Jurnal Sistemasi (OJS FTIK - UNISI, Fakultas Teknik dan Ilmu Komputer Universitas Islam Indragiri)
Not a member yet
946 research outputs found
Sort by
Optimizing Driver Drowsiness Detection: Evaluating CLAHE and AHE Enhancement Techniques
Full text linkDriver drowsiness is a critical factor in road safety, and early detection can be key to preventing accidents. This research focuses on improving the accuracy of drowsiness detection by enhancing the contrast of driver facial images using image processing techniques. Specifically, the study explores the effectiveness of Adaptive Histogram Equalization (AHE) and Contrast Limited Adaptive Histogram Equalization (CLAHE) in this context. The research utilizes the Drowsy Driver Detection (DDD) dataset, which includes facial images categorized into Drowsy and Non-Drowsy classes. AHE and CLAHE techniques are applied to preprocess these images, aiming to improve contrast and subsequently enhance drowsiness detection accuracy. Evaluation metrics such as Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), and Signal-to-Noise Ratio (SNR) are employed to assess the quality of the processed images. The findings indicate that CLAHE performs better than AHE in terms of image enhancement. CLAHE achieves significantly lower MSE (93.90) compared to AHE (103.92), along with higher PSNR (28.41 for CLAHE vs. 27.97 for AHE) and SNR (0.49 for CLAHE vs. 0.04 for AHE) values. These results suggest that CLAHE effectively enhances contrast and improves image clarity. The success of CLAHE as a contrast enhancement technique highlights its potential application in real-time driver monitoring systems. In conclusion, this research underscores the importance of image preprocessing techniques like CLAHE in advancing driver safety technologies, emphasizing their potential to enhance the performance of drowsiness detection systems in practical driving scenarios
Performance Analysis of Docker-based NFV Service Chaining Networks in a Single-Host Environment
Full text linkNetwork Function Virtualization (NFV) and Service Function Chaining (SFC) enable network functions to be deployed as Virtual Network Functions (VNFs) on flexible commodity servers. However, chaining multiple VNFs within a service chain may degrade data-plane performance, particularly in container-based environments. This study analyzes the performance of container-based SFC in a single-host Docker environment under three scenarios: (1) a direct client–server connection without VNFs (baseline), (2) the addition of a single Layer 3 (L3) VNF in the form of an iptables firewall, and (3) the integration of an L3 firewall VNF combined with a Layer 4 (L4) load balancer VNF based on HAProxy. Performance evaluation was conducted by measuring TCP throughput using iperf3, end-to-end latency using ping, and CPU utilization of each container using docker stats. The results indicate that adding the L3 firewall reduces throughput by approximately 33% and nearly doubles latency compared to the baseline. Meanwhile, incorporating the L4 load balancer causes throughput degradation of up to 92%. CPU utilization analysis shows that the kernel-space firewall introduces minimal additional overhead in user space, whereas the L4 VNF becomes the primary source of CPU saturation. These findings suggest that, in container-based SFC deployments on a single-host Docker environment, performance bottlenecks are primarily driven by user-space L4 VNFs rather than kernel-based L3 forwarding. Therefore, L4 VNFs require special consideration when designing service chaining architectures for resource-constrained edge nodes
Comparison of Support Vector Machine and K-Nearest Neighbor Algorithms on the Effectiveness of a Free Lunch Program
Full text linkThe Free Lunch Program is a government initiative aimed at ensuring adequate nutrition for the public. This study aims to examine public perceptions of the program through sentiment analysis and to compare the effectiveness of Support Vector Machine (SVM) and K-Nearest Neighbor (K-NN) models. A total of 6,532 public comments were collected from Twitter, YouTube, and TikTok. After preprocessing, including normalization, stopword removal, and stemming, features were extracted using Term Frequency–Inverse Document Frequency (TF-IDF), resulting in 5,992 clean data points. The dataset was split into 80% training and 20% testing sets. Model training was conducted with hyperparameter tuning using 3-fold GridSearchCV. The results indicate that negative sentiment dominated at 42.7%. In the model comparison, SVM with a linear kernel significantly outperformed K-NN, achieving an accuracy of 72%, while K-NN (k=3) reached only 48%. These findings suggest that the SVM algorithm is more effective in classifying public opinion sentiment on high-dimensional data compared to K-NN
Analysis and Improvement of an Agribusiness Web Information System Security using Grey-Box and White-Box Testing
Full text linkThis study aims to analyze and improve the security of the SawitGoDigi Palm Oil Harvest Recording Information System using grey-box and white-box testing approaches. The system is used by farmers, agents, drivers, and administrators to manage land data, harvest results, distribution, and transaction records, which makes it highly exposed to security risks if vulnerabilities are present. The security testing process was conducted based on the OWASP Web Security Testing Guide (WSTG) v4.2 and the OWASP Risk Rating Methodology. The testing stages included reconnaissance, automated scanning using OWASP ZAP, manual exploitation, risk evaluation, implementation of security improvements, and retesting. The results revealed several significant vulnerabilities, including SQL Injection in the search feature, weak session management through the trusted_device cookie, and the absence of a rate-limiting mechanism that enabled brute-force attacks during the login process. The risk assessment indicated that SQL Injection and session hijacking were classified as High risk, while brute-force attacks were categorized as Medium risk. Security improvements were implemented through the use of prepared statements, strengthening cookie attributes, adding security headers, and implementing rate limiting. Retesting results confirmed that all identified vulnerabilities were successfully mitigated and reduced to a Low-risk level. This study demonstrates that a comprehensive security testing approach, which includes exploitation, remediation, and verification through retesting, can significantly enhance the security of agribusiness web applications. Furthermore, the findings show that before remediation, the system contained four vulnerabilities with High and Medium risk levels, namely SQL Injection, Session Hijacking, Brute-Force Login, and Security Misconfiguration. After the remediation and retesting process, all High- and Medium-risk vulnerabilities were successfully reduced to Low risk or marked as Closed, indicating that the system is secure for operational use
Predicting Impulsive Buying in Tokopedia Flash Sales: A UTAUT2 Approach
Full text linkFlash sale events have become a dominant marketing strategy to trigger rapid purchasing decisions. However, despite the massive growth of e-commerce in Indonesia, it remains unclear whether consumer participation in these events is primarily driven by the thrill of the "hunt" (hedonic) or the rational calculation of discounts (price value), particularly in developing digital markets like Palembang City. This study investigates the determinants of impulsive buying behavior during Flash Sale events on the Tokopedia platform. Drawing upon a modified Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) framework, this study investigates how Hedonic Motivation and Price Value affect Behavioral Intention, and in turn, its effect on Impulsive Buying. A quantitative methodology was applied, leveraging survey responses from 144 participants in Palembang City who had engaged in Tokopedia Flash Sales. Analysis was conducted through Partial Least Squares Structural Equation Modeling (PLS-SEM) with SmartPLS 4 software. Findings reveal that both Hedonic Motivation and Price Value positively and significantly impact Behavioral Intention, with Price Value identified as the most influential predictor. Furthermore, a robust positive relationship was found between Behavioral Intention and Impulsive Buying, confirming that the intention to participate in Flash Sales significantly drives unplanned purchasing behavior. These findings suggest that while hedonic enjoyment is important, the perceived economic benefit remains the primary catalyst for consumers. Practically, platforms can optimize flash sale design by emphasizing perceived savings and enjoyable experience to effectively drive conversion
Functional Evaluation of the Virtual Batik and Mask Museum Application using Test Scenario Based
Full text linkThe development of digital technology has created new opportunities to introduce and preserve local culture through interactive media, one of which is virtual museums. This study develops a 3D virtual museum application aimed at introducing and preserving batik and mask cultural heritage in response to the need for technology-based educational media. The functionality of the application was evaluated using a scenario-based testing method involving 10 main test scenarios. The testing results indicate an overall excellent performance, with a success rate of 85%. Four activities—including 3D object interaction (mask), wall collider functionality, background sound, and lighting—achieved very good performance. The main weakness was identified in the notification trigger feature, where four users failed to complete the assigned task. The total time required by 10 users was 167.8 seconds, with an average of 16.78 seconds per task. Overall, these results demonstrate that the application has high functional stability and is suitable for use as an innovative and interactive medium for cultural learning
Novel Genre Classification based on Synopsis using the Random Forest Algorithm
Full text linkNovel genre classification based on synopses presents a significant challenge in text processing, as each genre exhibits distinct lexical characteristics. This study evaluates the performance of the Random Forest algorithm in classifying novel genres under conditions of imbalanced data distribution. The research stages include text preprocessing—comprising case folding, tokenization, stopword removal, and stemming—feature extraction using Term Frequency–Inverse Document Frequency (TF-IDF), and model training with Random Forest. In addition, manual data balancing was applied by increasing samples in minority classes through simple oversampling. The model was evaluated using accuracy metrics and confusion matrix analysis. The results indicate that Random Forest is able to identify most genres with moderate accuracy, particularly for classes with larger data volumes. The initial model achieved an accuracy of 42.11%, which increased to 46.67% after the application of data balancing. Misclassification primarily occurred in genres with limited samples that share similar vocabulary with dominant genres. These findings demonstrate that Random Forest can still be applied to synopsis-based novel genre classification without fully relying on balancing techniques. However, performance remains uneven across classes, highlighting the need for per-genre analysis to obtain a more comprehensive evaluation
Implementation of a Hybrid Filtering Approach in a Website-based Football News Recommendation System
Full text linkThe rapid growth of football news on digital portals has made it increasingly difficult for users to find information that matches their interests. This study develops a web-based news recommendation system by combining Content-Based Filtering and Collaborative Filtering through a feature-level Hybrid Filtering approach. The proposed hybrid approach constitutes the main novelty of this research, as it does not rely on score aggregation methods commonly used in previous studies, making it lighter, simpler, and more suitable for small datasets and limited user interactions. The system employs Term Overlap Matching to measure the similarity between news titles and Cosine Similarity to assess user preference similarity based on bookmark data. The evaluation results show that Content-Based Filtering achieves the best performance, with a Precision of 0.60, Recall of 0.75, and an F1-score of 0.67, while Collaborative Filtering performs poorly due to data sparsity in user interactions, resulting in a Precision of 0, Recall of 0, and an F1-score of 0. Overall, the feature-based hybrid approach is able to provide relevant recommendations from both content and preference perspectives, although system accuracy is still predominantly driven by Content-Based Filtering. These findings indicate that the proposed simple hybrid model can serve as an effective solution for small-scale sports news platforms and has the potential to be further improved through increased data availability, enhanced user interaction, and the adoption of more advanced NLP techniques
AI-Driven Fraud Detection in Digital Banking: A Hybrid Approach using Deep Learning and Anomaly Detection
Full text linkThe rapid digital transformation in the banking sector has introduced new opportunities for efficiency and customer convenience but has also amplified the risks of financial fraud. Traditional fraud detection mechanisms, often reliant on static rule-based systems, struggle to keep pace with the dynamic, evolving nature of fraudulent activities. This paper proposes a novel hybrid framework that integrates deep learning models with anomaly detection techniques to enhance the accuracy, robustness, and adaptability of fraud detection in digital banking. The proposed approach leverages a deep neural network (DNN) architecture trained under supervised learning to capture complex transactional patterns and combines it with autoencoder-based unsupervised anomaly detection to uncover previously unseen fraud strategies. Extensive experiments on benchmark financial datasets demonstrate that the hybrid system significantly outperforms state-of-the-art methods in terms of precision, recall, and false-positive reduction. Furthermore, the study highlights the scalability of the approach for real-time banking applications and its potential for multi-institutional deployment, enabling secure inter-bank fraud intelligence sharing without compromising data privacy. Extensive experiments on benchmark financial datasets demonstrate that the hybrid system significantly outperforms state-of-the-art methods in terms of precision, recall, and false-positive reduction. Furthermore, the study highlights the scalability of the approach for real-time banking applications. This work contributes to the growing field of AI-driven financial security by addressing both detection performance and adaptability to emerging fraud behaviors
WhatsApp Hybrid Chatbot Architecture Rasa-DeepSeek: Design and Performance Evaluation
Full text linkThis study designed and evaluated a hybrid chatbot for a domain-specific application by addressing two main issues: limited NLU coverage and the variability of latency and cost when all queries are routed directly to an LLM. The proposed solution integrates a deterministic Rasa-based pipeline with a DeepSeek fallback mechanism. In this architecture, Rasa handles NLU processing, rules, stories, and context storage for mk and jk, while the LLM is only invoked when the NLU confidence score falls below a defined threshold. The methodology includes end-to-end implementation through a Node.js bridge connected to Rasa, functional testing to validate the intent–entity–action flow, and performance testing using load (stress) testing across two access paths: the Rasa REST endpoint and the Node-to-Rasa bridge. Meanwhile, the LLM pipeline was profiled separately through instrumented action calls. The results indicate that domain-specific conversations were successfully answered using curated knowledge, and both deterministic access paths met the service level objective (SLO), achieving a median latency of approximately 32 milliseconds with no observed errors. This study contributes by demonstrating that a hybrid chatbot architecture separating deterministic and generative pipelines can maintain SLO compliance in domain-specific settings. In addition, it highlights limitations of LLMs in understanding domain ontologies, reinforcing the need for semantic guardrails