Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
Securing Electronic Medical Documents Using AES and LZMA
No full textWith increasing threats in cyberspace, maintaining the integrity of electronic medical data is crucial. This study aims to develop a method that integrates encryption using Advanced Encryption Standard (AES) and compression with the Lempel-Ziv-Markov Algorithm (LZMA) to protect DICOM files containing sensitive information. This method is designed to address two main challenges: the growth of file sizes after the encryption process and the efficiency in data storage. In this study, an experimental design with random sampling was applied, testing 427 DICOM files from open libraries ranging in size from 513.06 KB to 513.39 KB to evaluate the implementation of this method in reducing file size, encryption time, and maintaining data integrity. The results show that this method is able to reduce file size by between 40-50% with an average encryption time of about 0.2-0.3 seconds per file. In addition, the data remains intact before and after the encryption process, which indicates that the integrity of the data is well maintained. Further analysis revealed that CPU usage during the encryption process reached 94.05%, while memory usage was recorded at 92.95 KB. In contrast, in the decryption process, CPU usage decreased to 78.16% with a much lower memory consumption, which was 31.07 KB. The findings have significant implications for medical information systems, allowing developers to easily implement these methods through APIs. This research is expected to be a reference for future studies that focus on data security in health information systems and provide new insights into the combination of encryption and compression in the context of medical data
The Internet-of-Things-based Fishpond Security System Using NodeMCU ESP32-CAM Microcontroller
No full textFish theft in ponds is a common problem, especially in freshwater fish farms. To solve this problem, a security system that can detect human movement and provide real-time notifications is needed. This research aims to design and implement an Internet of Things (IoT)-based fishpond security system using NodeMCU ESP32-CAM Microcontroller equipped with HB100 Radar Sensor to detect human entity movement with NodeMCU ESP32-CAM to take pictures of the approaching human entity, as well as Arduino Uno R3 to control system inputs and outputs. The system also sends real-time notifications and can be managed independently by a social media application. The results show that the system can detect human movement well, provide real-time notifications, and be handled easily. The test results show that the HB100 Radar Sensor can detect entities with a maximum distance of 9 meters with overall accuracy of 90%, the Buzzer performs well according to the human entity detected by the sensor, the Arduino Uno R3 successfully sends a trigger signal to NodeMCU ESP32-CAM to activate the OV2640 camera to capture the detected human entities with a maximum distance of up to 60 meters with an optimal distance between 1 to 9 meters. Integrated system test results show that all components of the fishpond security syste
CNN Performance Improvement for Classifying Stunted Facial Images Using Early Stopping Approach
No full textStunting, a condition characterised by short stature, is a growth disorder caused by chronic malnutrition, which often begins in the womb. Children affected by stunting usually show different physical and cognitive characteristics compared to their peers. Research shows that these physical differences can also be observed in facial features. Because faces provide important information and are commonly studied in digital image processing, in this study, we will compare the facial image classification performance of stunted children versus normal children using various Convolutional Neural Network (CNN) architectures. The evaluated architectures include MobileNetV2, InceptionV3, VGG19, ResNet18, EfficientNetB0, and AlexNet. To improve the learning process, augmentation techniques with Haar cascade and Gaussian filters were applied so that the data set increased from 1,000 to 6,000 images. After adding the dataset, training is carried out with an early stop approach to minimise overfitting. The main aim of this research is to identify the CNN model that is most effective in differentiating facial images of stunted children from normal children. The results show that the EfficientNetB0 architecture outperforms other models, achieving 100% accuracy. Early stopping has been shown to improve training efficiency and help prevent overfitting
Measuring Factors of Trust in the Use of E-Government: A Multi-Factor Analysis of the E-Government in Indonesia
No full textThe implementation of dynamic records management applications within the Indonesian government remains relatively limited, with a lack of comprehensive integration between authorised institutions at both the central and regional levels. This research examines the impact of technical aspects, government agency variables, citizen variables, and risk indicators on trust in e-government. Furthermore, this study seeks to establish the effect of social factors and the advantages of trust in e-government. Finally, this research shows how trust in e-government influences satisfaction, willingness to use, and acceptance of e-government. The study examined 117 respondents using the integrated dynamic archival information system - SRIKANDI. Technical and risk factors were found to positively influence trust in e-government, with effects on satisfaction, intention to use, and adoption of e-government. Those who trusted SRIKANDI were more likely to utilize and implement the program. The findings indicate that for civil servants, trust in the government is also a factor influencing the utilisation of e-government services
Comparing Word Representation BERT and RoBERTa in Keyphrase Extraction using TgGAT
No full textIn this digital era, accessing vast amounts of information from websites and academic papers has become easier. However, efficiently locating relevant content remains challenging due to the overwhelming volume of data. Keyphrase Extraction Systems automate the process of generating phrases that accurately represent a document’s main topics. These systems are crucial for supporting various natural language processing tasks, such as text summarization, information retrieval, and representation. The traditional method of manually selecting key phrases is still common but often proves inefficient and inconsistent in summarizing the main ideas of a document. This study introduces an approach that integrates pre-trained language models, BERT and RoBERTa, with Topic-Guided Graph Attention Networks (TgGAT) to enhance keyphrase extraction. TgGAT strengthens the extraction process by combining topic modelling with graph-based structures, providing a more structured and context-aware representation of a document’s key topics. By leveraging the strengths of both graph-based and transformer-based models, this research proposes a framework that improves keyphrase extraction performance. This is the first to apply graph-based and PLM methods for keyphrase extraction in the Indonesian language. The results revealed that BERT outperformed RoBERTa, with precision, recall, and F1-scores of 0.058, 0.070, and 0.062, respectively, compared to RoBERTa’s 0.026, 0.030, and 0.027. The result shows that BERT with TgGAT obtained more representative keyphrases than RoBERTa with TgGAT. These findings underline the benefits of integrating graph-based approaches with pre-trained models for capturing both semantic relationships and topic relevance
Advancing Vehicle Logo Detection with DETR to Handle Small Logos and Low-Quality Images
No full textImage-based vehicle logo detection is an important component in the implementation of vehicle information recognition technology, which supports the development of intelligent transportation systems. Vehicle logos, as elements that represent the identities of vehicle brands and models, play a significant role in completing vehicle identity data. The information obtained from this logo can be utilized to solve various traffic problems, such as vehicle document counterfeiting and theft, and for better traffic planning and management purposes. However, the main challenge in developing an accurate logo detection system lies in the wide variety of shapes, sizes, and positions of logos in different types of vehicles. In addition, the generally small size of logos, especially on certain vehicles, often makes it difficult for computer-based detection systems to recognize logos consistently, thus affecting the overall performance of the detection model. In this research, the Detection Transformers (DETR) method is used to build a vehicle logo detection system that focuses on small-scale logo. The testing process was conducted using the VL-10 dataset, which was specifically designed for vehicle logo detection evaluation. The results show that the DETR model can detect vehicle logos very well, even for small-scale logos. The model achieved an AP50 value of 0.952, which indicates a high level of accuracy and reliability in detecting the vehicle logo in the dataset used
Breast Cancer Histopathological Image Classification with Convolutional Neural Networks Models
No full textEarly diagnosis and treatment can reduce mortality rates by preventing the progression of breast cancer. Owing to convolutional neural networks (CNN), breast cancer diagnosis can be performed faster and more objectively than humans using thousands of histopathological images. This study aimed to evaluate and compare the rapid and effective diagnostic performance of CNN models on breast tumor images, utilizing transfer learning through pre-training and fine-tuning on novel datasets. The study was performed in two ways on BreakHis and BACH datasets. First, fine-tuned VGG16, VGG19, Xception, InceptionV3, ResNet50, and InceptionResNetV2 models were used for classification. Second, these CNN models were used as feature extractors and support vector machines (SVMs) as classifiers. The success of all models in tumor classification was interpreted using performance metrics, such as accuracy, precision, recall, F1 score, and AUC. The models showing the best performance as a result of the analyses were as follows: InceptionResNetV2+SVM model with an accuracy of 99.3%, precision of 99.0%, recall of 100.0%, F1 score of 99.5%, AUC of 98.9% for BreakHis dataset; and InceptionResNetV2 model with accuracy of 96.7%, precision of 93.8%, recall of 100.0%, F1 score of 96.8%, AUC of 96.7% for the BACH dataset. As a conclusion, it has been seen that the CNN methods have good generalization abilities and can respond to clinical needs
Open-Set Recognition for Potato Leaf Disease Identification Using OpenMax
No full textTraditional methods for identifying potato leaf diseases rely on manual visual inspection, which is prone to human error and inefficiency. While machine learning models have improved automation, conventional closed-set classifiers fail to recognize unknown diseases outside their training scope, limiting real-world applicability. This study addresses this gap by implementing Open-Set Recognition (OSR) using the OpenMax framework to classify known potato leaf diseases while effectively rejecting unknown pathologies. Leveraging the Xception architecture with dual learning schedulers (ReduceLROnPlateau and StepLR), we optimized OpenMax parameters, including distance metrics (Euclidean, Eucos) and rejection thresholds. After rigorous tuning, the model achieved 86.8% accuracy and 86.4% F1-score under an openness score of 18.3%, with optimal performance using Euclidean distance and a 0.95 threshold. The results demonstrate robust discrimination between known classes (potato late blight, early blight, healthy leaves) and visually similar unknown classes (e.g., tomato diseases, healthy bell peppers). This work enhances AI-driven agricultural diagnostics by bridging the gap between closed-set precision and open-set practicality, offering a scalable solution for real-world disease identification where novel pathogens may emerge
Sentiment Analysis Optimization Using Ensemble of Multiple SVM Kernel Functions
No full textThis research aims to optimize sentiment analysis by leveraging the strengths of multiple Support Vector Machine (SVM) kernels—Linear, RBF, Polynomial, and Sigmoid—through an ensemble learning approach. This study introduces a novel model called SVM Porlis, which integrates these kernels using both hard and soft voting strategies to improve the classification performance on imbalanced datasets. Sentiment classification in this study involves two classes: positive and negative. Tweets related to the controversy over the naturalization of Indonesian national football players were collected using the official X/Twitter API, resulting in a dataset of 2,248 entries. The dataset was notably imbalanced, with significantly more negative samples than positive samples. Data preprocessing included cleaning, labeling, tokenization, stopword removal, stemming, and feature extraction using TF-IDF. To address the class imbalance, the SMOTE technique was applied to synthetically augment the minority class. Each SVM kernel was trained and evaluated individually before being combined into an SVM Porlis model. Evaluation metrics included accuracy, precision, recall, F1-score, and confusion matrix analysis. The results demonstrate that SVM Porlis with soft voting achieved the highest performance, with 98% accuracy, precision, recall, and F1-score, surpassing the performance of individual kernels and other ensemble approaches such as SVM + Chi-Square and SVM + PSO. These findings highlight the effectiveness of combining multiple kernels to capture both linear and non-linear patterns, offering a robust and adaptive solution for sentiment analysis in real-world, imbalanced data scenarios
BERT Model Fine-tuned for Scientific Document Classification and Recommendation
No full textThe increasing number of academic documents requires efficient and accurate classification and recommendation systems to assist in retrieving relevant information. This system is built using the "bert-base-uncased” model from Hugging Face, which has been fine-tuned to improve the classification accuracy and relevance of document recommendations. The dataset used consists of 2.000 academic documents in the field of computer science, with features including titles, abstracts, and keywords, which were combined into a single input for the model. Document similarity is measured using cosine similarity, resulting in recommendations based on semantic proximity. Unlike traditional approaches, which rely primarily on word frequency or surface-level matching, the proposed method leverages BERT’s contextual embeddings to capture deeper semantic meanings and relationships between documents. This allows for more accurate classification and more context-aware recommendations. Evaluation results show that the best model configuration (learning rate 3e-5, batch size 32, optimizer AdamW) achieved 89.5% training accuracy and an F1-score of 0.8947, while testing yielded 91% accuracy and 90% F1-score. The recommendation system consistently produced Precision@k values above 92% for k between 5 and 30, with Recall@k reaching 1.0 as k increased. These results indicate that the system not only performs reliably in classifying complex academic texts but also effectively recommends contextually relevant documents. This integrated approach shows strong potential for enhancing academic document retrieval and supports the development of semantically aware information management systems