Mobile and Forensics

Performance Analysis of Random Forest Algorithm with Smote for Multi-Class Attack Detection

2025-10-01T12:46:45+00:00

The increasing sophistication of cyberattacks necessitates the development of detection systems capable of accurately identifying various threat types. Data imbalance within attack logs presents a substantial challenge that can undermine the effectiveness of detection models. This study introduces a multi-class cyberattack detection model employing the Random Forest algorithm, optimized through the Synthetic Minority Over-sampling Technique (SMOTE) to address data imbalance. The innovative aspect of this research lies in integrating Random Forests and SMOTE to improve multi-class classification accuracy on local attack log datasets. This approach remains sparsely explored in academic research. The dataset consists of 3000 cyberattack logs from the Information Systems Bureau of Muhammadiyah University Purwokerto, spanning 10 cyberattack categories. The research process involved data collection, pre- processing, division, model training, and evaluation. Results indicate that the model achieved an average F1-macro score of 76% and a weighted average of 93%, with the " Threat Level Medium " feature identified as the most influential predictor. These findings suggest that the combination of Random Forest and SMOTE effectively enhances multi-class detection performance and presents promising prospects for log-based cybersecurity systems in educational and industrial environments.

Enhancing Early Diabetes Detection Using Tree-Based Machine Learning Algorithms with SMOTEENN Balancing

2025-10-01T12:53:19+00:00

Diabetes continues to be a critical global health issue, demanding accurate predictive systems to enable preventive interventions. Traditional diagnostic tests lack efficiency for large-scale early screening, which has led to growing interest in artificial intelligence solutions. This research proposed an effective methodology for diabetes classification based on tree-based algorithms enhanced with SMOTEENN balancing. The study employed the Kaggle Diabetes Prediction Dataset with 100,000 instances and eight medical and demographic features. Preprocessing steps included handling missing and duplicate values, encoding categorical variables, and scaling numerical attributes with Min-Max normalization. To address severe class imbalance, SMOTEENN was adopted, producing a cleaner and more balanced dataset. Model evaluation was performed using Stratified 5-Fold cross-validation on six classifiers: Decision Tree, Random Forest, Gradient Boosting, AdaBoost, XGBoost, and CatBoost. Experimental results indicated significant gains after balancing, with ensemble methods outperforming single-tree baselines. Random Forest delivered the best overall performance (98.93% accuracy, 98.96% F1-score, 99.16% recall, 99.94% AUC), followed by CatBoost and XGBoost with comparable results above 99% AUC. While Decision Tree benefited most from SMOTEENN in relative terms, it remained less competitive. Analysis of the importance of the analysis revealed HbA1c level and blood glucose level as dominant predictors, validating clinically meaningful learning. These findings suggest that integrating hybrid resampling with ensemble tree classifiers provides reliable and general predictions for diabetes risk. The approach holds promise for deployment in healthcare decision support systems.

Enhancing Offline Shopping Experiences With Real-Time Mobile Apps, Specifically in Batam City

2025-10-21T04:56:02+00:00

Information asymmetry in the offline retail market imposes substantial "search costs" on purchasing professionals, who frequently lack visibility into real-time product availability across physical stores. Existing solutions, such as generic store locators, fail to provide inventory context, while traditional e-commerce platforms are unable to meet immediate, same-day procurement needs due to logistical delays. This research addresses this gap by developing a Real-Time Location-Aware mobile artifact aimed at optimizing offline procurement efficiency in Batam City. Grounded in Design Science Research (DSR), the system employs a short-polling architecture implemented via Expo (React Native), Express.js, and PostgreSQL to ensure data freshness. Technical performance testing validated the system's "Near Real-Time" capabilities, achieving an average API response time of 180 ms and a stable synchronization interval of 5 seconds under 4G network conditions. Furthermore, a usability evaluation involving 40 purchasing professionals yielded an average System Usability Scale (SUS) score of 71.81, categorizing the application as "Good." These results empirically demonstrate that lightweight polling architectures can effectively mitigate cognitive load and search latency, offering a scalable software engineering solution for the "Offline-to-Offline" (O2O) retail sector.

Design and Expert Validation of AI-Supported Collaborative Digital Learning Model for Introductory Multimedia Course SPADA Indonesia

2025-12-22T03:26:01+00:00

This study develops and conceptually validates an AI-Supported Collaborative Digital Learning (AI-CDL) model for an Introduction to Multimedia course delivered through the national LMS, SPADA Indonesia. Using a Design and Development Research approach aligned with early-stage Design-Based Research, the study followed four phases: (1) contextual and needs analysis of course outcomes, commonly referred to as CPL (Capaian Pembelajaran Lulusan) and CPMK (Capaian Pembelajaran Mata Kuliah), existing learning activities, and available LMS affordances; (2) conceptual model design grounded in collaborative learning theory and multimedia learning principles; (3) development of project-based collaborative scenarios and supporting artefacts (learning paths, assessment rubrics, and responsible AI-use guidelines); and (4) conceptual validation through expert review and alignment with recent evidence syntheses on AI-supported collaboration in higher education. The resulting AI-CDL model operationalizes AI support across three layers intelligent content support, AI-supported collaboration, and AI-augmented production workflows mapped to key multimedia topics and implemented through SPADA activities. Expert feedback informed iterative refinements, particularly in task orchestration, assessment transparency, and ethical safeguards. This study contributes a validated design blueprint and transferable design principles for integrating AI into collaborative multimedia learning within a national-scale LMS. Future work will empirically evaluate learning processes and outcomes through classroom implementation and learning analytics.

Forensics of Low-Quality Facial Images from CCTV Using The Generative Adversarial Network (GAN) Method

2025-11-17T02:57:44+00:00

CCTV facial images often suffer from poor quality due to low resolution, motion blur, and poor illumination, complicating forensic investigation and identification. This situation necessitates more modern image restoration methods. This study proposes a Generative Adversarial Network (GAN)-based pipeline that combines two architectures, Real- ESRGAN for resolution enhancement and GFPGAN for more natural facial feature recovery. Experimental results show significant improvements in perceptual quality with a decrease in NIQE values from 12.56 to 7.81 and BRISQUE from 70.81 to 44.23, with an 82% image recovery success rate. Additional evaluations using texture entropy and gradient histograms demonstrate consistency in facial structure and edge sharpness. This study contributes by demonstrating an integrated two- stage GAN approach as an effective solution for face recovery in low-quality CCTV images, while highlighting the need for standardized forensic protocols and facial identity validation in real-world applications. Thus, this pipeline can serve as a pre-processing stage to improve image readability and has the potential to be used as a tool for forensic investigations.