Jurnal Ilmu Komputer dan Informatika

Comparative Analysis Of Support Vector Machine Kernels For DDOS Attack Detection: A Study Of Classification Performance And Computational Efficiency In Network Traffic

Libertino Felani Xavier Sarmento — 2026-08-08

Distributed Denial of Service (DDoS) attacks remain one of the most significant cybersecurity threats because they disrupt network services, degrade performance, and cause financial losses. Machine-learning-based intrusion detection systems have become a promising solution for automatically identifying malicious traffic. This study aims to compare the classification performance and computational efficiency of three Support Vector Machine (SVM) kernels: Linear, Polynomial, and Radial Basis Function (RBF). The experiment used 225,745 network traffic records consisting of 97,718 BENIGN flows and 128,027 DDoS flows extracted from the CICIDS2017 Friday-WorkingHours-Afternoon-DDoS subset. Data preprocessing included data cleaning, label encoding, feature selection, feature scaling using StandardScaler, and an 80:20 train-test split. The models were evaluated using Accuracy, Precision, Recall, F1-Score, Area Under the Curve (AUC), training time, and testing latency. Experimental results show that the RBF kernel achieved the best classification performance with 98.89% accuracy, 98.90% precision, 98.90% recall, 98.80% F1-score, and an AUC of 0.99. In contrast, the Linear kernel achieved the fastest computational performance with the lowest training and testing time. The novelty of this study lies in providing a kernel-level optimization analysis that simultaneously evaluates detection accuracy and computational efficiency, demonstrating that optimized traditional SVM models remain relevant for lightweight and real-time intrusion detection despite the increasing adoption of deep learning approaches.

Hybrid LBP–HOG Feature Extraction with Support Vector Machine and Deep Belief Network for Offline Signature Identification and Verification

Alvian Ardiansyah — 2026-08-04

Offline signature verification remains a challenging task due to high intra-writer variability and the presence of skilled forgeries, which can reduce the reliability of biometric authentication systems. This study aims to develop a robust offline signature identification and verification framework by integrating hybrid feature extraction and machine learning-based classification methods. The proposed approach combines Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) to capture complementary texture and shape characteristics from signature images. The extracted features are subsequently classified using Support Vector Machine (SVM) for writer identification and Deep Belief Network (DBN) for signature verification. The experimental dataset consists of signatures collected from 60 individuals, with each participant providing 8 genuine and 8 forged signatures. Prior to feature extraction, preprocessing stages including grayscale conversion, contrast enhancement, and Otsu thresholding were applied to improve image quality and segmentation consistency. Experimental results demonstrate that the proposed framework achieves an identification accuracy of 89.58% using SVM, while the DBN-based verification model attains an accuracy of 87%, an F1-score of 86%, and an AUC value of 0.85. The novelty of this study lies in the integration of hybrid LBP–HOG feature extraction with a dual-classification architecture combining SVM and DBN for offline signature authentication tasks. The combination enables the system to effectively represent both local texture patterns and structural signature characteristics while improving classification robustness against forged signatures. These findings indicate that the proposed framework provides a promising contribution to offline biometric authentication and computer vision-based signature verification research.

Application of CNN VGG16 for Digital Image-based Rice Leaf Disease Classification with Optimizer Analysis and Real-Time Validation

Lailia Rahmawati — 2025-10-10

Rice leaf diseases can significantly reduce yields, while manual identification takes a long time and risks producing errors. This research proposes a deep learning-based rice leaf disease classification system by utilizing the VGG16 Convolutional Neural Network (CNN) architecture through a transfer learning approach. The dataset was obtained from Kaggle with five categories, namely bacterial leaf blight, rice blast, brown spot, healthy, and tungro, which were processed through normalization and augmentation before being divided into training and validation data. The model was trained using two optimizers, Adam and Adadelta, with epoch variations of 20, 25, 35, and 50 to compare their performance. Experimental results showed that Adam produced the best validation accuracy of 96.24% and testing accuracy of 97.17%, with an average F1-score of 0.98, while Adadelta only achieved a maximum validation accuracy of 85.3%. Evaluation using confusion matrix and classification report further confirmed the reliability of the model, with Adam even achieving 100% accuracy on real-time testing. These findings show that the VGG16 CNN with Adam optimizer is capable of detecting rice leaf diseases quickly and accurately, thus contributing to early detection in precision agriculture and supporting food security.

Adult Gender Classification Based on Facial Images Using Convolutional Neural Networks

Salma Dewi — 2026-08-04

Gender classification in facial images is a critical component in biometric verification systems and digital service personalization. However, unconstrained environmental conditions such as lighting variations, pose, and occlusion remain major challenges in implementing reliable systems. This study aims to develop and evaluate a lightweight Convolutional Neural Network (CNN) model for gender classification specifically on individuals aged 17 and above using the filtered UTKFace dataset. The methodology includes preprocessing 19,633 images divided into training (80%) and validation data (20%), with a three-layer CNN architecture (filters 32-64-128), ReLU activation functions, MaxPooling, and 0.5 Dropout. Training applied real-time data augmentation and early stopping mechanisms to prevent overfitting. Evaluation results show a global accuracy of 93%, with female precision reaching 96% and male recall at 96%, indicating high reliability in detecting both gender classes in a balanced manner. Error analysis identified three dominant factors causing misclassification: extreme lighting (35%), occlusion (28%), and non-frontal pose (22%). These findings confirm that a lightweight CNN architecture with three convolutional layers can achieve competitive performance for gender classification in unconstrained environments with relatively low computational requirements (1.47 million parameters), making it feasible for implementation as the core unit of automated biometric verification systems on resource-constrained devices.

An Expert System for Determining Facial Treatments Using the Forward Chaining Method at Azalea Beauty Center

liana wahyu saputri — 2026-08-04

Facial skin problems such as acne, milia, warts, wrinkles, and hyperpigmentation often confuse customers in choosing appropriate treatments at Azalea Beauty Center. Manual consultations are time-consuming and heavily dependent on therapist availability. This study aims to develop an expert system using the forward chaining method to determine suitable facial skin treatments. The system was built using PHP, MySQL, and Bootstrap 5, implementing knowledge bases consisting of symptoms, skin problems, and treatment rules obtained from beauty experts. Black box testing showed all system functions operated correctlyindicating the system is highly acceptable. This expert system successfully provides fast, accurate, and consistent treatment recommendations, improving service efficiency at Azalea Beauty Center.