Earticle

다양한 분야에서 인공지능을 활용한 사례가 증가하면서 침입탐지 분야 또한 다양한 이슈를 인공지능을 통해 해결하려는 시도가 증가하고 있다. 하지만, 머신러닝을 통한 예측된 결과에 관한 이유를 설명하거나 추적할 수 없는 블랙박스 기반이 대 부분으로 이를 활용해야 하는 보안 전문가에게 어려움을 주고 있다. 이러한 문제를 해결하고자 다양한 분야에서 머신러닝의 결정을 해석하고 이해하는데 도움이 되는 설명 가능한 AI(XAI)에 대한 연구가 증가하고 있다. 이에 본 논문에서는 머신러닝 기반의 침입탐지 예측 결과에 대한 신뢰성을 강화하기 위한 설명 가능한 AI를 제안한다. 먼저, XGBoost를 통해 침입탐지 모 델을 구현하고, SHAP을 활용하여 모델에 대한 설명을 구현한다. 그리고 기존의 피처 중요도와 SHAP을 활용한 결과를 비교 분석하여 보안 전문가가 결정을 수행하는데 신뢰성을 제공한다. 본 실험을 위해 PKDD2007 데이터셋을 사용하였으며 기존의 피처 중요도와 SHAP Value에 대한 연관성을 분석하였으며, 이를 통해 SHAP 기반의 설명 가능한 AI가 보안 전문가들에게 침입탐지 모델의 예측 결과에 대한 신뢰성을 주는데 타당함을 검증하였다.

Efficient use of limited blood products is becoming very important in terms of socioeconomic status and patient recovery. To predict the appropriateness of patient-specific transfusions for the intensive care unit (ICU) patients who require real-time monitoring, we evaluated a model to predict the possibility of transfusion dynamically by using the Medical Information Mart for Intensive Care III (MIMIC-III), an ICU admission record at Harvard Medical School. In this study, we developed an explainable machine learning to predict the possibility of red blood cell transfusion for major medical diseases in the ICU. Target disease groups that received packed red blood cell transfusions at high frequency were selected and 16,222 patients were finally extracted. The prediction model achieved an area under the ROC curve of 0.9070 and an F1-score of 0.8166 (LightGBM). To explain the performance of the machine learning model, feature importance analysis and a partial dependence plot were used. The results of our study can be used as basic data for recommendations related to the adequacy of blood transfusions and are expected to ultimately contribute to the recovery of patients and prevention of excessive consumption of blood products.

Identifying informative features in bioinformatics is challenging due to their small proportion within large datasets. We propose a scalable and interpretable feature selection framework for cancer RNA-seq by transforming non-image bio-data into 2D formats and applying convolutional neural networks (CNNs) with transfer learning for efficient classification. Explainable artificial intelligence (XAI) techniques identify and prioritize important features, while principal component analysis (PCA) determines the optimal number of selected features, ensuring transparency and reliability. Comparative analysis of CNN and XAI highlights the effectiveness of our approach, providing a robust framework for high-dimensional genomic data analysis with applications in cancer diagnosis and prognosis.

Explainable AI (XAI) has found numerous applications in computer vision. While image classification-based explainability techniques have garnered significant attention, their counterparts in semantic segmentation have been relatively neglected. Given the prevalent use of image segmentation, ranging from medical to industrial deployments, these techniques warrant a systematic look. In this paper, we present the first comprehensive survey on XAI in semantic image segmentation. We analyze and categorize the literature based on application categories and domains, as well as the evaluation metrics and datasets used. We also propose a taxonomy for interpretable semantic segmentation, and discuss potential challenges and future research directions.

Digital twins (DTs) are an emerging digitalization technology with a huge impact on today’s innovations in both industry and research. DTs can significantly enhance our society and quality of life through the virtualization of a real-world physical system, providing greater insights about their operations and assets, as well as enhancing their resilience through real-time monitoring and proactive maintenance. DTs also pose significant security risks, as intellectual property is encoded and more accessible, as well as their continued synchronization to their physical counterparts. The rapid proliferation and dynamism of cyber threats in today’s digital environments motivate the development of automated and intelligent cyber solutions. Today’s industrial transformation relies heavily on artificial intelligence (AI), including machine learning (ML) and data-driven technologies that allow machines to perform tasks such as self-monitoring, investigation, diagnosis, future prediction, and decision-making intelligently. However, to effectively employ AI-based models in the context of cybersecurity, human-understandable explanations, and their trustworthiness, are significant factors when making decisions in real-world scenarios. This article provides an extensive study of explainable AI (XAI) based cybersecurity modeling through a taxonomy of AI and XAI methods that can assist security analysts and professionals in comprehending system functions, identifying potential threats and anomalies, and ultimately addressing them in DT environments in an intelligent manner. We discuss how these methods can play a key role in solving contemporary cybersecurity issues in various real-world applications. We conclude this paper by identifying crucial challenges and avenues for further research, as well as directions on how professionals and researchers might approach and model future-generation cybersecurity in this emerging field.

Considering factors such as illumination, camera quality variations, and background-specific variations, identifying a face using a smartphone-based facial image capture application is challenging. Face Image Quality Assessment refers to the process of taking a face image as input and producing some form of "quality" estimate as an output. Typically, quality assessment techniques use deep learning methods to categorize images. The models used in deep learning are shown as black boxes. This raises the question of the trustworthiness of the models. Several explainability techniques have gained importance in building this trust. Explainability techniques provide visual evidence of the active regions within an image on which the deep learning model makes a prediction. Here, we developed a technique for reliable prediction of facial images before medical analysis and security operations. A combination of gradient-weighted class activation mapping and local interpretable model-agnostic explanations were used to explain the model. This approach has been implemented in the preselection of facial images for skin feature extraction, which is important in critical medical science applications. We demonstrate that the use of combined explanations provides better visual explanations for the model, where both the saliency map and perturbation-based explainability techniques verify predictions.

Artificial Intelligence (AI) has been used extensively in all aspects of everyday life among people in recent times. Many techniques utilizing machine learning (ML) and deep learning (DL) models are being presented in this rapidly growing field of study. Most such models are generally regarded as “Black-Box” models since they are intrinsically complex and lack interpretable explanations for their decisions and conclusions. The lack of transparency increases the issue in the field of cybersecurity as implementing critical decisions in a system that cannot provide explanations for itself offers some evident risks. The lack of interpretability and transparency in existing AI techniques would make users distrust the models used to defend against cyberattacks, particularly given the increasingly complex and diverse nature of cyberattacks. Thus, Explainable Artificial Intelligence (XAI) must be utilized to construct cyber security models that are more understandable while keeping high accuracy and that enable users to understand, be reliable, and manage the future of cyber defence systems. This study provides a comprehensive survey of existing literature on using XAI to mitigate these challenges of cybersecurity black-box models. It emphasizes the significance of explainability in boosting faith and transparency in AI-driven systems and presents a thorough taxonomy of XAI techniques and technologies for cybersecurity applications. The study describes the evaluation criteria that are used to evaluate the effectiveness of XAI models, addresses different kinds of attacks like malware, phishing, and network intrusions, and shows how XAI techniques may mitigate these risks by providing a comprehensible understanding of model decisions. Along with the real-world case studies, it also explores the industrial applications of XAI in cybersecurity and examines the challenges in implementing XAI technology. The survey concludes with a review of the limitations of the existing XAI techniques and makes recommendations for future research, such as the requirement for more reliable XAI frameworks that can function in real-time and across a variety of cyber threat situations.

Data-driven learning is the most representative deep learning method. Generative adversarial networks (GANs) are designed to generate sufficient data to support such learning. The learning process of GAN models typically trains a generator and discriminator in turn. However, overfitting problems occur when the discriminator depends excessively on the training data. When this problem persists, the image created by the generator shows a similar appearance to the learning image. Images similar to learning images eventually lose the meaning of data augmentation. In this paper, we propose a limited discriminator GAN (LDGAN) model that explains the results of GAN, which is a model that can not be analyzed externally, such as a black box. The part explained in LDGAN becomes the discriminator model of GAN, and it is possible to check which area of the image is used as the basis for determining fake/real by the discriminator. In the end, a method for limiting the learning of discriminator is proposed based on the described results. Through this, it is possible to avoid the overfitting problem of the discriminator and to generate various images different from the learning image. The LDGAN method allows users to perform meaningful data augmentation with only specific objects except for complex images or backgrounds that require analysis. Compare the LDGAN method with the existing DCGAN and present the extensive simulation results. The extensive simulation result shows that the image generated by the proposed LDGAN including the estimation area is about 10% more.

Deep Reinforcement Learning (DRL) has emerged as a promising approach for solving complex tasks in optical networks. However, its black-box nature poses challenges for interpretability. For network operators, understanding the reasoning behind decisions is crucial for effective control and resource management. This paper addresses this gap by proposing a framework that generates explanations based on DRL agents’ decision-making processes. Using imitation learning, we train four classifiers to approximate a robust DRL agent designed for elastic optical networks. Our approach enhances explainability, enabling us to better understand and manage DRL-based decisions in optical network environments.

Explainable Artificial Intelligence (XAI) is crucial for the transition from the fourth to fifth industrial revolution, providing transparency and fostering user confidence in Artificial Intelligence (AI) powered systems. Since 2020, XAI applications demonstrate potential to transform manufacturing. This paper provides an extensive overview of XAI-based applications in Industries 4.0 and 5.0 by highlighting the trends regarding methods used, connecting XAI methods with important parameters and presenting XAI visualization approaches. The survey provides valuable insights for researchers, practitioners and industry leaders as it underscores the potential of XAI in shaping the future of manufacturing by enhancing transparency and user acceptance of AI-powered applications.

본 연구는 VFSS 시행 전, 환자의 기초 임상 정보와 기저 질환을 활용하여 Aspiration 발생 위험을 예측하는 머신러닝 모델을 개발하고, 설명 가능한 인공지능(XAI)인 SHAP을 통해 모델의 판단 근거를 정량적으로 분석하고자 하였다. 연하곤란으로 VFSS를 시행한 환자 500명의 후향적 데이터를 대상으로 성별, 연령, BMI, 뇌졸중, 파킨슨병, 식도암, 두경부암을 예측 변수로 설정하였으며, Logistic Regression, Random Forest, XGBoost 모델의 성능을 비교하였다. 분석 결과, XGBoost 모델이 정확도 0.824, 민감도 0.845, 특이도 0.810, F1 점수 0.798로 가장 우수한 성능을 보였다. 변수 중요도 분석에서는 뇌졸중, 파킨슨병, 두경부암 순으로 나타났으나, SHAP 분석에서는 뇌졸중에 이어 두경부암이 두 번째 주요 위험 요인으로 확인되었고, BMI 또한 낮은 중요도에도 불구하고 임상적으로 유의한 변수로 나타났다. 특히 저체중 상태에서 흡인 위험이 증가하는 비선형적 경향이 확인되었다. 이러한 결과는 XGBoost 기반 모델이 흡인 고위험군 선별에 효과적임을 시사하며, XAI를 통한 해석 가능성 확보로 영상 장비가 제한된 임상 환경에서도 신뢰할 수 있는 의사결정 지원 도구로 활용될 수 있음을 보여준다.

This study developed a machine learning model to predict aspiration risk prior to Video fluoroscopic Swallowing Study (VFSS) using basic clinical data and underlying disease information, and applied SHAP-based Explainable artificial intelligence (XAI) to interpret model decisions. A retrospective dataset of 500 patients with dysphagia was analyzed using Logistic Regression, Random Forest, and XGBoost models. XGBoost showed the best performance (accuracy 0.824, sensitivity 0.845, specificity 0.810, F1 score 0.798). Stroke was identified as the most influential predictor, followed by head and neck cancer and Parkinson’s disease. SHAP analysis further revealed that head and neck cancer had a greater impact than Parkinson’s disease, and that low BMI was a clinically meaningful factor. In particular, underweight status (BMI <18.5 kg/m²) was associated with a marked increase in aspiration risk. These findings suggest that the proposed model can effectively identify high-risk patients using non-imaging clinical data. The integration of XAI enhances interpretability, supporting its potential as a practical clinical decision support tool for early risk assessment and intervention.

Heart disease is a major cause of mortality in the world that is in dire need of accurate, interpretable predictive measures that could be utilized to manage it proactively. The writer of this paper proposes an Explainable AI (XAI) Ensemble Machine Learning model to predict heart disease using an 1,025 patient record dataset. To achieve methodological rigor and generalization, 5- Fold Stratified Cross-Validation (CV) was used to evaluate all models, such as LightGBM and Random Forest. LightGBM model was stable and better in performance as it had Mean CV Accuracy of ±0.9620 ±0.0178. Integration of XAI (SHAP/LIME) is the means of creating clinical trust; analysis has confirmed maximum heart rate (thalach) and type of chest pain (cp) as medically significant characteristics. This framework supports the sustainable smart city healthcare through a highly transparent decision-support system, which manages the resources in optimizing scalable public health programs.

The escalating importance of online reviews in customers' purchasing behavior has led to growing concerns about the prevalence of manipulated reviews, causing confusion in their decision-making process. However, despite several efforts to develop models for manipulated review detection, a critical aspect that has not been addressed is examining the connection between manipulated review writer personality traits and manipulated review detection. In this study, we examine the role of reviewer personality traits among the factors related to review characteristics in manipulated review detection using deep learning and explainable AI. The purpose of this research is to answer the question are reviewer personality matter in manipulated review detection and what features have an impact on detection. To rich our purpose, we apply deep learning twice: 1) to infer reviewers' personality traits from review text (CNN), 2) to create manipulated review detection models (RNN and CNN). Furthermore, we equip manipulated reviews detection model with one of the XAI techniques - SHAP, that generates visual explanations of the model's decision-making process, thereby enhancing the model's interpretability and transparency. We conduct our experiment by utilizing real-world review data from Yelp.com.

Sewer pipes are an essential public infrastructure of countries worldwide. They support wastewater transportation for processing or disposal. The harsh environments inside the sewer pipes can lead to the occurrence of various defects. Current crack detection approaches mainly focus on the surveillance camera (CCTV) to assess the condition of the sewer pipes. This process is considered a tiresome and laborious process. Therefore, a robust and efficient sewer defect detection system based on the transformer architecture is introduced in this manuscript. In addition, the system can provide explainable visualization for its predictions using the transformer's attention.

Artificial intelligence (AI) is a significant tool in modern military operations in that it helps to analyze a large volume of strategic, tactical, and operational data. On the other hand, current large language models (LLMs) like GPT-4 or Falcon have difficulty resolving problems in defense-specific contexts because of issues related to security, data confidentiality, and the lack of explainability. This document presents MilGPT, a secure and explainable LLM structure that aims at solving military problems only. To the model, fine-tuned open-source architectures with domain-specific defense datasets are integrated to elevate intelligence synthesis, decision-making, and threat prediction. On the benchmark, performance evaluation tasks show that MilGPT accounts for a 27% increase in contextual accuracy, an 18% reduction in hallucination rate, and an 33% improvement in explainability as measured by gradient-based feature attribution. In the proposed framework, military intelligence systems are not only secured but also made adaptive and humaninterpretable, thus, setting up a basis for the coming generation of AI models capable of defense-grade tasks.

Despite the widespread use of artificial intelligence (AI) in mobile healthcare apps, the need for more transparency in AI algorithms hinders their effectiveness by preventing users from understanding the reasons behind AI-based information provision. To address this challenge, various types of explainable AI (XAI) are adopted to offer transparent explanations of AI. Despite significant debates surrounding AI intervention, limited research has been devoted to whether and how various XAI types affect user behavior differently. In this study, we conducted a randomized field experiment to investigate the effectiveness of three XAI algorithms: 1) feature importance, 2) feature attribution, and 3) counterfactual explanation in promoting users' health behavior. Drawing on the self-regulated learning theory, we expect that XAI focusing on counterfactual explanation increases strategic planning and outcome expectancy, resulting in better self-regulation behavior. Our findings indicate that counterfactual explanation significantly improves users' action planning behavior, leading to a 16.5% increase in workout duration and a 3.49% increase in health records compared to the control group. Our results are salient for users with a high level of AI susceptibility due to age, goal weight loss, and AI outcome. Our finding sheds light on the potential of algorithmic explanations to improve the effectiveness of AI interventions in the healthcare industry, with practical implications for designing more transparent and user-friendly healthcare apps.

[연구목적] 본 연구는 온라인 리뷰 데이터를 기반으로 직원의 직무만족도를 예측하기 위한 새로운 설명가 능한 인공지능(XAI) 프레임워크를 제안한다. 본 연구의 참신성은 토픽 수준의 의미 정보(Latent Dirichlet Allocation(LDA))와 대형언어모델(LLM) 기반 언어 다양성 지표를 통합하여 직원 서술의 복잡성이 잠재된 직 무만족도를 어떻게 반영하는지를 규명하는 데 있다. 이러한 접근은 전통적인 피처 엔지니어링을 넘어, 미묘한 언어·심리적 패턴을 포착할 수 있도록 한다. [연구방법] 총 120개 기업의 Glassdoor 리뷰 141,854건을 수집하여 데이터 수집, 전처리, 모델링, 평가의 네 단계를 거쳐 분석하였다. 평점 4–5는 만족(1), 3 이하의 평점은 불만족(0)으로 구분하였다. LDA 토픽모델링을 통해 장점(pros)과 단점(cons) 섹션에서 각각 세 가지 주제를 도출하였으며, Sentence-BERT와 MPNet 임베딩 을 활용하여 cons_entropy, cons_sem_div, cons_pos_ttr, cons_mtld 등 언어 다양성 지표를 산출하였다. Synthetic Minority Over-sampling Technique(SMOTE)를 적용하여 Random Forest와 XGBoost 모델을 학습 하였으며, 해석력을 확보하기 위해 SHapley Additive exPlanations(SHAP) 분석을 수행하였다. [연구결과] Random Forest 모델이 가장 높은 예측 성능(정확도 및 F1 값 = 0.85)을 보였다. 핵심 예측 변수 에는 추천 여부, CEO 평가, 기업 전망뿐 아니라 복잡하거나 일관된 표현 패턴을 포착하는 언어 다양성 지표가 포함되었다. [연구의 시사점] 연구결과는 직무요구–자원(Job Demands-Resources, JD-R) 모형을 지지하며, 장시간 근 무와 갈등을 반영하는 언어는 만족도를 낮추고, 보상과 지원을 언급하는 긍정적 표현은 만족도를 높이는 것으 로 나타났다. 제안된 설명가능한 AI 프레임워크는 HR 분석의 투명성과 실무적 활용에 기여할 수 있는 이론적 및 실천적 시사점을 제공한다.

[Purpose] This study proposes a novel explainable AI framework designed to predict employee job satisfaction from online reviews. Explicitly highlighting the study’s novelty, the framework integrates topic-level semantics(Latent Dirichlet Allocation(LDA)) and LLM-based linguistic diversity indicators to reveal how employees’ narrative complexity reflects their underlying job satisfaction. This approach moves beyond traditional feature engineering to capture the subtle linguistic-psychological patterns. [Methodology] A total of 141,854 Glassdoor reviews from 120 firms were analyzed through four stages: data collection, preprocessing, modeling, and evaluation. Ratings of 4–5 were labeled as satisfied and 3 or below as dissatisfied. LDA topic modeling identified three themes from pros and cons sections, while linguistic diversity indicators such as cons_entropy, cons_sem_div, cons_pos_ttr, and cons_mtld were extracted using Sentence-BERT and MPNet embeddings. Random Forest and XGBoost models were trained with Synthetic Minority Over-sampling Technique(SMOTE)-based balancing, and SHapley Additive exPlanations(SHAP) analysis was applied for interpretability. [Findings] Random Forest achieved the best performance(accuracy and F1 = 0.85). Key predictors included employees’ company recommendations, CEO evaluations, business outlook perceptions, and linguistic diversity measures capturing complex or coherent expression patterns. [Implications] Results support the Job Demands-Resources(JD-R) model, showing that language reflecting workload and conflict lowers satisfaction, while positive references to pay and support enhance it. The explainable AI framework offers theoretical and practical insights for transparent HR analytics.

본 연구는 2015년부터 2024년까지의 한국 상장기업을 대상으로, 기계학습, 딥러닝, 그리고 설명가능한 인공지능을 활용하여 이익조정을 탐지하는 방법을 분석한다. 연구 결과, 모델과 과제 간의 전문화 현상이 확인되었다. 즉, 앙상블 소프트보팅은 발생액 기반 이익조정 예측에서 가장 우수한 성능을 보였으며, 순환 신경망은 실물활동 기반 이익조정 탐지에서 탁월한 성능을 보였다. 이는두유형의 이익조정이상이한이론적논리에기반하고있음을시사한다. 설명가능한인공지능 분석을 통해 본 연구는 이중 논리 프레임워크를 제시하며, 발생액 기반 이익조정과 실물활동 기반 이익조정을설명하는서로다른이론적메커니즘을규명하였다. 구체적으로, 발생액기반이익조정은 유동성위기에대한기회주의적생존대응으로나타났으며, 이는이자보상배율변수에의해주도되어 재무곤경 가설을 지지한다. 실물활동 기반 이익조정은 정치적 비용과 시장 가시성에 대한 전략적 대응으로 나타났으며, 이는 기업 규모(시가총액) 변수에 의해 가장 강력하게 설명되어 정치적 비용 가설을 지지한다. 이 프레임워크는 발생액 기반 이익조정과 실물활동 기반 이익조정의 내재된 이론적 논리를 동시에 밝혀냄으로써, 기존의 정확도와 해석가능성 간의 간극을 해소한다. 또한, 본연구의결과는감사인과규제당국이이론기반의감시및감독메커니즘을설계하는데실질적인 시사점을제공한다. 요약하면, 본 연구는모델성능과 설명-이론 정합성이라는 두가지연구질문에 대한 답을 제시한다.

This research explores how machine learning (ML), deep learning (DL), and explainable AI (XAI) can be applied to identify earnings management (EM) among Korean listed companies over the 2015–2024 period. We demonstrate model-task specialization: Ensemble SoftVoting best predicts AEM, while Recurrent Neural Networks (RNN) excel in REM detection, reflecting distinct underlying logics. Our XAI analysis introduces a dual-logic framework explaining distinct theoretical mechanisms behind AEM and REM. We find models independently learn two distinct theoretical logics: (1) AEM is identified as an Opportunistic response to Financial Distress and Liquidity Constraints (driven by Interest Coverage and Net Income, validating the Financial Distress Hypothesis); and (2) REM is identified as a Strategic response to Political Costs and Market Visibility (driven primarily by Firm Size and Profitability, validating the Political Cost Hypothesis). This framework bridges the long-standing accuracyinterpretability gap by uncovering distinct theoretical logics underlying AEM and REM. The findings provide actionable insights for auditors and regulators to design theory-grounded oversight mechanisms in high-stakes financial contexts. Specifically, this study answers two research questions concerning model performance and explainable-theory alignment.

The need to detect lung cancer accurately and early has been brought out through the fact that lung cancer has remained one of the top causes of cancer-related deaths in the world. A CT scan is usually interpreted manually which makes this time consuming and subjective. This paper hypothesizes that an explainable implementation of automated deep learning can be used to classify lung cancer based on the transfer learning models (VGG16, VGG19, ResNet50, and EfficientNetB3) on the IQ-OTH/NCCD dataset. The dataset was stratified by means of the SMOTE and was split into 80 and 20 percent training and validation subsets, respectively. All the pretrained CNNs were fine-tuned with the Adam optimizer and categorical cross-entropy loss. VGG16 performed the best and had a validation accuracy of 98.64, precision and recall of 98, and ROCAUC of 99. Visualization of tumor regions was done using explainable AI techniques (Grad-CAM, LIME), which are interpretable and have diagnostic transparency. The suggested framework proves that transfer learning combined with XAI is more effective in terms of accuracy and reliability in diagnosing medical images and is one of the steps to clinically reliable smart healthcare systems.