameters, achieved an accuracy of 0.753, outperforming baseline models trained with all features and models using features selected by deterministic methods. These results confirmed the effectiveness of combining evolutionary feature selection with model tuning to build efficient and interpretable predictive models for medical data classification. This approach demonstrated a practical solution for managing high-dimensional data in the context of chronic disease prediction.

The Application of Adaptive Neuro Fuzzy Inference System (ANFIS) Method in Estimating State of Charge (SOC) and State of Health (SOH) of Lithium-Ion Batteries

2025-07-05T12:23:29+00:00

The increasing reliance on lithium-ion batteries (LIBs) for electric vehicles and portable electronics demands accurate monitoring of battery performance, particularly the State of Charge (SOC) and State of Health (SOH). Conventional estimation methods—such as Coulomb counting, Kalman filtering, and equivalent circuit modeling—face challenges under dynamic conditions due to drift and limited adaptability. Recent studies have explored machine learning and neuro-fuzzy approaches to enhance prediction accuracy, yet many lack integration of real-time hybrid learning or struggle with high estimation error in noisy data environments. This research aims to apply the Adaptive Neuro-Fuzzy Inference System (ANFIS) to estimate SOC and SOH using experimental data from a 48V lithium-ion battery. The novelty lies in combining voltage, current, and capacity data within a MATLAB-based ANFIS framework that employs a hybrid learning algorithm integrating backpropagation and Recursive Least Squares Estimation (RLSE). Training data for SOC estimation used charging voltage and current, while SOH estimation incorporated discharging data and capacity. Results show that ANFIS achieved high accuracy with RMSE of 0.1466 and MAE of 0.021 for SOC, and RMSE of 0.012 and MAE of 0.0017 for SOH. The estimated SOH was 33.61%, closely aligned with actual values. These findings confirm ANFIS as a robust and adaptive method for real-time battery diagnostics. Future work will explore multi-input hybrid models, the integration of IoT-based BMS telemetry, and testing across diverse battery chemistries to generalize the model's performance and extend its application in smart energy systems

Implementation of Convolutional Recurrent Neural Network for Vehicle Number Plate Identification in Raspberry Pi Based Parking System

2025-05-23T03:18:23+00:00

The increasing number of vehicles in Indonesia has posed significant challenges in the management of parking facilities. This study proposes the development and implementation of an intelligent parking system based on automatic vehicle license plate character recognition. The proposed system employs the You Only Look Once version 8 (YOLOv8) model to detect the license plate region, and a Convolutional Recurrent Neural Network (CRNN) to recognize the alphanumeric characters contained within the plate. The system architecture integrates a Raspberry Pi, a camera module, and a servo motor to facilitate the automatic detection and recognition of license plates as vehicles enter and exit parking areas. The YOLOv8 model is responsible for identifying the license plate region by generating a bounding box through a convolutional layer, which is then used to isolate the license plate area from the original image. This cropped image undergoes a pre-processing stage to conform with the input specifications of the CRNN model. Subsequently, the CRNN model extracts visual features through convolutional layers and leverages recurrent layers to capture the sequential relationship among the characters on the license plate. The entire processing pipeline is deployed on the Raspberry Pi using TensorFlow Lite, ensuring efficient operation of both the YOLOv8 and CRNN models in a resource-constrained environment. Experimental results demonstrate that the YOLOv8 model achieved a detection accuracy of 94.69% for license plate localization, with a precision of 98.32%, recall of 96.25%, and an F1-score of 97.27%. In parallel, the CRNN model attained a character recognition accuracy of 93.8% across a test set comprising 30 license plates. Nevertheless, the system encountered some recognition errors, such as misclassification of the character 'G' as 'C', 'W' as 'H', and 'Q' as 'O'.

The Evolution of Image Captioning Models: Trends, Techniques, and Future Challenges

2025-05-05T05:34:31+00:00

This study provides a comprehensive systematic literature review (SLR) of the evolution of image captioning models from 2017 to 2025, with a particular emphasis on the impending problems, methodological enhancements, and significant architectural developments. The evaluation is guided by the increasing demand for precise and contextually aware image descriptions, and it adheres to the PRISMA methodology. It selects 36 relevant papers from reputable scientific databases. The results indicate a significant transition from traditional CNN-RNN models to Transformer-based architectures, which leads to enhanced semantic coherence and contextual comprehension. Current methodologies, such as prompt engineering and GAN-based augmentation, have further facilitated generalization and diversity, while multimodal fusion solutions, which incorporate attention mechanisms and knowledge integration, have improved caption quality. Additionally, significant areas of concern include data bias, equity in model assessment, and support for low-resource languages. The study underscores the fact that modern vision-language models, such as Flamingo, GIT, and LLaVA, offer robust domain generalization through cross-modal learning and joint embedding. Furthermore, the efficacy of computing in restricted environments is improved by the development of pretraining procedures and lightweight models. This study contributes by identifying future prospects, analyzing technical trade-offs, and delineating research trends, particularly in sectors such as healthcare, construction, and inclusive AI. According to the results, in order to optimize their efficacy in real-world applications, future picture captioning models must prioritize resource efficiency, impartiality, and multilingual capabilities.

Post Attack Mitigation on Open Journal System Services using Knowledge Understanding Assessment Defense (KUAD) Method

2025-03-26T10:24:20+00:00

This study was conducted to find evidence of attacks and restore data after an attack on the Open Journal System (OJS) service hosted on a computer server. The method used in this research is a new approach developed from the previous Network Forensic Digital Life Cycle (NFDLC) method. This new method, KUAD, consists of several stages for collecting evidence of cyber attacks and restoring data post-attacks. The stages in the KUAD method include initiation, acquisition, execution, mitigation, and disposition. Compared to the previous one, the novelty of this method lies in the mitigation phase, which aims to restore data or documents after an attack. The tool used to detect attacks and gather evidence is Tripwire, while the tool used to recover lost data is Crontab, which executes backup commands using rsync in four steps. Tripwire detects attacks by displaying the number of files added, deleted, or modified. This backup technique successfully recovered a hundred deleted files in .docx, .pdf, and .jpg formats. The success rate of this technique in performing post-cyber attack mitigation reaches 100%.

Development of a Web-Based Information System for Real-Time Fainting Detection Using YOLO in Smart Healthcare

2025-08-01T05:56:15+00:00

Loss of consciousness (fainting) is a critical condition that requires prompt treatment, especially in the context of elderly health services and independent patient care. This research aims to develop a web-based information system that is able to detect fainting events in real-time using the You Only Look Once (YOLO) algorithm version 11, which is one of the latest approaches in deep learning-based object detection. The system is designed to monitor video from the surveillance camera directly, make visual inferences of the patient's posture, and provide automatic notifications if a loss of consciousness condition is detected. The dataset used comes from the Roboflow platform with a total of 9,081 annotated images representing the fainting position. The YOLOv11 model was trained and tested using training data sharing, validation, and testing methods. The test results showed that the model achieved mAP, precison, recall and F1-Score values of 98.70%, 98.00%, 97.30% and 97.65%. The developed information system is able to display the detection visually through the bounding box on the dashboard and record the time of the incident. With this performance, this system shows great potential in improving patient safety through intelligent monitoring and automatic response in hospital, nursing home, and residential environments. This research also opens up opportunities for the development of more adaptive AI-based health monitoring systems and computer vision in the future.

Evaluation of Traffic Distribution Performance of ECMP and PCC+CAKE for Multi-ISP Load Balancing on Real Networks Based on Mikrotik

2025-09-12T22:52:30+00:00

Imbalance in bandwidth utilization among Internet Service Providers (ISPs) is a major challenge in network management in educational institutions, especially when differences in capacity between ISPs cause overload on one main path. To address this issue, this study proposes the application of load balancing methods using Equal Cost Multi-Path (ECMP) and Per Connection Classifier (PCC) optimized with the CAKE queue type. The implementation is carried out using MikroTik devices, which support the flexible configuration of both methods. Testing is conducted on a real network with a passive monitoring approach through the analysis of actual traffic and ISP utilization, and active monitoring. The evaluation results show that the ECMP method still produces an uneven traffic distribution, with a tendency to concentrate the load on one path. In contrast, PCC + CAKE is able to distribute traffic more evenly according to the ISP bandwidth ratio. In addition, PCC + CAKE shows more stable performance on throughput, RTT, and jitter parameters, and has very low packet loss. Therefore, PCC + CAKE is recommended as a more effective load balancing method to increase the efficiency of ISP utilization and overall network quality in a multi-ISP environment.