Latest Issue
Vol. 52,
No. 10,
Oct.
2025
All Issues
A Protected Key-based Table Generation Methodfor White-Box Cryptography in IoT Environments
http://doi.org/10.5626/JOK.2025.52.10.805
IoT environments, there is a significant risk of cryptographic key exposure, compounded by resource constraints that limit the use of traditional hardware-based key protection mechanisms. To address this issue, this paper introduces a method for generating protected key-based table for white-box cryptography tables. In this approach, an externally encoded protected key is inserted into the system to generate internal reference tables. The proposed algorithm ensures that the plain key does not reside in memory during the table generation process, effectively reducing the risk of key leakage during internal operations. The outputs of the algorithm are mathematically proven to be functionally equivalent to those produced by conventional table generation methods, while the size of the inserted protected key is reduced to approximately 1/270, enhancing both storage and transmission efficiency. Although the method is illustrated using AES-128, it is applicable to other SPN-based block ciphers as well. These findings indicate that the proposed scheme can serve as a secure and practical solution for key delivery and protection in resource-constrained IoT environments.
Simulation-Based Worst-Case Performance Estimation in Vehicle Application Handling Dynamic Events
Sungmin Kim, EunJin Jeong, Soonhoi Ha
http://doi.org/10.5626/JOK.2025.52.10.813
The rise of Software-Defined Vehicle (SDV) and the adoption of centralized electronic/ electrical (E/E) architectures have accelerated the integration of Electronic Control Units (ECUs), resulting in increased communication complexity and data throughput in automotive applications. In such environments, the dynamic nature of event-driven task models complicates the prediction of response times and memory requirements. This paper presents a conservative yet realistic methodology for analyzing worst-case scenarios. It combines system profiling with Envelope Arrival Curves, superset-based event path grouping, and statistical event distribution techniques. By utilizing profiling logs collected from an actual automotive embedded system, the proposed method constructs worst-case execution scenarios and evaluates both Worst-Case Response Time (WCRT) and memory requirements through simulation. Experimental results demonstrate the impact of profiling log composition and task priority adjustments on system performance.
Lowest Level Term Recommendation in MedDRA-Based Medical Coding Using Subword Tokenization
http://doi.org/10.5626/JOK.2025.52.10.825
In clinical trials, the terms reported by investigators for subjects’ medical history and adverse events are called Reported Terms (RT). Since the same symptom can be expressed differently by each investigator, a medical coding process that maps RTs to the MedDRA standard Low-Level Terms (LLTs) is essential. While English-based systems exist, domestic studies face challenges due to the mixture of Korean and English. This study proposes a method to convert RTs, including Korean expressions, into LLTs. We constructed a training dataset of 4,398 RT–LLT pairs collected from domestic clinical trials and applied subword tokenization algorithms—SentencePiece and WordPiece—to handle noise such as multilingual inputs, spacing, typos, and punctuation. Using the token vocabulary generated during training, we segmented new RTs and implemented an algorithm that recommends top-k LLTs based on matching scores. Test results showed that the correct LLT was included within the top five candidate list on average.
Anomaly Detection in Electrocardiogram using Bidirectional Long Short-Term Memory Residual Networks
Yinxian He, Bokyung Amy Kwon, Kyungtae Kang
http://doi.org/10.5626/JOK.2025.52.10.833
Recent advances in artificial intelligence (AI) have significantly enhanced classification tasks in medical domains, particularly in signal analysis, such as electrocardiograms (ECGs). This study focuses on detecting obstructive sleep apnea (OSA), a prevalent respiratory disorder, through signal analysis. OSA presents distinct yet irregular patterns in ECG signals, especially in the early stages, which pose challenges for detection using traditional methods or even deep learning approaches. To address these challenges, we propose a hybrid model that combines Bi-directional Long Short-Term Memory (Bi-LSTM) with a Residual Network (ResNet-18) specifically tailored for OSA detection. This model effectively leverages the abstract features and temporal dependencies of ECG signals to identify OSA patterns. Using ECG data from the PSG-Audio dataset, our approach achieved an accuracy of 94.65% (maximum 97.84%), precision of 94.65% (maximum 98.13%), sensitivity of 93.92% (maximum 100%), and an average F1 score of 0.94 (maximum 0.979), outperforming the Bi-LSTM-CNN combined model proposed in the previous study. Our findings demonstrate the potential of this model as a practical application for home-based health monitoring, providing an efficient, non-invasive solution for early OSA detection.
Korean Benchmark for Science and Technology Information Domain to Evaluate Large Language Models
Donghyeok Koh, Jeonghun Yuk, Byeongho Lee, Kyungtae Lim, Kyongha Lee, Taehoon Kim, Cheoneum Park
http://doi.org/10.5626/JOK.2025.52.10.841
We present a new Korean benchmark dataset for evaluating large language models (LLMs) in the science and technology information domain. We define eight specialized categories within scientific and technical fields, establishing levels and evaluation metrics for each question type to differentiate the science and technology information domain from the general domain. Our two-step synthetic data generation method enhances benchmark quality and domain specificity while reducing the time required for dataset construction. Experimental results with LLMs show accuracy scores ranging from 0.17 to 0.32 across multiple-choice question types, validating the benchmark’s effectiveness for the science and technology domains. This work provides a critical foundation for evaluating the domain-specific capabilities of language models trained for artificial general intelligence (AGI), marking Korea’s first specialized benchmark in this field.
Asynchronous-Parallel Sharpness-Aware Minimization for Efficient Deep Learning
Junhyuk Jo, Jihyun Lim, Sunwoo Lee
http://doi.org/10.5626/JOK.2025.52.10.851
Sharpness-Aware Minimization (SAM) is an optimization technique designed to enhance the generalization performance of machine learning models. However, its high computational cost associated with model perturbation has restricted its use in real-world applications. In this paper, we propose a novel asynchronous-parallel SAM, which decouples the data dependency between model perturbation and update steps, enabling efficient gradient norm regularization. By adjusting the perturbation batch size in a system-aware manner, our method fully hides the perturbation overhead and effectively utilizes heterogeneous resources such as CPUs and GPUs without sacrificing accuracy. Our experiments on the CIFAR and Oxford Flowers 102 benchmarks show that asynchronous SAM achieves 1-4% higher accuracy than SGD and slightly outperforms the original SAM. Additionally, it maintains accuracy comparable to recent SAM variants, while on CIFAR-10 with ResNet-20, it demonstrates a training speed comparable to that of SGD (about 1.02×).
Sylender: A Syllable-Enhanced Transformer Encoder Model Incorporating Korean Characteristics
Yumin Heo, Jiwon Heo, Minjun Choi, Youngjoong Ko
http://doi.org/10.5626/JOK.2025.52.10.860
While syllable-level tokenization better preserves grammatical and linguistic features, it is often less semantically informative, resulting in lower performance. This paper introduces Sylender, a model that enhances existing pretrained subword-based language models by incorporating syllable-level information. Sylender adds a syllable-level transformer module to each layer of the subword model, utilizing both subword and syllable embeddings. This parallel structure retains the benefits of subword representations while effectively integrating syllable-level information, thereby improving the model's ability to capture Korean linguistic characteristics. Experiments across multiple Korean NLP tasks demonstrate that Sylender outperforms strong baselines and even larger models, validating the effectiveness of combining subword and syllable-level representations tailored to the nuances of the Korean language.
Persona Extraction System via Quadruple Analysis
http://doi.org/10.5626/JOK.2025.52.10.869
In developing AI systems that generate personalized conversations, understanding persona is crucial for reflecting a user’s characteristics. In this study, we introduce a quadruple structure- comprising Core, Expression, Sentiment, and Category-as a new way to accurately extract persona information from utterances and leverage it for response generation. We construct a quadruple dataset using LLM-based automatic annotation and implement various extraction models for comparison. After training the model, we enhance data quality through additional verification before retraining to obtain the final model. By supplying the extracted quadruple information to a response generation model, we evaluate performance differences across various persona representation methods and observe improved outcomes. This work advances persona extraction by structuring dynamic persona information to encompass sentiment and category levels, resulting in a more fine-grained and effective extraction system.
GraphSAGE-based Embedding for Performance Enhancement in Time Series Classification
Sanghun Lee, Hong-Jun Jang, Yang-Sae Moon
http://doi.org/10.5626/JOK.2025.52.10.879
With the recent increase in graph research, studies analyzing time series within the graph domain have emerged. SimTSC is a novel approach that transforms time series into graphs for node classification; however, it does not account for the relationships between nodes in its feature embedding. To address this issue, we propose SbCM(SAGE-based Classification Model), which performs node feature embedding using GraphSAGE. Additionally, we introduce a new graph construction strategy for neighbor node selection in GraphSAGE, utilizing the original time series. GraphSAGE is an embedding model that leverages information from neighboring nodes, facilitating feature embedding by considering both the target node and its neighbors. Experimental comparisons using UCR data demonstrate that the proposed SbCM achieves up to 2.5 times better classification performance than SimTSC in large-scale data and multiple class scenarios.
LLM-based Conversational Recommender Systems Using User/Item Preference Reasoning Paths
Hyeri Lee, Heejin Kook, Seongmin Park, Jongwuk Lee
http://doi.org/10.5626/JOK.2025.52.10.890
Conversational recommendation systems seek to understand user preferences through interaction, providing personalized item suggestions. Recent advancements in large language models (LLMs) have improved the ability to infer latent preferences from conversations; however, including superfluous information can result in unintended recommendations. This study introduces RNCRS (Reasoning paths and Neighbor enhanced CRS), a conversational recommendation framework that develops reasoning paths for LLMs and captures latent preference information beyond surface-level context. The proposed method enables robust recommendations without requiring additional training by (1) utilizing item reasoning paths to represent multifaceted item characteristics, (2) leveraging collaborative knowledge to reflect the collective preference patterns of neighboring users, and (3) complementing explicit preferences that may be overlooked in reasoning paths through direct content similarity between conversations and items. Experimental results demonstrate that the proposed method achieves up to 12.7% improvement in performance over existing models based on Recall@50.
Search
Journal of KIISE
- ISSN : 2383-630X(Print)
- ISSN : 2383-6296(Electronic)
- KCI Accredited Journal
Editorial Office
- Tel. +82-2-588-9240
- Fax. +82-2-521-1352
- E-mail. chwoo@kiise.or.kr