In the wake of natural disasters, terrestrial communication infrastructure is often rendered inoperative, severely hampering emergency response efforts. This paper proposes a resilient communication framework leveraging High-Altitude Platforms (HAPs) to provide rapid, wide-area connectivity in disaster-stricken regions. We integrate a reinforcement learning (RL)-based algorithm that dynamically allocates bandwidth and adjusts coverage in real time, enabling efficient and adaptive resource management based on environmental feedback and user demand. The proposed system enhances communication reliability, reduces latency, and optimizes spectral efficiency under uncertain and evolving disaster scenarios. Simulation results demonstrate the effectiveness of the RL-driven HAP network in maintaining robust communication links and improving overall disaster response coordination.
Human Pose Estimation (HPE) aims to predict the positional coordinates of body keypoints in images. While significant progress has been made in HPE, certain challenges persist. For example, the potential for robust occlusion can result in considerable confidence in predicting false-positive keypoints. This can cause problems in applications where high detection accuracy is required, such as in a traffic environment. Most existing methods focus on predicting the positional coordinates of each human joint, but they fail to consider the keypoint occlusion problem and detect the visibility of keypoints in the image. In this paper, we propose a visibility-guided human pose estimation model, named ViPose. We extract the visibility information of keypoints to detect whether a specific keypoint of the subject in the image is obscured by other objects, visible or not captured. ViPose is composed of two modules: HPE keypoint coordinate predictor module, and keypoint visibility detector module. To verify the effectiveness of ViPose, we conduct experiments on both the general MSCOCO dataset and the diver monitoring DriPE dataset. The experimental results show that ViPose can achieve good performance with fewer parameters.
This paper presents a method for transmitting multiple synchronization signal blocks (SSBs) in 5G New Radio (NR) systems. When multiple SSBs with same physical cell identity are transmitted, the peak-to-average power ratio can increase. To reduce the PAPR, non-cell defining SSB is transmitted with a symbol delay relative to the cell-defining SSB. The simulation results show the reduction of PAPR under multiple SSBs transmission.
The objective of this study is to define a systematic taxonomy for assessing the fairness of Korean-based large language models (LLMs). To achieve this, various international documents, datasets, and academic papers are collected and analyzed for their presented fairness and bias classification methods. As a result, we propose a classification framework that includes superclasses for fairness assessment and subclasses reflecting the specific characteristics of Korean society. The proposed taxonomy provides a more systematic and detailed basis for assessing the fairness of Korean-based LLMs. This approach is expected to effectively address ethical issues in Korean language models and contribute to the development of AI technology suitable for Korean society.
Index modulation (IM) is a promising technique for enhancing system resilience against jamming attacks in orthogonal frequency division multiplexing (OFDM) systems. Dual-mode IM OFDM (DM-OFDM) offers notable potential among the various types of IM techniques. Motivated by this fact, in this paper, we explore the robustness of DM-OFDM under jamming attacks in comparison to the traditional OFDM and OFDM-IM systems. Our investigation focuses on a practical resource mapping scenario relevant to 5G cellular standards. The key finding is that while the DM-OFDM system demonstrates strong performance under limited jamming, its advantage over the compared systems decreases as the jamming power increases. Nevertheless, the DM-OFDM system continues to show robustness when the impact of the jamming signals is restricted to a small subset of subcarriers.
This paper introduces an example of AI(Artificial Intelligence) engine implementation for detecting events related to emergency situations such as collapse and assault, especially for elevator passengers. In contrast to the security cameras installed in a relatively open-space environment, those cameras installed in an elevator environment are likely to capture subjects from a high angle and at a relatively close distance. In an elevator environment, sometimes, only a part of subject may be captured. In addition to that, a part or whole of subject may be reflected in the wall or mirror and, moreover, those reflected things may be also captured. Considering those differences, to implement the above AI engine, the authors of this paper picked action/video recognition methods. And, the authors also built a prototype-level data-set to train and evaluate the related neural network models. This paper describes the tasks and methods to implement the above AI engine and also shows some demonstration results.
The rapid adoption of the Internet of Things (IoT) has led to the deployment of numerous devices communicating through protocols, with the Message Queuing Telemetry Transport (MQTT) being one of the most widely used. However, MQTT and IoT devices are susceptible to various cyber threats due to their inherent limitations in security methods. This study offers a brief analysis of intrusion detection methods designed especially for MQTT-enabled IoT networks. The study outlines the difficulties posed by the unique characteristics of MQTT, such as its resource-constrained environment and lightweight architecture, and it examines several detection techniques, such as hybrid, anomaly-based, and signature-based approaches. The paper highlights the effectiveness of these techniques in detecting various intrusion attacks including DoS attacks, unauthorized access, and data tampering, and concludes with suggestions for future research areas focusing on advanced machine-learning techniques and scalable security mechanisms tailored to the dynamic nature of IoT networks.
The increasing urgency for innovative cohesive networks stems from the necessity to integrate various networks, services, and modern technologies. Researchers are progressing in integrating diverse networks, modern technologies, and services through cohesive networks to facilitate more streamlined operations. A unified infrastructure enables effective data sharing from various sources, breaking down silos that typically hinder data exchange. Hydra radio access networks (H-RANs) are designed to achieve researchers' ambitious objectives by seamlessly combining diverse networks and technologies into a single, intelligent, perceptive, and dense infrastructure. However, the proliferation of dense deployments of traditional wireless local area networks (WLANs) is significantly contributing to excessive power consumption and encountering substantial interference challenges. In this paper, we propose a novel collaborative approach utilizing the H-RAN perceptual network architecture for distance-based adaptive power control and interface mitigation. By harnessing this integration into a cooperative multi-sparse input/multi-task learning-based federated learning (C-SMTL) framework. The cooperative architecture adeptly addresses the challenges associated with dynamic power control and interference coordination. Simulation results show that our strategy significantly improves network performance. This is done by achieving high throughput capacity, reducing interference by 90%, and maintaining high efficiency even with increasing user density.
Text classification is the process of categorizing text into predefined labels and is a core task in natural language processing. This process enhances system efficiency in various applications such as spam email filtering, news categorization, and sentiment analysis of comments. However, fine-tuning involves training all parameters, leading to increased spatial and temporal complexity. To address this issue, Low-Rank Adaptation (LoRA) has been proposed, which reduces complexity by significantly decreasing the rank of the training parameters. In this study, we apply LoRA to the embedding layer of bidirectional encoder representations from transformers (BERT), reducing the rank of the parameters and lowering complexity. This approach effectively operates in the embedding layer and can improve performance. Experimental results show that by adding and training a new token to appear at the end of sentences, we achieved excellent performance in text classification tasks. These results demonstrate that LoRA can effectively operate in the embedding layer, suggesting its potential contribution to enhancing the performance of Transformer models.
The widespread distribution of illegal pornographic material has become a significant societal concern due to the ease of sharing digital content over the internet. Existing detection methods often rely on manual review or keyword filtering, which are time-consuming and ineffective against evolving content. This paper proposes a novel real-time method for detecting duplicate images. The proposed method combines Convolutional Neural Network (CNN) based feature extraction with image hashing techniques for efficient image comparison. In this paper, we describe the dataset composition and procedures for training, and finally, we perform a performance evaluation.
The increasing aging population necessitates intelli-gent monitoring systems that can accurately detect and classify daily activities, enabling proactive healthcare interventions and promoting independent living for the elderly. To this end, this study presents ADL recognition system for elderly care, leveraging advanced 3D CNNs with skip connections and self-attention mechanism. Our proposed model processes RGB video data from the ETRI-Activity3D dataset. The architecture incorporates skip connections to mitigate the vanishing gradient problem and a self-attention mechanism to capture complex spatiotemporal features. Experimental results demonstrate the superiority of our approach, achieving 99.0% accuracy on the training set and 96.1% on the validation set. Notably, our model outperforms the baseline ETRI(FSA-CNN) implementation by 6%. The study provides both quantitative and qualitative evaluations, including accuracy metrics, loss curves, and visual representations of classification outcomes. These thorough results highlight the robustness and effectiveness of our proposed system in real-world scenarios. This research contributes to the evolving landscape of technology-assisted elderly care, offering a foundation for future developments in anomaly detection, risk management, and personalized alert systems. The proposed ADL recognition system has the potential to greatly improve the quality of life for elderly individuals and streamline healthcare resource allocation.
Additive Manufacturing (AM) technology allows for the rapid production of complex-shaped parts., but post- processing is essential to address issues of surface quality and mechanical properties. Specifically., the cleaning process is crucial for improving part quality., requiring accurate monitoring of the cleaning solution's condition. Traditional methods like chemical analysis or periodic replacement are costly., complex., and inefficient. This paper proposes a turbidity inspection method using Multi-Scale Laplacian Variance (MSL V) to enhance the efficiency of the cleaning process in post- processing in Additive Manufacturing Environments.
In this study, we analyze the performance of handover algorithms in scalable cell-free massive multiple-input multiple-output (CF-mMIMO) systems under mobility scenarios. To facilitate a comparative performance analysis considering mobility, we introduce a mobility-aware spectral efficiency met-ric. We then evaluate two baseline access point (AP) clustering methods and two mobility-aware handover algorithms. The simulation results reveal a significant decrease in minimum spectral efficiency across all algorithms and mobility scenarios, with differences of up to 7.5 times observed. This degradation is primarily due to existing algorithms prioritizing APs with the highest large-scale fading coefficients (LSFCs) when forming clusters, which prevents certain user equipments (UEs) from forming adequate AP clusters. The findings emphasize the need to design AP clustering algorithms that enhance minimum spectral efficiency, thereby ensuring the principle of CF-mMIMO in providing uniform performance.
In an environment of collaborative perception and cooperation intelligence, where objects autonomously perceive changes in their surroundings and perform tasks through autonomous collaboration with nearby objects, continually self-improving, the intelligence models required for the operation of intelligent things must undergo processes such as creation, registration, download, and updates before being deployed to the intelligent things. However, during these processes, if a malicious user intercepts and tampers with the models, resulting in the deployment of malicious models to the model registry and intelligent things, it can lead to inappropriate behavior of the intelligent things, causing severe personal or property damage or the failure of critical missions (e.g., rescue operations). This paper discusses methods to enhance the security and reliability of models by applying blockchain technology to the model exchange flow in a cooperation intelligence environment.
This paper proposes a two-stage AI-based method for automatically inspecting the fastener strength of generator stator wedges. This approach includes a ‘Noise removal stage’ employing a CNN-based autoencoder to eliminate industrial noise, and a ‘Classification stage’ extracting various features from the denoised signals to classify the fastener strength. Each stage of the proposed system demonstrates high accuracy and objectivity, significantly improving inspection efficiency by eliminating the rotor removal process. Moreover, the system effectively filters out industrial noise and is scalable, making it suitable for practical use in industrial sites with robotic inspection units.
The dyeing industry is known for its high energy consumption, but traditionally, more emphasis has been placed on reducing product defect rates rather than on energy savings. However, with the growing importance of reducing carbon emissions due to global warming, there is increasing interest in lowering energy consumption in dyeing factories. This paper discusses the installation of additional sensors for effective exhaust fan control in tenter machines and the setup of a demonstration site to implement these measures.
This paper analyzes the strength of Message Digest Algorithm (MD5) by performing deep learning-based Encryption Emulation (EE) and Plaintext Recovery (PR) attacks. We convert randomly generated S12-bit arrays, messages, into 128-bit arrays, digests, with MD5 in different numbers of steps. Furthermore, two different structures of deep learning models, fully-connected neural network and Bidirectional Long Short-Term Memory (BiLSTM), are used in attacks and trained to analyze MD5 automatically. As a result, the BiLSTM shows better prediction accuracy than the fully-connected neural network. Moreover, the PR attack is more challenging than the EE attack.
Constructing a training dataset is essential to train a deep learning model. However, limited budgets and the high cost of data annotation present significant challenges. Active learning has emerged to resolve this issue, and it incrementally selects useful samples to be labeled and added to the training dataset. Existing works primarily focus on expanding dataset without taking into account the usefulness of the generated annotations. Since annotation is inevitably imperfect when performed by human experts, it is essential to consider the reliability of the annotations. In this paper, we propose an enhanced active learning framework with selective removing for the efficient allocation of data. The proposed method gradually obtains valuable data while filtering out uninformative data on the training dataset and model knowledge. We evaluate the proposed method for the image classification task under a constrained data budget. Experimental results demonstrate that our approach outperforms existing active learning methods.
Federated learning is a distributed machine learning framework that enhances privacy and data security. However, model parameters communicated between server and clients are vulnerable to gradient leakage attacks (GLAs), posing a significant privacy risk. While differential privacy offers strong protection, it often comes at the cost of considerable performance degradation. To address this challenge, we explore a channel-level gradient leakage attack for the first time, discovering that successful attacks can be conducted using only partial channel information from certain parameters. Based on this insight, we leverage the empirical Fisher information matrix to estimate the information contained in each channel and propose a lightweight channel-level privacy protection method (FedLCP) to defend against GLAs and our newly explored channel-level gradient leakage attack. Extensive experiments on three datasets demonstrate that FedLCP provides privacy protection comparable to differential privacy, with up to a 37.78% improvement in accuracy and excellent convergence.
