Earticle

Phishing attacks have surged with the expansion of the internet, becoming a major cause of personal information leaks. A significant challenge in phishing detection lies in the shared characteristics between phishing and benign webpages, as attackers intentionally design phishing pages to closely resemble benign ones. These shared features often result in ambiguous representations in the embedding space, which complicates accurate detection. To address this issue, we propose a method that introduces feature disentanglement into deep learning models for phishing webpage detection. By leveraging both URL and HTML data, our method employs a triplet loss function to better separate phishing and benign classes in the embedding space, thus reducing the overlap of shared features. This disentanglement effectively decreases the rate of false positives and false negatives. Experimental results show that our approach improves phishing detection accuracy by up to 9.85 percentage points and increases the F1 score by up to 11.92 percentage points compared to existing methods.

This study aims to enhance the efficacy of anomaly detection techniques through the application of autoencoders. Autoencoders, neural network models that compress and reconstruct input data by learning patterns from normal instances, typically struggle with reconstructing anomalous data. To address this limitation, we propose integrating Mahalanobis Distance, a method for measuring the distance between a data point and the distribution center, into the autoencoder's latent space. Our approach diverges from conventional methods by treating reconstruction error as a vector rather than a scalar value, allowing for more granular outlier information. We evaluate the model's performance across multiple metrics, including accuracy, precision, recall, F1 score, and ROC-AUC, utilizing five different scaling techniques. Experimental results indicate that RobustScaler offers superior performance due to its resilience to outliers, ensuring consistent results across varied data distributions. This research contributes to the advancement of anomaly detection methodologies, potentially enhancing their applicability in realworld scenarios.

In this paper, we investigate the use of Generative AI to enhance and augment datasets within the context of smart factory metaverse platforms. Specifically, we propose a method for generating synthetic abnormal data using Generative Adversarial Networks (GANs) to address the inherent data imbalance issues in PCB (Printed Circuit Board) datasets, where normal data far exceeds abnormal samples. In our study, we demonstrate that generating synthetic data from a minimal set of abnormal samples significantly improves the performance of AI models, such as MobileNet-V3 Large. By augmenting the abnormal dataset from 20 to 500 images, the classification accuracy of the model increased from 74.9% to 97.3%, validating the effectiveness of this approach. This research highlights the potential of combining generative AI with metaverse platforms, enabling real-time guidance and training for users without the limitations of time and space, thus enhancing production efficiency and minimizing human error in smart factory environments.

Demonstration of Data Security has a significant role performed by blockchain technology. It has shown a crucial role in securing domains including finance, digital identity and supply chain. Scalability in blockchain remains a major challenge that limits the capability of adoption and performance. Existing research focuses more on external security of digital entity but may failed to enhance internal capabilities of blockchain technology. The proposed work explores an innovative module of blockchain technology that targets the scalability issue. The proposed approach introduces a hybrid framework includes the integration of cross-chain interoperability and Layer 2 solution to improve scalability of blockchain technology. The proposed work aims to mitigate current limitation in framework and provide novel solution to scalability issues. The cross-chain is a blockchain protocol which ensure the continuous interaction of information among the multiple blocks in the framework. The layer 2 solution offers the transaction processing facility to offload from the main chain to secondary protocol of the framework. The research presented the evidence of the effectiveness of the proposed solution that addresses the scalability concerns and offer an efficient blockchain ecosystem. The proposed approach avoids basic limitation of throughput and network congestion.

Since there is an invasion of privacy of elderly people's falls through cameras, we propose a fall detection system using millimeter wave (mmWave) sensors. To detect falls of the elderly, the system uses mmWave sensor data preprocessed with histogram of oriented gradient (HOG) followed by training and inference with a long short-term memory (LSTM) model. The fall detection rate tested using this system is very high at 95.4%.

In this paper, we investigate the application of Nash equilibrium strategy to enhance coordination and decision-making in multi-agent systems, specifically focusing on Automated Guided Vehicles (AGVs) systems. Traditional reinforcement learning methods often face challenges in multiagent environments due to the non-stationarity introduced by multiple learning agents and the complexity of coordinating actions among them. To address these challenges, we propose an approach that integrates game-theoretic principles of Nash equilibrium into existing multi-agent reinforcement learning frameworks. By incorporating Nash equilibrium considerations into the policy update mechanisms, agents can anticipate and respond to the strategies of other agents proactively. This integration reduces conflicts and improves cooperation without relying solely on reward shaping or penalization for undesirable behaviors, such as collisions. Additionally, we introduce a collaboration cost into the reward function to further incentivize cooperative behavior among agents. We validate the effectiveness of our approach in a flexible manufacturing system simulated using PyBullet, utilizing the default URDF models to create a realistic and standardized environment. Multiple AGVs operate as autonomous agents tasked with collaboratively optimizing production tasks. Experimental results demonstrate that our Nash equilibrium-based method significantly outperforms traditional algorithms—including MADDPG, NDQN, CQL, COMA, IQL, PPO, SAC, and DQN—in terms of cumulative reward, policy convergence speed, and overall system throughput.

This paper presents the design and implementation of a web-based system for real-time monitoring of underground pipeline conditions. The system utilizes sensor data from smart pipes, which are embedded with optical, potential, and ultrasonic sensors to detect anomalies in the pipeline infrastructure. By integrating sensor data with geographical mapping, the system provides comprehensive insights into the condition of underground pipelines, enabling effective maintenance and preventive actions. Key features include data visualization in the form of charts and a map-based interface that highlights the location of sensor-equipped pipelines.

This study investigates the use of machine learning techniques to estimate the value of studio apartments (Officetel), which are increasingly important as combined office and residential spaces for single-person households and freelancers. It aims to identify key variables affecting studio apartment prices and preprocess them for accurate predictions. Transaction data from studio apartments are used to compare the predictive performance of four methods: Multiple Regression Analysis, Random Forest, XG Boosting, and Deep Learning. The study seeks to determine the best-performing models for price estimation and aims to develop predictive models applicable to various real estate types and regions.

Hand hygiene is crucial for disease prevention, yet maintaining effective handwahsing habits remains challenging. Within this context, in this paper we build posture recongintion AI model able to automatically analyze, in real-time, the sequence of images acquired by a camera. We adopt YOLOv8 variants to classify the movement of the worker according to the six gestues defined by the World Health Organization and to evaluate the quality of the handwashing procedure. To test the our model, we use handwashing data provided by RoboFlow and an additional dataset builty by video sequences we directly capture. The results achieved on this dataset confirm the our model's effeciveness.

This study presents an innovative approach to bidirectional sign language translation using an AI-based 3D interactive Unity character. We have developed a system that receives Korean text input, translates it into Korean Sign Language(KSL), and then uses Unity3D to enable a 3D character to perform sign language gestures in real-time. The system aims to deliver natural character movements while accurately reflecting the vocabulary and grammatical structure of KSL. Although the representation of non-manual signals(NMS) remains a significant challenge, the combination of Unity3D characters with AI-based translation plays a crucial role in addressing the issue of information alienation among KSL users.

Following the COVID-19 pandemic, the use of kiosks surged, posing significant challenges for senior users. This study aims to improve kiosk user experience (UX) design for senior users by introducing a voice-based adaptive user interface (UI) alongside traditional touch input. This dualinput design intends to provide a more familiar and less visually cluttered experience for seniors. By incorporating personalized UI elements and a recommendation system, the process is streamlined, offering an experience akin to face-toface services. The results of this study aim to enhance kiosk accessibility and user experience for seniors. Future work includes the development of a prototype and conducting usability testing with senior user groups to validate the effectiveness of the proposed design.

The Metaverse is a socially influential platform that offers XR (Extended Reality) users access to multitudinous applications such as a website, a web application, or a desktop program. The handheld controllers, gestures, or voice commands are typically used by the XR-Users to navigate and access these applications. Considering voice commands are the fastest, they are used the most these days. We undertook an attempt to evaluate the security of voice commands for navigating browsers in the Metaverse to access websites. We studied the network analysis trends in depth for virtual environments. We observed URL (Uniform Resource Locator) whitelisting when performing tests on the Oculus Quest 2 voice assistant. We conducted partial network analysis on the network traffic collected during the URL voice command request processing using the OVR (Oculus Virtual Reality) device. With this partial network analysis of the network traffic, we propose a threat model for the OVR voice assistant for processing URLs. We convey, whether voice commands are secure even though they utilize URL whitelisting as a disguise through our research.

This paper presents a method to optimize 3D Gaussian Splatting, improving computational efficiency for real-time applications on edge devices like Jetson AGX Xavier. By applying pruning and quantization techniques, we enhance computing speed with minimal degradation on image quality. Our method enables efficient 3D scene reconstruction on resource-limited devices, making it suitable for AR/VR and autonomous driving.

This study proposes a smart camera system for detecting fires in Building-Integrated Photovoltaic (BIPV) systems. Traditional fire detection methods, such as smoke detectors, temperature sensors, thermal cameras, and videobased surveillance systems, each have specific advantages and disadvantages depending on the environment. However, due to the unique characteristics of BIPV systems, a more sophisticated and integrated detection system is required. In this study, we developed a smart camera system using an NVIDIA Jetson Orin Nano board, a YOLOv5 model, and a Lepton thermal camera to monitor and provide early warnings of fires in BIPV systems in real time. Experimental results across various environments demonstrated that the proposed system offers higher accuracy and reliability compared to existing methods.

The Internet of Things (IoT) has rapidly expanded across various sectors such healthcare, home, management, and agriculture, introducing significant security challenges due to the diverse and often insecure nature of IoT devices. An interesting solution is the specification-based intrusion detection system (IDS), which uses behavior rules to determine if the observed execution of a system conform with its intended operation modes and provide advisory alerts when deviations occur that indicate security breaches. This paper reviews the fundamentals of specification-based IDS, including the key design workflow and challenges in implementing these ideas, especially towards IoT ecosystem. It also summarizes existing use cases, introduce ideas to address identified challenges, and suggests future directions to enhance the effectiveness and efficiency of specification-based IDS in complex IoT systems.

Low-Rate Denial of Service (LDoS) attacks raise an increasingly frequent and significant threat to performancecritical and sensitive networks. Due to their slowly evolving nature, it is challenging –but crucial– to detect such attacks during their early phases in order to mitigate their impact on network performance, e.g. Quality of Service (QoS), in longterm operation. To this end, this paper investigates the prediction of QoS via a modified version of the Recurrent Trend Predictive Neural Network (rTPNN) and the use of the prediction towards detracting LDoS attacks. The presented rTPNN-based QoS predictor is evaluated and compared against benchmark models for five scenarios using an open-access dataset. The results have shown that the modified rTPNN model can predict QoS with under 2% SMAPE, and the QoS prediction is a promising approach for developing LDoS attack detectors in future works.

Today, the network becomes the core element in all that is done efficiently and effectively. They include block transfer, linear transfer, and asynchronous transfer. Optical Burst Switching (OBS) is also classified with them. By picking on data sent with OBS, some security failures occur, and these comprise Replay Attacks, Spoofing, and Burst Header Packet (BHP) flooding attacks which are among these threats. The addressed methodology incorporates the application of the Support Vector Machine (SVM) algorithm to fight down BHP attacks. The simulation outcomes reveal that the performance which is obtained from the actual learning algorithm is the best at 97.7% in all four classes of flooding attacks which include NB-No Block, NB-Wait, No Block, or Block. This proposed Intelligent Identification of BHP Flooding Attack on OBS utilizing Machine Learning Technique (I2BHPOBSML) shows that it is giving better results than the past Works.

This study compared the performance of the FCM(C-Means)-based ANFIS(Adaptive Neuro-Fuzzy Inference System) model and the CFCM(Context-based Fuzzy C-Means) clustering-based ANFIS model. The FCM-ANFIS model sets the initial Fuzzy Rule through FCM clustering and optimizes the rule through neural network learning. The CFCM-ANFIS model generates more sophisticated rules through CFCM clustering that considers the input and output variable space and learns the neural network. As a result of the experiment, the verification RMSE of the FCM-based ANFIS model was 3.5654 when the number of clusters was 6, and the RMSE of the CFCM clustering-based ANFIS model was 3.3954 in the parameters (P = 6, C = 2), which was higher than the FCM-based ANFIS model. It was confirmed that the CFCM method had better prediction performance than the FCM method, and this study proved that the CFCM-based ANFIS model was more effective in predicting body fat percentage.

While the gut microbiota is increasingly implicated in the pathogenesis of Parkinson's Disease (PD), the majority of existing research predominantly focuses on Western populations, with limited studies addressing Eastern cohorts. This study aims to elucidate the differential composition of gut microbiota between Eastern and Western PD patients by utilizing advanced machine learning techniques. 16S ribosomal RiboNucleic Acid (rRNA) sequencing data from stool samples are obtained from the Sequence Read Archive (SRA), comprising a random selection of 70 individuals (35 healthy controls and 35 PD patients) from four countries: Korea, Japan, the United States, and Italy. Recursive Feature Elimination (RFE) is employed for feature selection, and four machine learning models— Support Vector Machine (SVM), Random Forest (RF), k- Nearest Neighbors (KNN), and eXtreme Gradient Boosting (XGBoost)—are applied to classify PD patients by geographic origin. RFE identifies 15 key microbial taxa that distinguish between healthy controls and PD patients. Among the models trained on these taxa, the RF model exhibits the highest predictive accuracy, achieving 0.83 ± 0.048. Despite the relatively small sample size, this study underscores the necessity for larger-scale investigations and contributes to a more comprehensive understanding of gut microbiota disparities between Eastern and Western populations in the context of PD.

Parkinson’s Disease (PD) patients were reported to exceed approximately 10 million worldwide, and the number has been consistently increasing every year. Factors causing PD are genetic and environmental factors. We focus on the effects of environmental factors on epigenetic. Specifically, the α- synuclein (SNCA) gene is a risk factor causing PD, and we analyze it based on its regulation by the methylated 3’ Untranslated Region (UTR). The analysis data consists of Deoxyribonucleic Acid (DNA) methylation by NCBI GEO datasets, including 40 samples from PD cases and 38 from cognitively healthy subjects. The total number of CpG islands (CGI) is 863,159. To select the distinctive CGI from the DNA methylation data, we apply the Recursive Feature Elimination (RFE) method. Consequently, we have the results from selecting the 20 CGIs. Based on 20 CGI, we employ four models, Support Vector Machine (SVM), RandomForest (RF), Extreme Gradient Boosting (XGBoost), and Neural Network (NN), to divide the control and PD. Among the results, the RF model achieved the highest accuracy at 87.5%. Our findings indicate that changes in DNA methylation levels of the 20 selected CGIs are associated with PD occurrence. Based on these results, we suggest that PD occurrence is related to environmental factors such as chemical stress, eating habits, and general stress. Furthermore, our paper provides insights into the relationship between epigenetics and neurodegenerative disease.

This study analyzes the performance of ANFIS(Adaptive Neuro-Fuzzy Inference System) based on the input space partitioning method. Using body fat datasets and concrete compressive strength datasets, various fuzzy system configuration methods, such as Grid Partitioning, Subtractive Clustering, and FCM(Fuzzy C-Means), are compared. The results show that the FCM-based ANFIS model demonstrated superior performance, recording the lowest RMSE value. It is confirmed that the initialization method of the fuzzy system significantly influences the performance of ANFIS, and the optimal configuration method may vary depending on the data distribution and complexity.

This study compares and analyzes the performance of LIME-based machine learning methods (Gaussian Naive Bayes (GNB), Highly-Efficient Logistic Regression (LR), Linear Support Vector Machine (SVM), and Triple-layer Neural Network (TNN)) using three medical datasets. High-dimensional data increases the likelihood of overfitting in learning algorithms due to the curse of dimensionality. To address this, LIME is utilized to compute the importance of key features contributing to the model's predictions. Based on this, features are selected. The LIME technique generates multiple samples by perturbing the data in the local region. Subsequently, a simple linear model is used to evaluate the impact of each feature on the predictions. Features with high importance derived from this process are selected for model retraining. As a result, it was confirmed that learning time could be reduced while maintaining or even improving performance with a smaller number of features. Consequently, by selecting necessary features, the curse of dimensionality issue is alleviated, and accuracy can be maintained or improved using fewer features in the Hepatitis C Prediction Dataset, Breast Cancer Wisconsin (Prognostic) Dataset, and Glioma Grading Clinical and Mutation Features Dataset.

In low-visibility conditions, such as night-time driving or during bad weather, the risk of vehicle collisions rises dramatically. This is mainly because it becomes much harder to spot pedestrians and other obstacles on the road. To address this challenge, we developed a real-time vehicle collision prevention system using a Jetson Nano equipped with infrared cameras and the Yolov8n model. The system works by capturing heatemitting objects using thermal imaging technology, making it easier to detect obstacles even in difficult lighting conditions. To enhance its reliability, the system was mounted on a vehicle and tested in a range of environments, including both day and night. Throughout these tests, the system consistently proved capable of detecting potential hazards in real time, showcasing its potential to significantly improve driver safety and reduce collisions in challenging driving conditions.

Skin cancer, particularly melanoma, poses significant risks due to its high metastatic potential and challenges in early diagnosis. Accurately detecting skin lesions through automated systems is crucial for improving survival rates. This paper does not merely propose a detection method but analyzes the effectiveness of feature extraction for accurate skin lesion classification. Utilizing a dataset from Kaggle, this paper compares the performance of various deep learning models, including Convolutional Neural Networks (CNN), Deep Neural Networks (DNN), and ResNet-18. We evaluate the ability to classify skin lesions by training three models on 10,015 images across seven classes. ResNet-18 achieved the highest accuracy of 81.6%, demonstrating its potential for the development of automated diagnostic systems. In contrast, CNN and DNN attained lower accuracies of 72.9% and 70%, respectively, likely due to limitations in their feature extraction capabilities. These results underscore the superior performance of ResNet-18, particularly in its ability to handle complex patterns and deep feature learning, which are critical for skin lesion classification. In addition, we explored the potential integration of Large Language Model(LLM) to enhance the interpretability of diagnostic outcomes. By utilizing the Llama2 model API provided by Hugging Face, we explained the feasibility of interpreting ResNet-18's predictions to provide users with more transparent and higher-level medical insights. This suggests a promising future direction for improving the explainability and clinical applicability of AI-driven skin lesion diagnosis.

The recent introduction of information and communications technology (ICT) in the maritime sector has increased the number of contacts with external networks, increasing cybersecurity threats. Accordingly, the importance of cyber resilience to ships is being emphasized. The International Association of Classification Societies (IACS) emphasized holistic cyber resilience in its Unified Requirements (UR) E26, which lays out five functional elements and requirements for ships: identify, protect, detect, respond, and recover. However, UR E26 provides requirements but no practical techniques. This paper proposes a framework that integrates IACS UR E26 and the MITRE Cyber Resilience Engineering Framework (CREF) to provide a systematic way to enhance and manage ship cyber resilience. Since the proposed framework provides a practical description of the requirements of UR E26, it can be used for training cyber resilience response to ships and can improve cyber resilience through continuous improvements.