Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
Unifying Knowledge, Reasoning, and Hierarchy for Classifying Harmful Content
No full textThe spread of negative, engagement-driven content online causes significant societal harm, requiring advanced automated moderation tools. However, current classification systems often treat harmful content subtypes as independent, "flat" categories, which hinders their ability to thematically overlap content. This study designed and validated a novel integrated framework to accurately and transparently classify such complex cases. We proposed KG-DToT-HTC, a hybrid framework that synergistically combines three methodologies: a predefined Hierarchical Text Classification (HTC) taxonomy to structure the decision-making process; a domain-specific Knowledge Graph (KG) to provide factual, real-world context; and Decision Tree-of-Thought (DToT) prompting to guide a Large Language Model through an explicit, step-by-step reasoning process. On a real-world dataset of harmful Indonesian news, the proposed framework achieved a state-of-the-art Macro-F1 score of 0.934, representing a nearly 15-percentage point improvement over a zero-shot baseline. Ablation studies confirmed that each component—hierarchy, knowledge, and reasoning—provided a distinct and critical contribution to the final performance. The major conclusion of this study is that a synergistic architecture is essential for the accurate classification of complex harmful content. This work demonstrates a viable path toward "glass-box," interpretable AI moderation systems whose decisions are not only highly accurate but also fully auditable
Investigating Shallow Learning Methods for Optical Character Recognition of Indonesia’s Nusantara Scripts
No full textIndonesia has numerous regional scripts—or so-called Nusantara scripts—and recognizing them is important to preserve Indonesia's cultural heritage. The advances of AI and computer vision technologies make it possible for a machine to optically read the handwritten scripts through the Optical Character Recognition (OCR) technique. However, collecting some of the top OCR solutions and comprehensively investigating their performances on the Nusantara scripts is currently lacking. This study investigates and evaluates some shallow learning-based methods on our newly introduced datasets, consisting of more than 38,000-character images across 80 letter classes in total; here, we focus on three regional scripts: Javanese, Sundanese, and Balinese. The methods include Random Forest, SVM, Logistic Regression, and Gaussian Naïve Bayes, as well as boosting techniques such as XGBoost, Light GBM, and CatBoost. A 5-fold cross-validation approach assessed model performance based on accuracy, precision, recall, and F1-score. Based on the experimental results, the methods demonstrated their competitiveness in reaching the best models for scripts; in particular, XGBoost, Light GBM, and Random Forest-Gini were the winners for Javanese, Sundanese, and Balinese scripts, respectively. These findings demonstrate the effectiveness of ensemble learning methods for diverse handwritten scripts. Comparative analysis to prior deep learning studies is also discussed in this paper. In addition, this research also contributes to preserving Indonesian traditional scripts, as well as offers insights for future regional OCR in other countries
Optimizing Transformer Model FlanT5 for Multi-Question Answering with Context-Aware Learning Rate
No full textThis study investigates the performance of FlanT5-based transformer models in handling Multiple-Question Answering (M-QA) tasks, in which multiple semantically related questions must be addressed with a single cohesive answer. Unlike traditional QA systems that focus on one-to-one question-answer pairs, the M-QA approach challenges the model to understand contextual relationships across several questions tied to the same topic. A custom dataset was developed with shared context, grouped questions, and a unified answer to train and evaluate the model. The FlanT5 architecture was fine-tuned using different learning rates (0.0001, 0.0002, 0.0003) to explore the effect of training configurations on model performance. The evaluation was conducted using the ROUGE-1, ROUGE-2, ROUGE-L, and ROUGE-Lsum metrics. The results indicate that a learning rate of 0.0003 provides the optimal performance, achieving a ROUGE-Lsum score of 0.7390. This study confirms the capability of instruction-tuned transformers to manage complex summarization scenarios that require contextual coherence. The findings are relevant for real-world applications such as intelligent digital assistants, clinical decision support, and educational chatbots. Furthermore, this study emphasizes the importance of hyperparameter tuning in improving the effectiveness of question-driven summarization systems for scalable and efficient deployment
Detecting Language Anxiety in Indonesian Students: Deep Learning and Traditional Classification Methods for English Learning Anxiety
No full textMastery of the English language represents a fundamental determinant of professional achievement, particularly for individuals seeking to develop their careers and participate in international contexts. However, when learning a foreign language such as English, Indonesian students may experience language anxiety that cause them to hesitate to practicing English in the class, both in orally and in writing. For English teachers, it is crucial to identify students experiencing language anxiety early on, so that they may provide appropriate teaching strategies and interventions from the first class meeting. To address this issue, this study compares machine learning methods to provide a solution for early detection of students experiencing language anxiety. Furthermore, these methods are classification models, including LSTM, GRU, decision tree, naïve Bayes, logistic regression, and SVM. The implementation of each of these models is combined with different text representation techniques, such as Word2Vec, BERT, FastText, Glove, and TF-IDF. The advantage of our model is that despite the imbalance, limited, and smaller than baseline dataset size, this research finds that GRU with focal loss achieves the highest F1 score of 0.89. This result outperforms our baseline and thus suggests that this method is effective in detecting students who experience language anxiety
Enhancing Face Authentication for Online Examination Systems Using Median Filtering and MobileNetV2
No full textDigital transformation in higher education is driving the uptake of online tests, which require academic integrity, security, and robust user experience. In the context of authentication of users, deep learning based face recognition, in particular the Convolutional Neural Network (CNN) architectures, such as MobileNetV2, combined with intermediate filter, promises to deliver a consistent performance across a wide range of devices and imaging environments. However, there are limited comprehensive studies evaluating the final integration of the median filter and MobileNetV2 in high-value test scenarios. This study contributes by proposing an effective end-to-end Face Authentication Pipeline, assessing the median impact of filtering on MobileNetV2 performance, and validating it with a prototype application. The authentic face dataset was collected using the Teachable Machine, preprocessed with cropping, resizing, and median filtering, and then augmented through rotation, shift, shear, zoom, reversal, and brightness adjustment. The MobileNetV2 model was trained with Adam in a stepwise manner, starting with 0.001 and then 0.0001 for 20 epochs in a batch size of 32, and was evaluated for accuracy, precision, recall, and F1 score. Results show that the accuracy curve has remained stable at almost 95 percent during the 20th epoch; most grades achieved 1.00 in both precis, recall and F1, with some classings showing a limited decrease due to facial similarity or expression differences. These findings confirm that MobileNetV2 median filtering can be the basis for an effective, accurate and ready to integrate face recognition in online testing applications on a wide range of devices
Speech-Based Virtual Assistant for Mental Health Support Through Natural Interaction
No full textMental health is a significant global concern. Indonesia has reported high rates of depression and anxiety, compounded by limited emotional outlets. Although AI virtual assistants are prevalent in e-commerce and education, their application in mental health remains underexplored. Existing solutions are predominantly text-based and transactional, which restricts empathetic and natural interactions. This study develops a voice-based assistant by integrating Automatic Speech Recognition (ASR), a generative AI for empathetic responses, and a Text-to-Speech (TTS) module fine-tuned on an Indonesian dataset to adapt accent and prosody. The system underwent both technical evaluation and human testing to assess its feasibility and user experience. The results showed that the TTS model converged effectively with low loss. Human evaluation indicated 'good' interaction (MS = 3.91, SD = 0.02), 'good' AI responses (MS = 3.83, SD = 0.26), and 'fair' TTS naturalness (MOS = 3.27, SD = 0.05). Most participants found the assistant's responses meaningful, pleasant, and helpful in managing low to moderate anxiety. These results suggest that a voice-based assistant has the potential to support mental health in Indonesia. Future work should enhance speech naturalness and utilize a larger participant pool for evaluation
Securing NFT Copyright with Robust DWT-Hessenberg-SVD Watermarking and RSA Signatures
No full textIn the digital era, protecting visual content from misuse and forgery is essential. This study proposes a robust image watermarking method by integrating Discrete Wavelet Transform (DWT), Hessenberg Decomposition (HD), and Singular Value Decomposition (SVD), aiming to enhance watermark imperceptibility and resilience against common image attacks. Additionally, the system incorporates RSA digital signatures within the watermark metadata to ensure verifiable authenticity in NFT (Non-Fungible Token) applications. The method was implemented using Python and tested on multiple grayscale images across various attack scenarios, including noise addition and compression. Experimental results demonstrate high SSIM and PSNR values, confirming the method's effectiveness in maintaining both visual fidelity and embedded watermark integrity. These findings support the potential of this approach for secure and scalable NFT copyright protection
AI-Generated Narratives and Infographic Synthesis for Visualizing Climate Temperature Anomalies
No full textCommunicating long-term climate trends to non-specialist audiences remains a persistent challenge, despite the availability of well-validated global temperature datasets. While existing climate visualizations and reports provide accurate information, they often rely on expert-driven interpretations or static representations that limit accessibility and scalability. This study presents a proof-of-concept system that integrates analytical processing, rule-based narrative generation, and infographic synthesis to transform structured climate data into coherent public-facing communication artifacts. The proposed framework uses the NASA GISTEMP v4 dataset, covering annual global temperature anomalies from 1880 to 2024. It applies linear trend estimation and deterministic anomaly highlighting to extract salient temporal patterns. These analytical outputs are then translated into traceable natural language summaries and integrated with visual encodings within a single reproducible pipeline. The results confirm a persistent long-term warming trend, with several recent years exceeding high-anomaly thresholds, and demonstrate that analytical values, narrative descriptions, and visual emphasis can be generated consistently from a shared data source. Rather than introducing new climate indicators or predictive models, this study’s contribution lies in system-level integration: coupling data analysis, narrative synthesis, and visual composition into a unified, communication-oriented workflow. The framework is explicitly positioned as a proof of concept and does not claim causal attribution or empirical validation of user impact. Nonetheless, it demonstrates how transparent automation can reduce reliance on expert mediation while preserving scientific fidelity, supporting scalable climate communication systems
Automated Young Children’s Pain Detection via Facial Expressions with YOLO v11
No full textThis study demonstrates that pain detection in young children using a YOLO v11-based deep learning model can be performed effectively. By utilizing image data taken from video recordings of immunization and IV infusion procedures, then processed into photo frames and labeled using Roboflow, the model is able to provide good evaluation results. The dataset was divided into 70:20:10 for training, validation, and testing. Model performance evaluation uses accuracy, precision, recall, and F1-score metrics, and is visualized through a performance curve and confusion matrix. The results show that YOLO v11 has great potential as a pain detection method, with an [email protected] achievement of 0.893, an accuracy of 78%, a precision of 89.3%, a recall of 97%, and an F1-score of 83%. The high recall value indicates the model's excellent ability to recognize pain expressions, making it relevant for use in clinical contexts to ensure pain symptoms are not overlooked. Overall, this performance demonstrates that YOLO v11 can be a more objective and accurate approach than manual instruments, and has the potential to be developed as a tool for healthcare professionals in pediatric pain assessment
Analysis of Backdoor Shells in Web Servers Using Splunk and SPL-Based Machine Learning
No full textBackdoor shell attacks pose a critical threat to web server security, allowing attackers to bypass authentication and gain persistent, unauthorized control. Conventional signature-based detection methods often fail against these attacks due to their polymorphic and obfuscation techniques. To address this, we propose an integrated detection approach leveraging Splunk as a log management platform combined with Search Processing Language (SPL)-based machine learning (ML) models. This study collected and preprocessed web server log data using SPL queries, transforming it into structured features for classification. We evaluated two supervised learning algorithms, Logistic Regression and Random Forest, on a labeled dataset comprising both normal traffic and simulated backdoor shell attacks. The evaluation showed that while Logistic Regression achieved a solid performance with 93.5% accuracy and 87.8% recall, the Random Forest model significantly outperformed it. Random Forest reached an accuracy of 97.2%, with a precision of 95.8%, recall of 94.1%, and an F1-score of 94.9%. Crucially, it also reduced the false negative rate (FNR) to 2.3% and the false positive rate (FPR) to 3.5%, making it more reliable for real-time applications. Our findings demonstrate that Random Forest, when integrated with Splunk's SPL, provides a highly robust and practical detection mechanism that effectively distinguishes malicious activities. The primary contribution of this research is an end-to-end architecture that combines scalable log management, effective feature engineering, and advanced ML detection, offering a scalable and practical solution for enterprise-level security monitoring