Latest Issue
Vol. 53,
No. 1,
Jan.
2026
All Issues
Monitoring and Controlling Internal Container Activity Using LSM + eBPF in a Multi-Container Environment
Yejune Ko, Hyeonseok Kim, Mingyu Jeong, Changhyun Lee, Harksu Lim, Sunghyun Jeon
http://doi.org/10.5626/JOK.2026.53.1.1
This paper explores real-time monitoring and control techniques utilizing an eBPF (extended Berkeley Packet Filter) and the LSM (Linux Security Module) in multi-container environments and Kubernetes-based orchestration systems. Traditional security methods struggle to maintain consistent policies due to the dynamic nature of container creation and termination, limiting fine-grained control at the individual container level. In this study, we employ eBPF to monitor system calls, network activities, and file accesses at the kernel level, while also implementing mechanisms to restrict specific container behaviors. Furthermore, we assess the feasibility of applying consistent security policies in Kubernetes environments, experimentally validating policy management and monitoring techniques at the namespace, pod, and label levels. Our experimental results indicate that eBPF-based monitoring and control functions efficiently in multi-container environments with minimal performance overhead, allowing for flexible and scalable security policy enforcement in orchestration systems like Kubernetes. This research advances the development of cloud-native security solutions that leverage utilizing eBPF.
AI Tax: Performance Analysis of AI Inference Serving
http://doi.org/10.5626/JOK.2026.53.1.8
With the rapid advancements in artificial intelligence (AI), smart applications powered by compute- and memory-intensive AI models now make up a significant portion of modern datacenter workloads. To meet the growing demands of AI workloads, specialized accelerators are increasingly deployed in datacenters to enhance AI inference efficiency. However, most previous studies on AI inference acceleration have focused primarily on the performance of neural network computations in isolation. In addition to these computations, an AI inference server typically handles other essential infrastructure tasks, such as web serving to send and receive inference requests and responses, as well as application-specific pre- and post-processing. In this paper, we refer to these additional operations as the AI Tax. We analyze the AI Tax in a representative modern AI inference server that runs various image classification models using Nvidia's industry-standard AI serving software stack. Our findings reveal that the AI Tax can lead to up to 55% degradation in end-to-end server performance compared to standalone neural network compute and consumes an average of 25 CPU cores.
Zero-shot Referring Image Segmentation using Referring Expression Augmentation and Mask Aggregation
Seungheon Song, Sungsik Kim, Junghyeon Seo, Jaekoo Lee
http://doi.org/10.5626/JOK.2026.53.1.15
With advancements in computer vision technology, image segmentation tasks are increasingly utilized across various fields. Among these, reference image segmentation is particularly valuable for achieving precise regional segmentation based on user instructions. In this paper, we introduce a reference image segmentation framework inspired by human cognitive processes, which utilizes prior knowledge to recognize objects. Our method employs a large language model to infer various visual attributes of objects and integrates these inferences with mask proposals. Consequently, our approach enhances oIoU and mIoU performance by 0.41% and 0.74%, respectively, compared to existing methods on the RefCOCO, RefCOCO+, and RefCOCOg benchmark datasets.
LLM-based Extractive and Abstractive Korean Meeting Summarization with Topic Decomposition
Dongwon Noh, Donghyeok Koh, Hyun Kim, D asol Kim, T aehoon Kim, Cheoneum Park
http://doi.org/10.5626/JOK.2026.53.1.22
We propose a framework for summarizing meeting content based on specific topics using a large language model that combines extraction- and generation techniques. Each topic sencapsulates the central concept of the meeting, but its concise nature can complicate direct application during model inference. To overcome this challenge, we utilize a Chain-of-Thought (CoT) decomposition technique to interpret the topic and guide the extraction- and generation of summaries. Additionally, we employ an encoder with a long-context retriever to select meaningful sentences from extensive meeting content for topic summarization. Experimental results show that our extraction-generation framework achieves a ROUGE-1 score of 43.65, demonstrating its effectiveness in producing meeting summaries aligned with the specified topics.
Training Liquid State Machine using Reward-Modulated Spike-Timing-Dependent Plasticity
Youngseok Joo, Minsu Lee, Byoung-Tak Zhang
http://doi.org/10.5626/JOK.2026.53.1.32
The Liquid State Machine (LSM) is a recurrent spiking neural network model rooted in computational neuroscience, characterized by its rich temporal dynamics, low training complexity, and biological plausibility. While traditional LSMs typically use fixed reservoir weights, which limits their adaptability, incorporating spike-timing-dependent plasticity (STDP) into the reservoir has been shown to improve performance. Additionally, conventional LSMs employ external classifiers, such as linear regression or gradient-based methods, to train the output layer, which is not suitable for online, real-time learning. In this paper, we introduce a reinforcement learning framework for training an LSM using dopamine-modulated spike-timing-dependent plasticity (DA-STDP). Our system enables a biologically inspired, reward-based learning mechanism that adjusts synaptic weights based on feedback signals. We validate our approach under various training conditions using the MNIST dataset, demonstrating the applicability of DA-STDP as a training mechanism within the LSM framework.
Supporting Novice Researchers to Write Literature Review using Language Models
Kiroong Choe, Seokhyeon Park, Seokweon Jung, Hyeok Kim, Ji Won Yang, Hwajung Hong, Jinwook Seo
http://doi.org/10.5626/JOK.2026.53.1.40
Literature reviews are vital components of academic research, showcasing the unique perspectives of researchers. Although there is a growing number of large language model (LLM)-based services designed to help users understand and utilize academic papers, their effectiveness in aiding novice researchers to cultivate independent viewpoints on literature remains largely unexamined. This study employs semi-structured interviews to identify the barriers novice researchers encounter before, during, and after the writing process. It also proposes a prototype system based on LLM technology that systematically supports the literature review writing process. A series of workshop studies revealed that novice researchers could seamlessly transition into the writing phase, collaboratively generate satisfactory content, and build agency and confidence through a long-term, dynamic partnership with AI.
Integrating Self-Verification with External Knowledge Retrieval for Reducing Hallucination in Large Language Models
http://doi.org/10.5626/JOK.2026.53.1.58
Large Language Models (LLMs) have transformed the field of natural language processing. However, they face a significant challenge known as hallucination—the generation of text that appears plausible but is factually incorrect. Previous methods aimed at reducing hallucinations often struggled with limitations, such as relying solely on the LLMs’ inherent knowledge, which can lead to insufficient information, and failing to fully utilize the reasoning capabilities of LLMs. To overcome these challenges, we introduce CoVe-RAG, a novel technique that combines Chain-of-Verification (CoVe) with Retrieval-Augmented Generation (RAG). CoVe-RAG improves the factual accuracy of LLM outputs by integrating external knowledge retrieval into a self-verification process. Our experiments demonstrate that CoVe-RAG significantly outperforms both CoVe and RAG in reducing hallucinations. Our main contributions include: (1) identifying the limitation of CoVe in mitigating hallucinations, (2) presenting CoVe-RAG as a more effective solution, and (3) providing empirical evidence of the synergistic benefits of combining self-verification with external knowledge retrieval. Based on these findings, we highlight the complementary relationship between self-verification and external information in enhancing the reliability of LLMs.
UWB-Based Multi-Drone Collaboration System for Target Localization
TaeHyeon Jo, YeoGyeom Kim, Su-Hwan Yun, SungTae Moon
http://doi.org/10.5626/JOK.2026.53.1.65
Drones are increasingly used in search and rescue operations across large areas. However, traditional methods encounter two main challenges: the inaccuracy of vision and GNSS sensors, and the reliance of Ultra-Wideband (UWB) technology on fixed anchors. To address these issues, this paper introduces a novel collaborative search system in which multiple drones, accurately localizing themselves with RTK-GPS, act as mobile UWB anchors. In this system, the drone swarm initially conducts a distributed search over the target area. When a UWB tag signal is detected, the swarm reorganizes into a formation optimized for ranging using a potential field algorithm, and then localizes the target's position with a particle filter algorithm. Implementation on a PX4-ROS2-based hardware platform, along with outdoor flight experiments, demonstrates the feasibility of achieving centimeter-level positioning accuracy in expansive, infrastructure-less environments. These findings confirm the practical effectiveness of our system and underscore its potential to revolutionize autonomous search operations.
Hierarchical Object Detection Method for Automated Object-based Place Recognition
Won-Seok Choi, Byoung-Tak Zhang
http://doi.org/10.5626/JOK.2026.53.1.74
This study proposes a hierarchical object detection method (HOD-SAM) designed to effectively eliminate geometric and semantic duplications among masks produced by the Segment Anything Model (SAM), while organizing them into a hierarchical structure. The method constructs a tree structure based on the inclusion relationships between object masks and utilizes self-supervised feature extractors to remove semantically duplicated masks. This approach not only addresses duplication issues in SAM-based instance detection but also enhances downstream task performance. We validate the effectiveness of HOD-SAM in multiple tasks such as object instance segmentation and place classification within the context of place recognition. Experiments conducted on the COCO and Places365 datasets reveal that the proposed hierarchical recognition structure outperforms the original SAM model in terms of both accuracy and efficiency, indicating its potential as a general-purpose perception module for broader applications in understanding complex environments.
A Novel Evaluation Method and Learning Approach for Identifying and Addressing Interaction Type Recognition Issues in Drug-Drug Interactions Prediction
Youngbin Cho, Dasom Noh, Gyoung Jin Park, Minji Seo, Sunyoung Kwon
http://doi.org/10.5626/JOK.2026.53.1.82
Drug-drug interaction prediction is a task aims to identify interactions between two drugs to prevent potential side effects from polypharmacy. Previous studies have employed a binary classification approach, where drug pairs and their interaction types are provided to the model to determine whether a specific interaction occurs. However, this method has limitation: the model often struggles to learn interaction types adequately, and the standard evaluation method does not highlight this issue. In this paper, we introduce a new assessment called the "Interaction Type Recognition Test" to evaluate the model's ability to identify interaction types. Additionally, we propose a learning method that incorporates negative pairs (interaction changes) to enhance model’s ability to learn these types effectively. Experiments conducted on datasets with varying structural characteristics, specifically DrugBank and Twosides, demonstrate that our proposed method significantly improves interaction type recognition performance in both datasets, validating the effectiveness of our approach in learning interaction types.
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