Earticle

We compared empirically the forecast accuracies of the LSTM model, and the GRU model, because two models were commonly used in deep learning time series forecast. LSTM compensates for the vanishing gradient problem of RNN and provides good performance, but takes a long time to calculate. GRU was proposed as a model that simplifies the structure of LSTM to reduce computation time. Data used in the model is 163 days. We compared the forecast results for 33 days. We collected the stock closing prices of the top 4 companies by market capitalization in Korea such as “Samsung Electronics”, and “LG Energy”, “SK Hynix”, “Samsung Bio”. The collection period is from January 2, 2024, to August 30, 2024. The paired t-test is used to compare the accuracy of forecasts by the two methods because conditions are same. The null hypothesis that the accuracy of the two methods for the four stock closing prices were the same were not rejected at the significance level of 5% except Samsung Electronics. However, in four cases, the averages of the GRU errors were lower than those of the LSTM errors. And we can find that GRU shows similar performance while requiring less computation than LSTM. Graphs and boxplots confirmed the results of the hypothesis tests. Because in other empirical studies the results may vary, many additional empirical studies are needed.

This paper presents a design of an OOK (On-Off Keying) envelope detector with one and a half octave bandwidth in UHF band, along with the analysis on the third-order intermodulation (IMD3) performance. A Schottky diode is employed for high-speed switching, while T and Pi networks are used for the wideband matching at UHF frequency range. To minimize ripples within the bandwidth, capacitor and resistor combination is accurately modeled and optimized at the output stage for an input resistance of 50Ω and the high output resistance. To evaluate the performance of the designed OOK receiver, mathematical analysis in the frequency domain is conducted to verify its modulation effect on the envelope signal throughout the conversion. In order to further assess the performance of the diode envelope detector, two-tone signal is applied to analyze the modulation bandwidth and IMD3 of the envelope signal through nonlinear diode and low-pass filter (LPF). As a result, the bandwidth of the detector shows 340MHz ~ 980MHz, and its IMD3 is - 44.39 dBc which is better than the typical value of -40dBc for RFID communication systems.

A Markov-modulated Bernoulli process (MMBP) is a mathematical framework used to describe the state of a system over time, where a binary state (either success or failure) is determined by the environment of the system, which evolves randomly over time according to a discrete-time Markov chain. In this paper, we present an MMBP-based mathematical modeling and analysis for evaluating the performance of the slotted Aloha random access protocol from both terminal and network perspectives. In particular, we investigate channel usage patterns in the presence of fading and present a unified analytic approach that benefits from the MMBPbased modeling. Through numerical studies, we verify the efficacy of our analysis.

To cater to the growing usage of e-commerce, understanding how consumers respond to the core features of websites is needed. Personalized recommendation is a crucial element of online shopping that differentiates online shopping from offline experience. Two experiments examined the effects of interplay between how the recommendations are made and the type of product consumers shop for. Study 1 revealed that when the source of recommendations is explained, the recommendations are perceived as more credible. Study 2 showed that the effect of personalized recommendations is not uniform; participants have a more positive website attitude for identity-signaling than non-identity-signaling products. However, the different sources of recommendations did not have a direct effect on website attitude.

This paper proposes a 4-pole tunable absorptive dual-notch filter. The absorptive responses at two notch frequencies are achieved using a coupled-line with a lumped inductor and a finite Q varactor. Additionally, the two notches can be independently controlled by adjusting three loading capacitances of the resonator. A 4-pole filter was designed and fabricated on a substrate with a height of 25 mils. The filter demonstrated measured return losses of 5.5–10.2 dB and 8.1–20.2 dB at 0.68–0.96 GHz and 1.11–1.39 GHz, respectively. Applications include cognitive radio systems and carrier aggregation systems utilizing two transmission frequencies. In the future, RF MEMS switching capacitors are expected to significantly improve insertion loss, power handling, and linearity.

Recent advancements in artificial intelligence (AI) technology are creating new opportunities for evolving brand design methodologies. AI possesses the ability to analyze intricate data and propose innovative solutions that may be overlooked by human designers. In this light, this study seeks to investigate the development of brand design concepts in tandem with AI advancements and explore the potential of integrating Generative AI into brand design through practical workshop case studies. The researchers organized a rebranding workshop for 'Goubuli (狗不理),' a renowned Chinese snack brand, involving students in the use of AI technology to generate design concepts. This study examined how AI can be incorporated into brand design processes and the changing role of designers. The key findings revealed that while AI tools excel at rapid concept generation and creative ideation, they require significant human oversight for cultural sensitivity and brand alignment. The findings revealed both the effectiveness and limitations of AI in brand design, highlighting specific methodologies for its application. This research contributes practical guidelines for integrating AI tools into brand design workflows and provides a framework for balancing AI capabilities with human expertise in commercial design projects. It was found that AI-generated images have inherent stylistic and structural limitations, underscoring the ongoing necessity for human designers to refine and enhance AI-generated content.

As cyberthreats continue to rise rapidly, there is an urgent need for high quality cybersecurity AI datasets. These datasets are essential in training advanced AI models that enhance cybersecurity measures. The construction of such datasets is often faced with data quality, diversity and ethical consideration issues. Moreover, current datasets suffer from bias, incompleteness, and real-world representations. Given the dynamic nature of emerging cyber threats, there is also need for real time updates that traditional methods often fail to avail. This results in outdated cybersecurity AI datasets. Another issue is the ethical handling of sensitive data where compliance with regulations such as General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPPA) are often overlooked, thus endangering ethical handling of data in cybersecurity AI systems. Thus, our paper proposes a conceptual framework to address the challenges by considering state of the art technologies such as edge computing, real time processing and machine learning for enhanced data collection, processing, labeling and feature extraction. We integrate in diverse data sources and other innovative methods making our framework achieve high quality datasets that are highly needed for enhanced AI model performance in cybersecurity AI applications. We also consider data privacy and compliance thus contributing to a achieving a more secure and resilient cyberspace.

Ransomware is one of the most significant cybersecurity threats facing the world. In this research we designed and constructed a custom cybersecurity AI dataset for ransomware detection. We then evaluated the dataset using different machine learning models. The dataset was constructed using Cuckoo Sandbox where raw ransomware samples were analyzed to extract key features such as API calls, DLL usage, file operations, network activity, process creation and registry changes. These were then carefully labeled as either ransomware or benign. For evaluation purposes, the custom cybersecurity AI dataset was utilized to train and test various machine learning models. The dataset was split into 80% for training and 20% for testing. Logistic Regression, Random Forest, K-Nearest Neighbors (KNN), and XGBoost models were used to evaluate the resulting custom Cybersecurity AI Dataset. We obtained higher results of accuracy, precision, recall, and F1 scores evaluation metrics. Moreover, our results demonstrate the robustness of a combination of well-designed custom Cybersecurity AI Datasets and machine learning techniques in enhancing ransomware detection mechanisms as well as providing a framework for future cybersecurity applications

In many stream ciphers, Linear Feedback Shift Registers (LFSRs) are a fundamental component. Even though they are fundamental, their nature of inherent linearity can be exploited by cybercriminals through cryptanalytic attacks. In this research, we explore the importance of non-linearity in LFSR-based stream ciphers. We then propose techniques for achieving enhanced security through incorporation of strong non-linear elements. Specifically, we explore the application of S-boxes, Permutation boxes, Full Adder, and other non-linear operations in combining function and address limitations of traditional approaches. Finally, security analysis of our design is performed. We analyze Period, Linear Complexity, Randomness and Correlation Immunity. The results obtained are compared with the National Institute of Standards and Technology (NIST) requirements. All the results passed the NIST tests, indicating that our proposed enhancements offer a robust solution to the fundamental weaknesses of traditional LFSRs thereby contributing to more secure cryptographic systems.

In this paper, we investigate the mediating role of artificial intelligence (AI)-based technologies in the relationship between entrepreneurial characteristics and career intentions, focusing on the entrepreneurial and employment intentions of undergraduate design students. Results indicate that innovation, autonomy, and proactiveness significantly enhance AI utilization skills, which are, in turn, strongly associated with employment intention. Innovation emerged as the most influential trait for entrepreneurial intention, with proactiveness and risk-taking playing smaller roles. Notably, AI skills had the strongest impact on employment intention, suggesting that technological proficiency enhances employability within design fields. Moreover, AI utilization skills acted as a mediator between entrepreneurial traits and employment intention, underscoring the role of AI in bridging entrepreneurial attributes with career outcomes. These findings highlight the importance of integrating AI skill development into educational programs for students. Supporting both entrepreneurial and AI competencies prepares students for career flexibility, whether in entrepreneurship or traditional employment, and enhances their competitiveness in creative and tech-driven industries.

Submarine security systems are pivotal in modern military strategies and are critical for ensuring the operational security of military forces. This article examines submarine security technologies' historical evolution and assesses their adequacy in the face of contemporary threats. The study comprehensively addresses traditional security measures such as acoustic and magnetic signature reduction technologies, countermeasure systems, cybersecurity strategies, and energy security issues. In particular, it emphasizes the impact of anechoic coatings and degaussing methods used to complicate submarines detection. It also notes that autonomous underwater vehicles and AI-supported systems pose new threats to submarine security. In this context, an analysis of the strengths and weaknesses of current security systems is provided, along with recommendations for future technological adaptation strategies. The proposed strategy for strengthening submarine security encompasses the integration of technological innovations and measures in the field of cybersecurity. Ultimately, this study highlights the importance of developing submarine security systems and their adaptation to modern threats, providing a guiding framework for future research.

In this paper, we explore the third-person effect, focusing on people’s perception of the negative impacts of ChatGPT and their intention to support its regulation. A survey was conducted to verify third-person effect in the context of generative AI’s positive and negative sides. The survey results revealed that respondents tend to rate ChatGPT’s negative effects on other users and on society in general higher than those on themselves, confirming the presence of third-person as well as the social distance effect. It was revealed that positive and negative evaluations about ChatGPT’s functional features can influence the magnitude of the third-person effect and the social distance effect. This finding is in parallel with the result that respondents who were largely using ChatGPT for functional purposes than hedonic purposes were showing huger social distance effect, while those without clear use purpose was more prone to perceptive bias regarding negative influence of ChatGPT over other people and the society. Lastly, it was also revealed that the magnitude of social distance effect was influencing the user’s intention to support any regulation on ChatGPT usage in response to any potential harmful outcomes of using the media.

Recently, due to the occurrence of safety accidents in areas with high pedestrian density, local governments have been focusing on introducing systems that can prevent such accidents in advance. Specifically, they have plans to prevent these accidents by predicting pedestrian density in real-time beforehand. There are various methods to measure human density, but using CCTV to measure density has been gaining attention. This is because the infrastructure for CCTV is already well established, eliminating the need to build additional hardware infrastructure. In other words, only software enhancements are needed on the backend. Many algorithms for measuring pedestrian density have been developed, and recently, deep learning-based methods have been particularly prominent. However, most deep learning-based density measurement methods either count the number of pedestrians in the entire input video frame or predict their locations. In practice, though, even if the footage is obtained from the same CCTV camera, it is necessary to measure the density for specific areas within the video separately. This is because the distance from the camera differs for each region within the video, leading to potential discrepancies between the visible density in the video and the actual density. Therefore, this paper proposes a density measurement system that compensates for these variations in distance from the camera across different areas.

The advancement of generative AI technologies has significantly impacted various domains in software engineering, particularly in automating test case generation. As software systems become increasingly complex, manual test case creation faces limitations in terms of efficiency and coverage. This study analyzes the capabilities and limitations of major generative AI models—ChatGPT, Copilot, and Gemini—in generating software test cases. We focus on evaluating their performance in boundary value analysis, exception handling, and property-based testing. Using the ArrayUtils.indexOf() function from the Apache Commons Lang library as the test subject, we conducted experiments to compare the quality and effectiveness of the test cases generated by each model. Our findings indicate that while generative AI can efficiently produce a substantial number of high-quality test cases, there are instances of incorrect test cases and test codes. To address these issues, we propose guidelines for developers to enhance the reliability and consistency of test case generation using generative AI. Future research will explore the application of these models to more complex software systems and further methods to improve their test generation capabilities.

The trend of Small Large Language Models (LLMs) has been developing significantly in recent years. Lightweight LLMs are designed to operate efficiently on mobile devices or edge computing environments, and perform well even with limited resources. These models are optimized for specific domains and provide results that meet the needs of specific industries. In addition, they are easily accessible to non-developers due to their user-friendly interfaces, and are utilized in various fields. The purpose of this paper is to analyze the performance, functionality, and usability of Small Large Language Models (LLMs) to understand how they can be effectively used in various natural language processing (NLP) tasks. In particular, the key goal is to evaluate what advantages and disadvantages small models have compared to large models, and whether they can be optimized for specific tasks. Through this analysis, we aim to provide useful insights for developers and researchers in selecting and utilizing LLMs.

Generative AI is being used in software development for automated code generation, bug detection and optimization, natural language-to-code transformation, and automated documentation. Tools like GitHub Copilot and OpenAI Codex provide real-time code suggestions, boosting productivity and enabling nondevelopers to create code from simple descriptions. The purpose of this paper is to analyze various generative AI frameworks employed in software development to investigate their specific characteristics, advantages, and disadvantages. By conducting this analysis, the study aims to furnish software developers with essential data that could streamline their use of generative AI, thereby enhancing productivity and facilitating better decision-making regarding which tools to adopt for their projects. Understanding these attributes will enable developers to select frameworks that best align with their needs and technical environments, optimizing their workflow and outcomes.

The recent surge in e-commerce has led to an increase in demand for parcel logistics, contributing to social costs such as traffic congestion and environmental issues. Various methods have been researched and developed to address these problems, and as part of this effort, research and development on community logistics parcel terminal systems is currently underway. This study aims to analyze the interrelated factors among operational systems from an operational perspective before the actual establishment of a community logistics parcel terminal system. To achieve this, we identified the functions and requirements of the operational systems within the community logistics parcel terminal and analyzed the interconnection elements based on this foundation, presenting an architecture. The results of this study provide foundational data for improving the operations of the community logistics parcel terminal system and are expected to contribute to future research and practice in the logistics field.

Large language models (LLMs) have emerged as powerful tools in the field of natural language processing (NLP) and have recently attracted considerable attention in the field of recommendation systems (RSs). In this regard, we investigated a method to simultaneously improve the accuracy of real-time recommendations and user satisfaction by combining LLMs and session-based recommendation systems. We propose the LReLLM4SBR model, which combines lightweight LLMs and reflective reinforcement learning to improve the performance of session-based recommendation systems. Through experiments on MovieLens and Amazon review datasets, LReLLM4SBR showed improved performance compared to existing models in Precision@K, Recall@K, MAP@K, and NDCG@K indices. This study suggests that combining lightweight LLM-based models and reinforcement learning techniques can improve the performance of session-based recommendation systems, and suggests the possibility of contributing to improving real-time personalized services of recommendation systems.

We explored motor control impairments in older adults with mild cognitive impairment (MCI) using virtual reality environment. In particular, we aimed to investigate the specific impact of MCI on both feedforward and feedback motor control functions, including visual and proprioceptive feedback mechanisms. To do that, we specifically designed an experimental setup within an immersive virtual reality environment to isolate the effects of aging-induced MCI on motor control. We compared healthy older adults to adults with MCI who were instructed to perform a reaching task in an immersive virtual reality environment under two conditions: (1) using a visible cursor and (2) using an invisible cursor. We demonstrated that while peak speed errors were not significantly different between the groups, the MCI group showed a marked difference between visible and invisible cursor conditions. Furthermore, movement times in the MCI group increased exponentially with displacement, whereas the healthy group maintained consistent times across different displacements. These findings highlight proprioception as a key factor contributing to the motor control impairments associated with MCI, offering new insights into the interaction between cognitive decline and motor functions.

A rotation angle estimation method tailored for low-spec systems, such as embedded systems, was proposed, which is essential for image alignment in the initial stages of image analysis and recognition. While estimating the rotation angle using the phase of complex moments defined in polar coordinates can achieve high accuracy, it typically demands significant memory and longer processing times. Therefore, the proposed method narrows the search range for angle estimation using Angular-Radial Transform (ART) coefficients and simplifies the computation by converting rotation angle estimation using only the lowest frequency coefficient into a vector sum of projection profiles in polar coordinates. The memory usage, execution time, and estimation error of the proposed methods were also controlled to provide recommendations for selecting the most suitable method based on the performance and requirements of the target system.

While conventional LLMs provide uniform services to various users, Personal LLMs can improve user experience by optimizing for the needs and environments of specific users. The purpose of this paper is to identify how the personalized features of Personal LLM can improve user satisfaction and efficiency, and the challenges associated with its application. The results of the study show that Personal LLM significantly improves work efficiency by providing customized responses and reflecting the specific needs of users. In addition, LLM showed progressively better performance over time through learning, and it was confirmed that it can be gradually improved through interaction with users. However, it was confirmed that there are technical and ethical limitations, such as data privacy issues, which remain important challenges in commercializing Personal LLM. This paper suggests the possibility that Personal LLM can provide customized services to users and provides important basic data for the development of personalized AI systems in the future.

This study systematically analyzes the impact of face angle and lighting changes on eye state recognition technology and compares the performance of three technologies: CNN, MediaPipe, and Dlib. Specifically, the CNN-based approach utilizes a transfer learning model, Inception, to assess eye state recognition accuracy. With recent advancements in AI and computer vision technology, eye state recognition has become crucial in applications like driver drowsiness detection, user authentication, and medical monitoring. However, the performance of these technologies is greatly influenced by face angle and lighting conditions. This research evaluates the recognition accuracy of the three technologies under various face angles and lighting conditions, finding that CNN demonstrates robust performance against both lighting and angle variations. This study aims to provide fundamental data to improve the reliability of eye state recognition technology and to suggest future research directions.

This study constructed a real-time image processing and object recognition system using MobileNet and evaluated its performance on custom robots. The lightweight architecture of MobileNet enabled high performance in resource-constrained environments, demonstrating its capability for real-time object recognition. The findings validate that a MobileNet-based system is well-suited for optimizing the performance of custom robots, especially in low-power settings. This research highlights the potential of efficient real-time image processing and object recognition implementation to enhance the functionality of robotic systems.

This study aims to design a system that can further improve two-way communication based on user activity data of an online Learning Management System (LMS). Most online education systems used in domestic educational institutions have the limitation of simply providing lecture videos. If LMS system only have a simple information provision function without improving communication, the educational effect of online education cannot be expected. To solve the limitations of this LMS, this study seeks to find a system design methodology that systematically activates mutual communication in online education. It has research value in that it overcomes the functional limitations of existing studies and current LMS solutions through platform design based on user action scenarios for students, professors, and system operators. In this study, we aim to improve this limitation from the engineering aspect of information system design.

ZNS (Zoned Namespaces) SSD is an emerging type of SSD that divides its internal space into zones, which are logical units with a sequential write constraint. By allowing the host to manage these zones directly, it enables more efficient data placement and reduces garbage collection overhead. F2FS (Flash-Friendly File System) is a log-structured file system optimized for NAND flash memory and is one of the default Linux file systems that support ZNS SSDs. Currently, ZNS-based F2FS allocates zones based on file extensions, sizes, and other characteristics. However, it does not support zone allocation policies that consider the level-based lifetime of SST (Static Sorted Table) files, which are the primary data structure of LSM (Log Structured Merge) tree-based key-value stores like RocksDB. Therefore, this paper proposes a technique to reduce the garbage collection overhead of ZNS SSDs by employing a new zone allocation policy that considers the levels of the LSM-tree.

This research developed a Zigbee mesh network topology to assess the data transmission efficiency utilizing the Ad-Hoc On-Demand Distance Vector (AODV) routing protocol. AODV is Zigbee's low-power, low-datarate wireless ad hoc network routing protocol. A physical mesh network was created using Zigbee components, namely the Zigbee Coordinator (ZC), Zigbee Router (ZR), and Zigbee End Device (ZED). Data transmission experiments were carried out among the Zigbee modules with the AODV routing protocol, and evaluated over four distinct versions. Modifications were made to key aspects of the AODV protocol, including mesh selfhealing and link cost management. The experimental results revealed that during the self-healing phase, all modules in the network were engaged, resulting in a transmission time of 0.4 seconds. After the direction was established, data was sent in only 0.01 seconds, achieving a speed that was 40 times faster than the previous experiment.

In the past two years, generative AI painting has not only been brilliant in the field of design, but also has been widely used in other fields. However, just generating a single AI photo is no longer enough to meet the growing needs of design. With the continuous development of AI video generation technology, design teams have more AI tools to choose from. In this paper, through the research of vid2vid style transfer technology, the basic 3D animation video can be quickly converted into a variety of animation video style, the analysis shows that the technology can be better applied in the field of 3D animation, and has a wide range of development prospects and application potential. Although some problems were found in the testing phase, they have been solved, and by exploring the application of new technologies, we can better promote the creative expression of 3D animation and improve the productivity of designers.

The study investigates gender differences in the direct and indirect effects of media exposure and interpersonal communication on HPV vaccination intention, mediated by social norms. An online survey of college students was conducted. Data analyses revealed that for male students, interpersonal communication directly influenced their HPV vaccination intention, while media exposure indirectly influenced it through injunctive norms. For female students, interpersonal communication influenced vaccination intention both directly and indirectly through injunctive norms. These findings emphasize the importance of injunctive norms in HPV vaccination contexts and offer insights for developing evidence-based strategies to boost HPV vaccination rates among both men and women.

Generative AI technology is an innovation that saves time and cost and promotes creativity in the process of discovering ideas for advertising production, creating storyboards, and producing advertising videos and voices through various technologies such as text-image generation, image-image conversion, and voice synthesis. established as a tool. It provides differentiated advantages and synergies compared to existing advertising production methods, and shows the potential to accelerate innovation in the advertising industry by giving advertising creators more options and flexibility. The use of these technologies does not simply increase efficiency, but serves as an important turning point that requires changes in creative methods, suggesting the possibility of transitioning to a future-oriented advertising production model that combines creative ideas and technical efficiency.

In response to growing global concerns over environmental sustainability, electric vehicles (EVs) have emerged as a crucial technology in reducing greenhouse gas emissions, reflecting the broader transition towards ecofriendly transportation. This study investigates the influence of brand personality and self-image congruence on the purchase intentions of Chinese consumers within the rapidly expanding EV market, with a focus on two significant brands: Tesla, a global leader, and Xiaomi, an emerging Chinese contender. By analyzing how congruence between a consumer's self-image—both actual and ideal—and the perceived brand personality influences brand attachment and purchase intentions, this research uncovers critical factors driving consumer behavior in the context of green technology adoption. The moderating roles of patriotism and price sensitivity are also explored, revealing their significant impact on brand attachment and purchase intentions for both Tesla and Xiaomi. These insights contribute to the understanding of consumer behavior in green markets and provide actionable strategies for promoting sustainable vehicle adoption.