Advanced Future Information Technology

Inhaltsverzeichnis

1. Frontmatter

2. Classification of Diseases of Bean Plant Leaves Using Data Augmentation and Deep Neural Networks

   Bui Hai Phong, Phuong Anh Nguyen, Tran Dang Quyet, Phung Duc An, Phan Duc Tri, Le Anh Ngoc

   Abstract

   Bean plants have provided us with a rich source of food and industrial materials. However, diseases have significantly affected the production of bean plants. The recognition of bean plant diseases is an important step in the treatment of diseases. In recent years, deep neural networks have shown the leading performance for the image classification task. The paper presents a method for the recognition of diseased bean leaves. The vision of transformer (ViT) network has been applied to improve the accuracy of the recognition of diseased bean leaves. Moreover, data augmentation techniques have been applied to enlarge the image dataset to efficiently train the network. Obtained results of 97% for the classification of diseased bean leaves have shown the effectiveness of the proposed method.

3. Classification of Waste Bottles Using Image Processing and Deep Neural Networks

   Hai Phong Bui, Van Son Nguyen

   Abstract

   Waste management is becoming an important issue to protect our environment. The issues in the field have attracted many researchers in recent years. The automatic classification of waste is one of the most important issues. In recent years, with the advances in computer vision and deep learning, the classification of waste has obtained promising results. The paper presents a method to classify waste bottles using image processing and advanced deep neural networks. The image processing aims to improve the quality of input images. The Vision of Transformer (ViT), an advanced deep neural network, is applied to improve classification accuracy. Moreover, dimensional reduction techniques are used to efficiently visualize the feature extraction of the deep neural networks. Obtained results of 91% for the classification of bottles and the performance comparison with existing methods has shown the effectiveness of our proposed method.

4. Integrating NLP Platform to Build a Chatbot in the Hospitech System

   Tam Dinh Thi, Thang Ta Quoc, Khoai Vu Tien, Huyen Trinh Thi Thu, Thuy Nguyen Thu, Thai Nguyen Dinh, Hanh Nguyen Thi Phuong, Trung Tran

   Abstract

   Integrating an NLP platform into the Hospitech system leverages natural language processing (NLP) technology to develop chatbotAI capable of analyzing symptoms and providing online health consultations. The chatbot comprehends context and delivers accurate and flexible responses, improving communication quality between patients and doctors. It enables users to access information, schedule appointments, and receive support efficiently. With the ability to learn from new data, the chatbot continuously enhances accuracy and better meets patient needs. This reduces the workload of medical staff, optimizes schedule management, and improves user experience.

5. A Comparative Study of LGPMA and Table Transformer in Table Structure Recognition

   Tang Van Nguyen, Dang Hai Bui, Phuong Anh Nguyen

   Abstract

   Currently, there are numerous approaches to tackle the problem of recognizing table structures in unstructured documents, in which deep learning stands out as an efficient one. However, applying the deep learning algorithms in real-world scenarios poses significant challenges due to the diversity of tables found in documents evolving daily.

   Therefore, researching and comparing various TSR (TSR) models for identifying tables in practical documents is not only academically significant but also holds practical implications for applying Artificial Intelligence (AI) to real-world scenarios. Discovering a suitable or superior model for table recognition can help research groups and businesses save time, costs, and optimize product development pathways. In this paper, we perform a comparative study between the two state-of-the-art models through the years, namely Local and Global Pyramid Mask Alignment (LGPMA) and Table Transformer (TATR), using two datasets that we created to represent practical tables. These datasets include the Borderless Table dataset, representing tables commonly found in research articles, and the Vietnamese Tax Liability dataset. To evaluate the performance of the above models, we used the Tree-Edit-Distance-Based Similarity metric, considering only the table structure (TEDS-Struc.). Finally, we draw conclusions and suggest some recommendations for the practical applications of these methods.

6. An Assessment of Formative and Reflective Constructs in Ethnic Consumer Behavior Research

   Thi Thu Cuc Nguyen, Van Tang Nguyen

   Abstract

   This study employs Confirmatory Tetrad Analysis (CTA) to evaluate the specification and validation of formative and reflective constructs within a Partial Least Squares Structural Equation Modeling (PLS-SEM) framework. While the formative-reflective distinction is critical in construct development, its implications are often overlooked, leading to issues such as construct misspecification, identification challenges, and validation errors. Using a systematic approach, this research compares the utility of formative and reflective modeling approaches, demonstrating that certain constructs align better with a reflective framework, while one other exhibit stronger validity when modeled formatively. The findings highlight the importance of adopting a hybrid formative-reflective approach to accurately capture the complexity of latent constructs and their observed measures.

7. Pothole Detection and Classification Using YOLO Models

   Thien Thanh Tran Ngoc, Tien Thanh Do, Ngoc Anh Le, Thi Nhung Vuong, Ngoc Thanh Pham, Doan Dong Nguyen, Phuong Anh Nguyen

   Abstract

   The study introduces an advanced approach to pothole detection and classification implemented by YOLOv12 and shows greater performance improvements than YOLOv8. A new data set was created containing 1,000 pothole images (water and no water). The images, along with staggered ground truths, were recorded manually. We ran five different models on this data merged from YOLOv8 and YOLOv12. While, YOLOv12 surpasses YOLOv8 in detection and generalization, YOLOv8 outperformed YOLOv12 in classifying ponds. These results establish a foundation for integrating YOLO-based pothole detection into intelligent transportation systems, enhancing road safety and infrastructure maintenance. Future research may focus on further architectural optimizations and real-world deployment strategies to enhance system efficiency.

8. An Efficient Data Auditing Scheme for Multi-cloud Multi-replica Environments

   Haiyan Yu, Yuxin Cui, Chen Wang

   Abstract

   With the rapid development of information technology, cloud computing has become increasingly popular, attracting a large number of users to outsource personal or corporate data to the cloud for storage and management. However, with the growth of data outsourced to the cloud, issues of data integrity and security have become increasingly prominent. These issues have become a focus of concern for both users and cloud service providers. We propose a novel auditing scheme for multi-cloud multi-replica storage. This scheme leverages a cloud combiner for coordination. Additionally, we design a bit-map-based file storage mapping mechanism. This scheme can audit multiple replicas of files stored on different cloud servers and provide feedback on the integrity and storage locations of the corresponding replica files. Therefore, the proposed scheme can enhance data security and auditing efficiency in a multi-cloud environment. Security analysis and performance evaluation both indicate that the scheme provides reliable and efficient auditing for outsourced data in a multi-cloud setting.

9. DeepWaste: Deep Learning-Based Waste Classification

   Duc-Kien Bui, Tien-Thanh Do, Ngoc-Thanh Pham, Ngoc-Anh Le, Kim-Thai Dinh, Phuong-Thao Le, Phuong-Thao Nguyen, Minh-Quan Vu, Phuong Anh Nguyen

   Abstract

   This paper is based on the research process of applying the YOLO model to identify and classify waste. In the self-collected dataset, we tested 2 versions: YOLOv11 and YOLOv12. The test results show that two models have a reasonably high average accuracy of about 71–72% mAP50-95 to identify and classify waste types in the test dataset, a low error rate, and stable performance. Through the research, the reader can see the great uses of artificial intelligence in general and the YOLO model in particular in the field of waste identification and classification. Therefore, this research can become a reference for future research on the application of computer vision in image identification and classification.

10. Federated Unlearning: Efficient Data Removal Strategies and Challenges in Privacy-Preserving Machine Learning

    Qingyu Tan, Yan Li, Jungho Kang, Byeong-Seok Shin

    Abstract

    In recent years, the Right to Be Forgotten (RTBF) has emerged as a critical issue in promoting digital trust and AI safety, highlighting the necessity of effectively removing identifiable information from trained machine learning models. This necessity has stimulated the development of Machine Unlearning (MU), enabling ML models to selectively remove specific data contributions without retraining from scratch. Building upon MU, Federated Unlearning (FU) has been developed to address data deletion challenges inherent in Federated Learning (FL) scenarios, granting federated models the capability to selectively remove the contributions of specific clients or identifiable client-related data. However, the unique characteristics of federated learning pose additional challenges that cannot be fully addressed by existing centralized unlearning approaches. In this survey, we first provide a concise overview of the FU framework, followed by a detailed introduction to the concepts and definitions related to MU and FU. We then summarize the key challenges currently faced in federated unlearning research. Furthermore, we systematically classify existing federated unlearning algorithms according to their specific tuning objectives. Finally, we outline several promising research directions for future investigations in the field of federated unlearning.

11. Wetland Vegetation Identification Model Based on Improved Deeplabv3+ and Contrastive Learning

    Xiuhe Yuan, Han Li, Guoqing Ni, Zitong Liu, Sheng Miao, Chao Liu

    Abstract

    Aiming at the problem of insufficient classification accuracy of traditional methods in complex wetland environments, this paper proposes an improved DeeplabV3+ model (Ct-Deeplabv3+) integrating a comparative learning encoder and a dual attention mechanism: a clustering-based comparative learning mechanism is designed at the encoder side to optimize the feature representation in an unsupervised manner, and an improved inflated convolution strategy is combined to expand the sensory field. The decoder side introduces the dual modules of gated context transformation and channel-coordinate attention to strengthen the ability of multi-scale feature fusion and spectral weight assignment. The results of the ablation experiments show that the comparative learning encoder and the dual attention mechanism module can effectively improve the benchmark model accuracy. The Ct-Deeplabv3+ model is applied to the task of semantic segmentation of high-resolution multispectral UAV remote sensing images, and the experimental results show that the accuracy of Ct-Deeplabv3+ on the test set reaches 93.5%, and the recall and Kappa coefficients are 88.4% and 0.9318, which is better than that of the mainstream model. This paper provides a new technical model for high-precision dynamic monitoring of wetland vegetation, which has important application value for ecological restoration and sustainable development.

12. MetaAdvisor: An AI-Driven Metahuman System for Personalized Admissions Counseling

    Tung Vu, Phuong Anh Nguyen, Cong Duan Truong, Ngoc Le

    Abstract

    University admissions offices face mounting challenges in providing consistent, personalized guidance to thousands of prospective students while managing limited human counselor resources, creating an urgent need for intelligent digital solutions that maintain high-quality interactions. Current approaches including text-based chatbots and basic virtual assistants fail to establish meaningful emotional connections with users, struggle with complex contextual conversations, and typically operate as isolated tools rather than integrated components of a comprehensive admission ecosystem. We propose the Smart Metahuman Admissions Consulting System that addresses these limitations through three integrated technologies: a Retrieval-Augmented Generation pipeline achieving over 95% query accuracy with only 0.76-s response time, an Ernerf-based metahuman-rendering system creating photorealistic facial animations that maintain strong user visual engagement, and a multimodel edge computing framework that optimizes processing of multilingual interactions with only a 3.5% accuracy decrease for bilingual conversations. Experimental evaluations with our system demonstrated exceptional performance in handling both general admission inquiries (96% accuracy) and complex contextual questions, while creating such a convincing human-like experience that many participants reported moments where they forgot they were interacting with an AI system, confirming our approach provides an immersive consultation experience while offering institutions a scalable solution for modernizing their admission processes.

13. NormSoftmax Attention: Improving Transformer Model Performance

    Phuong Anh Nguyen, Anh Ngoc Le

    Abstract

    This paper addresses the indoor positioning problem with the objective of enhancing localization accuracy using data from WiFi signals. The performance of the positioning system is highly sensitive to fluctuations in WiFi signal strength, which can negatively impact accuracy. To address this challenge, we propose a transformer-based algorithm, leveraging its ability to effectively model complex dependencies in data and mitigate the impact of signal fluctuations. Attention weights play a crucial role in enabling the transformer model to focus on relevant features; however, traditional softmax-based attention mechanisms often lack the flexibility needed to adapt to variable signal environments. By exploiting the unique characteristics of WiFi fingerprint data, we investigate Dirichlet distribution-based self-attention to replace the original mechanism for improved performance. Incorporating the Dirichlet distribution allows for greater control over attention scores, enabling the model to assign more robust, context-dependent weights to each token. This customization dynamically adjusts the attention focus on various signal features, even amidst fluctuating WiFi signals, thereby enhancing the model’s adaptability and overall positioning accuracy. Experimental results illustrate the efficiency of the proposed methodology, which outperforms existing approaches and achieves an accuracy improvement of up to 64%.

14. Probabilistic Transformer Model for Uncertainty-Aware Learning

    Phuong Anh Nguyen, Anh Ngoc Le

    Abstract

    Predictive modeling in complex, uncertain environments poses significant challenges across diverse domains, where traditional machine learning approaches often fail to effectively quantify inherent uncertainties. This research introduces Probabilistic Transformers, an innovative architectural approach that fundamentally transforms uncertainty estimation by generating probabilistic distributions instead of single-point predictions. Through a comprehensive case study in indoor localization using Received Signal Strength Indicator (RSSI) data, we demonstrate the model’s superior capabilities in handling signal variability and environmental complexity. Our Probabilistic Transformer model achieves unprecedented performance by explicitly capturing and quantifying prediction uncertainties, outperforming state-of the art methods in both accuracy and uncertainty estimation. Experimental results validate the model’s ability to provide more precise location predictions, adapt to noisy indoor environments, and offer nuanced uncertainty assessments that traditional approaches cannot match. By bridging the gap between deterministic prediction and probabilistic reasoning, this research contributes to the emerging field of uncertainty-aware machine learning, presenting a flexible framework that promises significant advancements in decision-making systems across signal processing, robotics, and other domains requiring robust, reliable predictive modeling.

15. Federated Learning for Greenhouse Temperature Prediction: A Privacy-Preserving Approach Using N-BEATS

    Trang Ha, Phuong Anh Nguyen, Tung Vu, Ngoc Le

    Abstract

    This paper presents a federated learning model for greenhouse temperature prediction based on the N-BEATS deep learning model. Precise greenhouse temperature forecasting facilitates the improvement of greenhouse environmental control and thereby reduces energy consumption as well as maximizes agricultural yield. In addition to addressing the prediction issue, the model addresses problems concerning data privacy and decentralized learning models in general for smart agriculture. The test's findings show that the federated learning model with the N-BEATS structure achieves full-size predictive power with Mean squared error (MSE) of 0.0053 and Root mean square error (RMSE) of 0.0975, and with stable convergence. Empirical proof validates the use of the federated learning model in the greenhouse climate regulation as proof of strong predictive capacity as well as the maintenance of data confidentiality. There's a lot of promises with the current model for actual use in smart agriculture, as well as with comprehensive machine learning models running in distributed environments.

16. Rumor Detection Method Based on Multi-modal Mixture of Experts and Cross-modal Enhancement

    Jianyong Yu, Xiuyu Li, Xue Han

    Abstract

    Rumor spreaders are increasingly using multimedia content on social platforms to attract and mislead the public. To alleviate this issue, research on rumor detection has received extensive attention in recent years. Although current multimodal rumor detection models utilize both textual and visual features of rumors for detection, they overlook the problem of text-image mismatch in rumors and the differences in the extraction processes between textual data and image data (such as temporal information and spatial structure). This paper proposes a new rumor detection method based on Multi-modal Mixture of Experts with Cross-modal Enhancement (MMoE). This method uses a contrastive language-image pre-trained model to calculate the similarity between text and images. Through a gating unit, different experts are selected based on the similarity to ensure that the model's predictions are not misled by incorrect information. In addition, a visual encoder-decoder is employed to convert post images into descriptive text for data augmentation, leveraging additional text data to optimize the model's performance. Experimental results show that the rumor detection results of MMoE outperform six baselines and achieve an accuracy rate of 92.49% on a Chinese dataset.

17. Soft Porn Identification Based on Social Media-- User Engagement Perspective

    Jingmeng Dou, Xuepeng Xu, Xiaofeng Yu

    Abstract

    Pornography has existed since the early days of the Internet. With the rapid growth of video-sharing platforms, a new form of pornography, online soft pornography, has emerged. While its harms are evident, few studies have focused on its conceptualization or identification. This paper proposes a deep learning method to identify soft pornography using a new perspective—user engagement. We utilize video data from Bilibili's dance section to measure user engagement from both content creators and users based on interactive characteristics and build a learning model. The results show that the model based on user engagement outperforms the text-based model (87.6% vs. 83.5%) in identifying soft pornography, and the fusion model further improves the performance (88.5%). Principal component analysis reveals that in selective attention-driven participation, users of soft pornography exhibit a significantly higher interest index. This study provides a detailed understanding of user engagement with soft pornography on video platforms and offers new insights for content regulation.

18. Non-Fungible Token Verification Scheme Based on Core Peripheral Sharding

    Zijuan Chen, Jianyong Yu, Yulong Wang

    Abstract

    As the fusion of the real world and the digital space, the metaverse is constructing a new interaction paradigm and economic system. Non-Fungible tokens (NFTs) have become a key technology to solving the digital asset ownership problem by their uniqueness and indivisibility. There are three core flaws in existing NFT verification schemes regarding privacy protection. First is the risk of identity exposure. Second, compatibility is insufficient. Traditional protocols rely on the homogenization property of currency, which is difficult to adapt to the indivisibility of NFT. Third, there is a trade-off between verification efficiency and privacy. These problems require new schemes to achieve anonymity, data confidentiality and resource consumption optimization while ensuring ownership verifiability.

    To solve these challenges, we propose a privacy-preserving NFT scheme based on dynamic commitment and fragmentation architecture (NFTCP). Its innovation is reflected in three aspects: sharded Merkle tree architecture, a two-level cryptographic commitment mechanism, and a lightweight zero-knowledge proof circuit to verify commitments. Experiments show that the NFTCP is significantly better than the traditional NFT scheme regarding anonymity and confidentiality. Although the processing time of NFTCP is higher than that of ERC-721, its effective transaction ratio is increased by 106 percent than that of ERC-721 and 4.5 percent than that of other schemes, this cost is verified to be a reasonable compromise for privacy protection.

19. Integrating TCM Constitution Theory with Deep Learning for Tongue Image Analysis

    Zixuan Chen, Jin Wang, Yuzhen Liu, Xiaolan Zhou, Xiaoliang Wang

    Abstract

    This paper presents a comprehensive tongue diagnosis system that integrates Traditional Chinese Medicine (TCM) constitution theory with advanced deep learning techniques. Tongue diagnosis, a vital component of TCM's “inspection” method, has long relied on practitioners’ subjective experience, leading to inconsistent diagnostic standards and susceptibility to external factors. To address these challenges, we propose a system combining YOLOv11 for detection, U2-Net for segmentation, and ResNet50 for classification of tongue images. Our system leverages the strengths of these models to overcome limitations in traditional tongue diagnosis, offering an efficient and accurate solution. We constructed a standardized dataset of 3,000 tongue images across nine constitutional types, collected under varied lighting conditions to ensure robustness. The experimental results demonstrate the system's effectiveness in automatically detecting, segmenting, and classifying tongue images, with significant improvements in diagnostic accuracy and efficiency. This study paves the way for the objectification and standardization of TCM inspection, providing a promising tool for mobile-based TCM diagnostic assistance.

20. Intelligent Pulse Diagnosis in Traditional Chinese Medicine: Signal Processing and Deep Learning for Health Status Classification

    Zixuan Chen, Jin Wang, Yuzhen Liu, Xiaolan Zhou, Xiaoliang Wang

    Abstract

    The intelligent pulse diagnostic instrument, developed by Jianwei TCM Studio, integrates signal processing and deep learning for standardized pulse wave acquisition and recognition. This device combines hardware, including a sensor, pressure device, signal amplification module, and STM32 microcontroller, with software featuring Alibaba Cloud servers, HVG networking, and an attention-based ACRNN for pulse wave analysis and physique recognition. It offers two primary functions: standardized pulse wave data acquisition and cloud storage, enabling health monitoring through a user-friendly, portable design; and pulse wave preprocessing and physique recognition using HVG and ACRNN, achieving high diagnostic accuracy for sub-healthy populations. A prototype has been successfully developed with all functional modules operational.

21. KITE-MRE: Knowledge-Infused and Transformer-Enhanced Multimodal Relation Extraction

    Junjie Li, Wenti Huang, Yu He, Xinjie Mo

    Abstract

    Multimodal Relation Extraction (MRE) is a task that identifies semantic relations between two entities from image-text pairs. Existing methods often suffer from insufficient extraction accuracy and low cross-modal fusion efficiency due to inadequate auxiliary information. In this paper, we propose a multimodal relation extraction framework KITE-MRE,which integrating large model generation and knowledge graph (KG) to provide abundant additional auxiliary information. Specifically, prompts guide large language models to construct image semantic entity triplets and generate image graphs, while knowledge graph entity linking technology is applied in text graphs to build a unified graph structure. Additionally, we introduce the Graph Information Bottleneck (GIB) technique combined with Graph Convolutional Networks (GCN) to optimize, and perform feature fusion by integrating semantic and structural multi-perspective information to predict relations. Experimental results demonstrate that this approach provides a new path for multimodal data processing, effectively addressing issues of knowledge graph incompleteness and ambiguous relation generation by large models in cross-modal scenarios.

22. A Lightweight Data Indexing Mechanism with Low Query and Maintenance Overhead

    Jingyu Zhang, Di Lan, Zisang Xu, Jin Wang

    Abstract

    Edge storage systems aim to store data on edge nodes to provide users with reliable and low-latency data retrieval services. This is especially important for latency-sensitive applications, where users need to locate the target data and retrieve it quickly under multi-hop delay constraints. Although users can query neighboring nodes through serial or parallel retrieval methods, these approaches often lead to significant redundant communication and resource consumption across the edge network. Therefore, designing an efficient indexing mechanism is essential. However, current research on indexing mechanisms for local retrieval in edge storage systems remains limited. In this paper, we propose a tree-based indexing structure built upon the probabilistic data structure known as the Counting Bloom Filter (CBF). By incorporating the advantages of Delaunay Triangulation (DT) graph, the proposed indexing structure enables efficient data query and retrieval. Experimental results show that, compared with existing local retrieval indexing structures such as EDIndex-CBF and EDIndex-HCBF, our approach achieves lower maintenance and query overhead.

23. A Novel PSO-Based Adaptive Container Scheduling Strategy for Edge Computing

    Jingyu Zhang, Hanbo Jiang, Zisang Xu, Jin Wang

    Abstract

    Edge computing has emerged as a critical technology for achieving low latency and real-time data processing, especially in applications with the Internet of Things (IoTs) and smart cities. With their lightweight and portable nature, containers provide an effective solution for deploying and managing edge applications. The Particle Swarm Optimization (PSO) algorithm, inspired by the collective behavior of bird flocks, offers an efficient approach to complex optimization problems by iteratively updating particle positions within a multidimensional solution space. PSO's low computational complexity and strong global search capability make it highly suitable for resource allocation and task scheduling in edge environments. This paper presents an adaptive scheduling strategy based on PSO. In this approach, a time-series-based Dynamic Request Prediction Algorithm (DRPA) anticipates request patterns to dynamically adjust container placement, minimizing response time. Concurrently, a threshold-based PSO (TPSO) optimizes container scheduling based on node load, enhancing resource utilization and reducing scheduling delays. Experimental results show that DRPA-TPSO effectively lowers latency, boosts resource utilization, and improves system stability in high-load, complex environments, outperforming traditional scheduling methods.

24. A Tree-Based Pipeline Consensus with High Scalability and Throughput

    Jingyu Zhang, Sheng Jiang, Zisang Xu, Jin Wang

    Abstract

    Edge computing has become an increasingly popular paradigm as part of distributed computing architecture. It accomplishes this by facilitating data from end devices to be stored and processed at the edge of the network close to the data. Unfortunately, consensus scalability and throughput become a big issue in collaborative edge scenarios. Byzantine consensus is a promising consensus solution that is feasible in small-scale collaborative edge systems. However, most of the existing consensus algorithms cannot meet the increasing number of nodes in collaborative edges, which poses a big challenge to the throughput of consensus algorithms. In this paper, we propose a tree-based pipeline consensus, a variant BFT (Byzantine Fault Tolerant) that maintains high throughput as the system size grows, utilizing a novel pipelining technique that performs scalable splitting and recombination on the tree to increase the efficiency of BLS (Boneh-Lynn-Shacham). We conduct experiments through simulations. The experimental results show that the proposed consensus mechanism can effectively improve the performance of collaborative edges, including throughput and consensus latency.

25. Design and Evaluation of P-GAN for Privacy Protection and Generative AI Regulatory Compliance

    Seo Jeong Min, Ji Su Park, Jin Gon Shon

    Abstract

    This study deals with the design and implementation of a learning model to prevent the leakage of sensitive information of the original data during the learning process of a model that generates high-quality data that is difficult to distinguish from the original using a generative adversarial network (GAN). Existing European General Personal Information Protection Act (GDPR) and domestic Personal Information Protection Act had limitations in that it was difficult to effectively respond to the privacy leakage occurring in the characteristics of the Generative AI model and the learning method. In order to improve this problem, AI Act that applies differential regulations for each risk level of the system was introduced. Therefore, this study aims to introduce personal information enhancement technology (PETs) into the GAN model design process to promote a balance between the usefulness of the learning model and privacy enhancement, and to propose a new P-GAN model that can quantitatively evaluate the learning results. In this paper, by describing the design principles and implementation results of the proposed model in detail, we seek the direction of improvement of the future generative model's usefulness and privacy protection function enhancement system.

26. A Survey on AI-Based Anomaly Detection for Cloud Security in IIoT Environments

    Jungho Kang

    Abstract

    The Industrial Internet of Things (IIoT) has revolutionized industrial operations with seamless connectivity and real-time decision-making, but the reliance on cloud computing poses significant security risks, such as data breaches and advanced persistent threats (APTs). Traditional anomaly detection methods are inadequate for the dynamic and complex IIoT-cloud ecosystem. This survey examines AI-based anomaly detection techniques, evaluating their performance in data confidentiality, real-time capabilities, scalability, and robustness against evolving cyber threats. Identified challenges include computational inefficiencies, scalability constraints, and insufficient privacy-preserving mechanisms. Based on these findings, future research directions include developing Federated Hybrid AI Frameworks that integrate Federated Learning, Reinforcement Learning, and Generative Adversarial Networks to address security and scalability challenges. Additionally, incorporating explainable AI techniques is crucial to enhance transparency and trust in detection systems. Aligning these efforts with stringent IIoT security demands is essential for building robust and adaptive frameworks for Industry 4.0.

27. Backmatter