Ardytha Luthfiarta
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ardytha Luthfiarta has a total of 44 co-authors. Here are the top co-authors by number of shared publications:
- Adhitya Nugraha
Universitas Dian Nuswantoro
6 shared publications Mohammad Wahyu Bagus Dwi Satya
Universitas Dian Nuswantoro
2 shared publications- Nauval Dwi Primadya
Universitas Dian Nuswantoro
2 shared publications - Sri Winarno
Universitas Dian Nuswantoro
2 shared publications Adhitya Nugraha
Informatics Engineering Faculty of Computer Science Dian Nuswantoro University Indonesia
1 shared publication- Affandy Affandy
Universitas Dian Nuswantoro
1 shared publication
Latest Publications
- 6 October 2026
- Published by Springer Nature
- p. 349-357
Retails Sales Demand Forecasting Using Bidirectional Long Short-Term Memory Technique
- 6 October 2026
- Published by Springer Nature
- p. 349-357
Retails Sales Demand Forecasting Using Bidirectional Long Short-Term Memory Technique
- Open Access
- 28 December 2025
- Published by Universitas Andalas in Jurnal Nasional Teknologi dan Sistem Informasi
- Vol. 11 (3) , 263-272
Analisis Komparatif Kinerja Llama Murni, Rag Native, dan Rag Fine-Tuning
Penelitian ini mengatasi masalah halusinasi pada Large Language Models (LLM) seperti LLaMA ketika menjawab kueri domain spesifik. Tujuan penelitian adalah membandingkan kinerja tiga arsitektur chatbot: LLaMA murni, Retrieval-Augmented Generation (RAG) berbasis LangChain (RAG Native), dan RAG Fine-Tuning. Implementasi RAG pada penelitian ini menggunakan framework LangChain sebagai sistem penghubung antara model LLaMA dan sumber pengetahuan eksternal (vector database). Framework ini menyediakan pipeline retriever-reader yang memungkinkan integrasi antara model bahasa dan data kontekstual melalui embedding serta pencarian vektor. Metode evaluasi dilakukan secara kuantitatif menggunakan metrik ROUGE-L dan BLEU pada dataset studi kasus. Hasil penelitian menunjukkan peningkatan kinerja yang progresif: arsitektur LLaMA murni (baseline) memperoleh skor ROUGE-L sebesar 46.48, implementasi RAG Native (LangChain) meningkat menjadi 61.42, dan model RAG Fine-Tuning (LangChain Optimized) mencapai kinerja tertinggi dengan skor 83.06. Penelitian ini menyimpulkan bahwa integrasi arsitektur RAG melalui framework LangChain secara signifikan meningkatkan akurasi respons chatbot, dan proses fine-tuning pada konfigurasi RAG merupakan langkah optimasi krusial untuk mencapai performa terbaik pada domain spesifik.
- 28 December 2025
- Published by Universitas Andalas in Jurnal Nasional Teknologi dan Sistem Informasi
- Vol. 11 (3) , 263-272
Open AccessAnalisis Komparatif Kinerja Llama Murni, Rag Native, dan Rag Fine-Tuning
Penelitian ini mengatasi masalah halusinasi pada Large Language Models (LLM) seperti LLaMA ketika menjawab kueri domain spesifik. Tujuan penelitian adalah membandingkan kinerja tiga arsitektur chatbot: LLaMA murni, Retrieval-Augmented Generation (RAG) berbasis LangChain (RAG Native), dan RAG Fine-Tuning. Implementasi RAG pada penelitian ini menggunakan framework LangChain sebagai sistem penghubung antara model LLaMA dan sumber pengetahuan eksternal (vector database). Framework ini menyediakan pipeline retriever-reader yang memungkinkan integrasi antara model bahasa dan data kontekstual melalui embedding serta pencarian vektor. Metode evaluasi dilakukan secara kuantitatif menggunakan metrik ROUGE-L dan BLEU pada dataset studi kasus. Hasil penelitian menunjukkan peningkatan kinerja yang progresif: arsitektur LLaMA murni (baseline) memperoleh skor ROUGE-L sebesar 46.48, implementasi RAG Native (LangChain) meningkat menjadi 61.42, dan model RAG Fine-Tuning (LangChain Optimized) mencapai kinerja tertinggi dengan skor 83.06. Penelitian ini menyimpulkan bahwa integrasi arsitektur RAG melalui framework LangChain secara signifikan meningkatkan akurasi respons chatbot, dan proses fine-tuning pada konfigurasi RAG merupakan langkah optimasi krusial untuk mencapai performa terbaik pada domain spesifik.
- Open Access
- 12 December 2025
- Published by Politeknik Negeri Batam in JOURNAL OF APPLIED INFORMATICS AND COMPUTING
- Vol. 9 (6) , 3798-3809
Evaluation of the Accuracy and Efficiency of Deep CNN Architecture in Feature Extraction for Guava Disease Classification
This study analyzes and compares several Deep Convolutional Neural Network (DCNN) architectures to evaluate the balance between classification accuracy and computational efficiency in guava fruit disease detection. A hybrid DCNN–Machine Learning (ML) approach was applied to 3,784 images from the Guava Fruit Disease Dataset using a 10-fold cross-validation scheme and undersampling techniques to address data imbalance. Six DCNN architectures were systematically tested, and the combination of ResNet50 with Artificial Neural Network (ANN) showed the best performance with an accuracy of 0.9979 and an F1-score of 0.9975, surpassing the InceptionV3 baseline (0.9974). In addition to being the most accurate, ResNet50 was also 2.5 times faster in feature extraction than DenseNet201, demonstrating an optimal balance between accuracy and time efficiency. These findings emphasize the importance of analyzing the accuracy-efficiency trade-off in selecting a DCNN architecture and open up opportunities for developing more efficient models for future agricultural image classification applications.
- 12 December 2025
- Published by Politeknik Negeri Batam in JOURNAL OF APPLIED INFORMATICS AND COMPUTING
- Vol. 9 (6) , 3798-3809
Open AccessEvaluation of the Accuracy and Efficiency of Deep CNN Architecture in Feature Extraction for Guava Disease Classification
This study analyzes and compares several Deep Convolutional Neural Network (DCNN) architectures to evaluate the balance between classification accuracy and computational efficiency in guava fruit disease detection. A hybrid DCNN–Machine Learning (ML) approach was applied to 3,784 images from the Guava Fruit Disease Dataset using a 10-fold cross-validation scheme and undersampling techniques to address data imbalance. Six DCNN architectures were systematically tested, and the combination of ResNet50 with Artificial Neural Network (ANN) showed the best performance with an accuracy of 0.9979 and an F1-score of 0.9975, surpassing the InceptionV3 baseline (0.9974). In addition to being the most accurate, ResNet50 was also 2.5 times faster in feature extraction than DenseNet201, demonstrating an optimal balance between accuracy and time efficiency. These findings emphasize the importance of analyzing the accuracy-efficiency trade-off in selecting a DCNN architecture and open up opportunities for developing more efficient models for future agricultural image classification applications.
- Open Access
- 16 October 2025
- Published by Infinite Corporation in Jurnal Teknik Informatika (Jutif)
- Vol. 6 (5) , 3018-3032
Single-Image Face Recognition For Student Identification Using Facenet512 And Yolov8 In Academic Environtment With Limited Dataset
Face recognition has become one of the most significant research areas in image processing and computer vision, mainly due to its wide applications in security, identity verification, and human and machine interaction. In this study, FaceNet512 and YOLOv8 models are used to overcome the challenges in face recognition with a limited dataset, which is only one formal photo per individual. The application of image augmentation to the model achieved 90% accuracy and ROC curve of 0.82, while the model without augmentation achieved 89% accuracy and ROC curve of 0.79. FaceNet512 showed superiority in producing more accurate and detailed facial representations compared to other models, such as ArcFace and FaceNet, especially in handling minimal facial variations. Meanwhile, YOLOv8 provides efficient face detection across various lighting conditions and viewing angles. The main challenge in this research is the low quality of the original image, which can reduce the accuracy of face recognition. These results show the great potential of using deep learning-based face recognition systems in the real world, especially for automatic attendance applications in academic environments.
- 16 October 2025
- Published by Infinite Corporation in Jurnal Teknik Informatika (Jutif)
- Vol. 6 (5) , 3018-3032
Open AccessSingle-Image Face Recognition For Student Identification Using Facenet512 And Yolov8 In Academic Environtment With Limited Dataset
Face recognition has become one of the most significant research areas in image processing and computer vision, mainly due to its wide applications in security, identity verification, and human and machine interaction. In this study, FaceNet512 and YOLOv8 models are used to overcome the challenges in face recognition with a limited dataset, which is only one formal photo per individual. The application of image augmentation to the model achieved 90% accuracy and ROC curve of 0.82, while the model without augmentation achieved 89% accuracy and ROC curve of 0.79. FaceNet512 showed superiority in producing more accurate and detailed facial representations compared to other models, such as ArcFace and FaceNet, especially in handling minimal facial variations. Meanwhile, YOLOv8 provides efficient face detection across various lighting conditions and viewing angles. The main challenge in this research is the low quality of the original image, which can reduce the accuracy of face recognition. These results show the great potential of using deep learning-based face recognition systems in the real world, especially for automatic attendance applications in academic environments.
- Open Access
- 31 August 2025
- Published by Asosiasi Prakarsa Indonesia Cerdas (APIC) in Jurnal Sistem Cerdas
- Vol. 8 (2) , 191-202
Leveraging BERT and T5 for Comprehensive Text Summarization on Indonesian Articles
One of the main challenges in the field of Natural Language Processing (NLP) is developing systems for automatic text summarization. These systems typically fall into two categories: extractive and abstractive. Extractive techniques generate summaries by selecting important sentences or phrases directly from the original text, whereas abstractive techniques focus on rephrasing or paraphrasing the content, producing summaries that resemble human-written ones. In this research, models based on Transformer architectures, including BERT and T5, were used, which have been shown to effectively summarize texts in various languages, including Indonesian. The dataset used was INDOSUM, consisting of Indonesian news articles. The best results were achieved with the T5 model, using the abstractive approach, recorded ROUGE-1, ROUGE-2, and ROUGE-L scores of 69.36%, 61.27%, and 66.17%, respectively. On the other hand, the extractive BERT model achieved ROUGE-1, ROUGE-2, and ROUGE-L scores of 70.82%, 63.99%, and 58.40%.
- 31 August 2025
- Published by Asosiasi Prakarsa Indonesia Cerdas (APIC) in Jurnal Sistem Cerdas
- Vol. 8 (2) , 191-202
Open AccessLeveraging BERT and T5 for Comprehensive Text Summarization on Indonesian Articles
One of the main challenges in the field of Natural Language Processing (NLP) is developing systems for automatic text summarization. These systems typically fall into two categories: extractive and abstractive. Extractive techniques generate summaries by selecting important sentences or phrases directly from the original text, whereas abstractive techniques focus on rephrasing or paraphrasing the content, producing summaries that resemble human-written ones. In this research, models based on Transformer architectures, including BERT and T5, were used, which have been shown to effectively summarize texts in various languages, including Indonesian. The dataset used was INDOSUM, consisting of Indonesian news articles. The best results were achieved with the T5 model, using the abstractive approach, recorded ROUGE-1, ROUGE-2, and ROUGE-L scores of 69.36%, 61.27%, and 66.17%, respectively. On the other hand, the extractive BERT model achieved ROUGE-1, ROUGE-2, and ROUGE-L scores of 70.82%, 63.99%, and 58.40%.
- Open Access
- 1 March 2025
- Published by Forum Kerjasama Pendidikan Tinggi (FKPT) in Building of Informatics, Technology and Science (BITS)
- Vol. 6 (4) , 2193-2203
Classification of Key and Time Signature in Western Musical Notation by using CRNN Algorithm with Bounding Box
This research seeks to employ the Convolutional Recurrent Neural Network (CRNN) algorithm to develop a method for classifying key and time signatures from sheet music images. The research design involved compiling a dataset of 285 sheet music images, which includes 15 types of key signatures and 19 types of time signatures. The methodology encompasses annotation using the bounding box technique, image preprocessing, and applying the CRNN model for classification using K-Fold Cross Validation because of the limited dataset. Then, the model is evaluated using the Multi Class Confusion Matrix and performance metrics. The primary findings of this study reveal that the developed model achieves 96% accuracy in key signature classification and 95% in time signature classification when utilizing bounding boxes. Conversely, the absence of bounding boxes substantially negatively impacted the accuracy of key signature classification, resulting in only a 58% accuracy rate. Time signature classification performed even worse, with an accuracy of just 19%. This research highlights the substantial accuracy enhancements achievable by incorporating bounding boxes. Therefore, we anticipate that this research will help singers, especially those in choirs, to understand and express music better using existing technologies while enhancing the accuracy of optical music recognition using the CRNN model.
- 1 March 2025
- Published by Forum Kerjasama Pendidikan Tinggi (FKPT) in Building of Informatics, Technology and Science (BITS)
- Vol. 6 (4) , 2193-2203
Open AccessClassification of Key and Time Signature in Western Musical Notation by using CRNN Algorithm with Bounding Box
This research seeks to employ the Convolutional Recurrent Neural Network (CRNN) algorithm to develop a method for classifying key and time signatures from sheet music images. The research design involved compiling a dataset of 285 sheet music images, which includes 15 types of key signatures and 19 types of time signatures. The methodology encompasses annotation using the bounding box technique, image preprocessing, and applying the CRNN model for classification using K-Fold Cross Validation because of the limited dataset. Then, the model is evaluated using the Multi Class Confusion Matrix and performance metrics. The primary findings of this study reveal that the developed model achieves 96% accuracy in key signature classification and 95% in time signature classification when utilizing bounding boxes. Conversely, the absence of bounding boxes substantially negatively impacted the accuracy of key signature classification, resulting in only a 58% accuracy rate. Time signature classification performed even worse, with an accuracy of just 19%. This research highlights the substantial accuracy enhancements achievable by incorporating bounding boxes. Therefore, we anticipate that this research will help singers, especially those in choirs, to understand and express music better using existing technologies while enhancing the accuracy of optical music recognition using the CRNN model.
- Open Access
- 10 February 2025
- Published by Infinite Corporation in Jurnal Teknik Informatika (Jutif)
- Vol. 6 (1) , 11-22
AN ENHANCED MULTI-LAYERED IMAGE ENCRYPTION SCHEME USING 2D HYPERCHAOTIC CROSS-SYSTEM AND LOGISTIC MAP WITH ROUTE TRANSPOSITION
In the rapidly evolving digital era, image encryption has become a crucial technique to protect visual data from the threat of information leakage. However, the main challenge in image encryption is improving security against cryptanalysis attacks, such as brute-force and differential attacks, which can compromise the integrity of the encrypted image. Additionally, the creation of efficient and fast encryption schemes that do not degrade image quality remains a significant challenge. This research proposes a multi-layer image encryption scheme that integrates the Logistic Map algorithm, Cross 2D Hyperchaotic (C2HM) system, and Route Transposition techniques. The method aims to enhance the security of digital image encryption by combining chaotic and hyperchaotic systems. The Logistic Map is used to generate a sequence of random values with high chaotic properties, while C2HM contributes to increasing complexity and variability. The Route Transposition technique is applied to scramble pixel positions, further strengthening the encryption’s randomness. The encryption key is derived from a combination of the image hash and user key, which are then used to calculate the initial seed in the chaotic algorithm. Experiments were conducted using standard images with a resolution of 512×512 pixels. The security analysis includes evaluations of NPCR, UACI, histogram analysis, and information entropy. The experimental results show that NPCR consistently exceeds 99.5%, while UACI ranges between 33.23% and 33.56%, indicating high sensitivity to minor changes. Histogram analysis demonstrates an even intensity distribution, and the information entropy value of 7.999 reflects an exceptionally high level of randomness. Robustness tests also indicate that this method can maintain image integrity even when subjected to damage or data loss.
- 10 February 2025
- Published by Infinite Corporation in Jurnal Teknik Informatika (Jutif)
- Vol. 6 (1) , 11-22
Open AccessAN ENHANCED MULTI-LAYERED IMAGE ENCRYPTION SCHEME USING 2D HYPERCHAOTIC CROSS-SYSTEM AND LOGISTIC MAP WITH ROUTE TRANSPOSITION
In the rapidly evolving digital era, image encryption has become a crucial technique to protect visual data from the threat of information leakage. However, the main challenge in image encryption is improving security against cryptanalysis attacks, such as brute-force and differential attacks, which can compromise the integrity of the encrypted image. Additionally, the creation of efficient and fast encryption schemes that do not degrade image quality remains a significant challenge. This research proposes a multi-layer image encryption scheme that integrates the Logistic Map algorithm, Cross 2D Hyperchaotic (C2HM) system, and Route Transposition techniques. The method aims to enhance the security of digital image encryption by combining chaotic and hyperchaotic systems. The Logistic Map is used to generate a sequence of random values with high chaotic properties, while C2HM contributes to increasing complexity and variability. The Route Transposition technique is applied to scramble pixel positions, further strengthening the encryption’s randomness. The encryption key is derived from a combination of the image hash and user key, which are then used to calculate the initial seed in the chaotic algorithm. Experiments were conducted using standard images with a resolution of 512×512 pixels. The security analysis includes evaluations of NPCR, UACI, histogram analysis, and information entropy. The experimental results show that NPCR consistently exceeds 99.5%, while UACI ranges between 33.23% and 33.56%, indicating high sensitivity to minor changes. Histogram analysis demonstrates an even intensity distribution, and the information entropy value of 7.999 reflects an exceptionally high level of randomness. Robustness tests also indicate that this method can maintain image integrity even when subjected to damage or data loss.
- Open Access
- 2 January 2025
- Published by Politeknik Ganesha in sinkron
- Vol. 9 (1) , 55-64
Leveraging Label Preprocessing for Effective End-to-End Indonesian Automatic Speech Recognition
This research explores the potential of improving low-resource Automatic Speech Recognition (ASR) performance by leveraging label preprocessing techniques in conjunction with the wav2vec2-large Self-Supervised Learning (SSL) model. ASR technology plays a critical role in enhancing educational accessibility for children with disabilities in Indonesia, yet its development faces challenges due to limited labeled datasets. SSL models like wav2vec 2.0 have shown promise by learning rich speech representations from raw audio with minimal labeled data. Still, their dependence on large datasets and significant computational resources limits their application in low-resource settings. This study introduces a label preprocessing technique to address these limitations, comparing three scenarios: training without preprocessing, with the proposed preprocessing method, and with an alternative method. Using only 16 hours of labeled data, the proposed preprocessing approach achieves a Word Error Rate (WER) of 15.83%, significantly outperforming the baseline scenario (33.45% WER) and the alternative preprocessing method (19.62% WER). Further training using the proposed preprocessing technique with increased epochs reduces the WER to 14.00%. These results highlight the effectiveness of label preprocessing in reducing data dependency while enhancing model performance. The findings demonstrate the feasibility of developing robust ASR models for low-resource languages, offering a scalable solution for advancing ASR technology and improving educational accessibility, particularly for underrepresented languages.
- 2 January 2025
- Published by Politeknik Ganesha in sinkron
- Vol. 9 (1) , 55-64
Open AccessLeveraging Label Preprocessing for Effective End-to-End Indonesian Automatic Speech Recognition
This research explores the potential of improving low-resource Automatic Speech Recognition (ASR) performance by leveraging label preprocessing techniques in conjunction with the wav2vec2-large Self-Supervised Learning (SSL) model. ASR technology plays a critical role in enhancing educational accessibility for children with disabilities in Indonesia, yet its development faces challenges due to limited labeled datasets. SSL models like wav2vec 2.0 have shown promise by learning rich speech representations from raw audio with minimal labeled data. Still, their dependence on large datasets and significant computational resources limits their application in low-resource settings. This study introduces a label preprocessing technique to address these limitations, comparing three scenarios: training without preprocessing, with the proposed preprocessing method, and with an alternative method. Using only 16 hours of labeled data, the proposed preprocessing approach achieves a Word Error Rate (WER) of 15.83%, significantly outperforming the baseline scenario (33.45% WER) and the alternative preprocessing method (19.62% WER). Further training using the proposed preprocessing technique with increased epochs reduces the WER to 14.00%. These results highlight the effectiveness of label preprocessing in reducing data dependency while enhancing model performance. The findings demonstrate the feasibility of developing robust ASR models for low-resource languages, offering a scalable solution for advancing ASR technology and improving educational accessibility, particularly for underrepresented languages.
- Open Access
- 29 November 2024
- Published by Fakultas Teknik Elektronika dan Komputer Universitas Kristen Satya Wacana in Techné : Jurnal Ilmiah Elektroteknika
- Vol. 23 (2)
Optimizing Imbalanced Data Classification: Under Sampling Algorithm Strategy with Classification Combination
The security of Internet of Things devices is a factor that must be considered because device damage and data theft can occur. Internet of Things devices are very useful in various sectors, such as health, transportation, and industrial sectors. Attacks on Internet of Things devices increase every year. To overcome this, it is necessary to take a research approach with machine learning. The dataset used is CIC IoT Attacks 2023 from the University Of New Brunswick. To be able to produce good data, it is necessary to do random under sampling as a way to overcome data imbalance. Then, modeling is done using the KNN algorithm, Random Forest, Logistic Regression, Adaboost, And Perceptron. The result of this research is that random forest has the best accuracy result of 99.73%. From these results, it can be concluded that the random under-sampling technique can improve the accuracy of data imbalance.
- 29 November 2024
- Published by Fakultas Teknik Elektronika dan Komputer Universitas Kristen Satya Wacana in Techné : Jurnal Ilmiah Elektroteknika
- Vol. 23 (2)
Open AccessOptimizing Imbalanced Data Classification: Under Sampling Algorithm Strategy with Classification Combination
The security of Internet of Things devices is a factor that must be considered because device damage and data theft can occur. Internet of Things devices are very useful in various sectors, such as health, transportation, and industrial sectors. Attacks on Internet of Things devices increase every year. To overcome this, it is necessary to take a research approach with machine learning. The dataset used is CIC IoT Attacks 2023 from the University Of New Brunswick. To be able to produce good data, it is necessary to do random under sampling as a way to overcome data imbalance. Then, modeling is done using the KNN algorithm, Random Forest, Logistic Regression, Adaboost, And Perceptron. The result of this research is that random forest has the best accuracy result of 99.73%. From these results, it can be concluded that the random under-sampling technique can improve the accuracy of data imbalance.
- Open Access
- 29 November 2024
- Published by Fakultas Teknik Elektronika dan Komputer Universitas Kristen Satya Wacana in Techné : Jurnal Ilmiah Elektroteknika
- Vol. 23 (2)
Optimizing Performance of AdaBoost Algorithm through Undersampling and Hyperparameter Tuning on CICIoT 2023 Dataset
The increasing prevalence of the Internet of Things (IoT) in various sectors presents new challenges related to security and protection against cyberattacks. The connection of IoT devices to the Internet network makes them vulnerable to various types of attacks. One approach to attacking IoT devices is to perform analysis based on network traffic using machine learning algorithms such as AdaBoost. An IoT device attack prediction model was created for the purpose of predicting IoT device attacks based on network traffic. Based on research and discussion regarding optimization of the n_estimator value and algorithm in the AdaBoost algorithm on the CICIoT 2023 dataset that has been undersampled and using the grid search cv method, the most optimal n_estimator value is 500 and the most optimal algorithm value is SAMME with an accuracy rate of 0.78 and a recall value of 0.78. This optimization underscores the significance of finetuning parameters in machine learning algorithms to enhance the effectiveness of cybersecurity measures for IoT devices.
- 29 November 2024
- Published by Fakultas Teknik Elektronika dan Komputer Universitas Kristen Satya Wacana in Techné : Jurnal Ilmiah Elektroteknika
- Vol. 23 (2)
Open AccessOptimizing Performance of AdaBoost Algorithm through Undersampling and Hyperparameter Tuning on CICIoT 2023 Dataset
The increasing prevalence of the Internet of Things (IoT) in various sectors presents new challenges related to security and protection against cyberattacks. The connection of IoT devices to the Internet network makes them vulnerable to various types of attacks. One approach to attacking IoT devices is to perform analysis based on network traffic using machine learning algorithms such as AdaBoost. An IoT device attack prediction model was created for the purpose of predicting IoT device attacks based on network traffic. Based on research and discussion regarding optimization of the n_estimator value and algorithm in the AdaBoost algorithm on the CICIoT 2023 dataset that has been undersampled and using the grid search cv method, the most optimal n_estimator value is 500 and the most optimal algorithm value is SAMME with an accuracy rate of 0.78 and a recall value of 0.78. This optimization underscores the significance of finetuning parameters in machine learning algorithms to enhance the effectiveness of cybersecurity measures for IoT devices.