The increasing frequency of natural disasters and climate change events presents significant challenges for emergency response operations. Area monitoring for civil protection purposes is critical to safeguard resources, track and survey assets, and optimize decisionmaking. Coordination of resources is crucial in crisis situations where timely and accurate information is essential. Aviation traffic plays a key role in civil protection missions such as firefighting or search and rescue operations. Monitoring aviation and satellite data is of major importance and is one of the primary goals of this research. Real-time monitoring and alerting can be extremely useful to local authorities, NGOs, and the general public. In this paper, we introduce an online visualization tool that uses data aggregation/fusion to provide command capabilities and informative visualizations along with great functionality for real-time command and control operations. This system could be particularly beneficial for multiple stakeholders, including municipalities and the Ministry of Climate Crisis and Civil Protection.
Early detection of Coronary Artery Disease (CAD) and Non-Small Cell Lung Cancer (NSCLC) is crucial for improving patient outcomes. In this study, RGB-CNN (Convolutional Neural Network) was implemented, and trained from scratch using Polar Maps for CAD diagnosis and Computed Tomography (CT) images for NSCLC diagnosis. The CNN predictions were then integrated with clinical data into a Fuzzy Cognitive Map (FCM) classifier for each type of diagnosis. Nuclear medicine experts provided linguistic values in the form of fuzzy sets to define the relationships between input and output concepts, which were later converted into interval values. Extreme Learning Machine (ELM) and Genetic Algorithm (GA) were applied to the FCM learning process to refine the interconnections based on expert knowledge. To ensure the robustness of the results, 10-fold cross-validation was employed. The DeepFCM-ELM model demonstrated superior performance, achieving 80.4% similar to 4.97% accuracy for CAD diagnosis, and 91.9% similar to 3.07% for NSCLC diagnosis using CT images. Heatmaps were generated to interpret CNN predictions by highlighting pathological regions. These heatmaps were then used in GPT, along with DeepFCM weights, CNN, and DeepFCM prediction and input clinical values, employing Natural Language Generation to translate DeepFCM results into humanreadable language, enhancing the model's overall explainability. All these techniques have been integrated into a Medical Decision Support System (MDSS) designed to effectively manage both medical classification challenges.
Software accessibility reveals how easy it is to use an interactive system, regardless of disabilities and devices used. Website accessibility is particularly important due to the popularity of the Web. This is why there are web accessibility standards, such as the Web Content Accessibility Guidelines (WCAG), and technologies available that can help in developing accessible websites. WCAG includes many criteria and techniques that need to be followed to ensure website accessibility. As a result, software tools, often referred to as automated accessibility evaluation tools, have been developed to facilitate this rather demanding task. These tools do, however, report many potential accessibility issues that call for human judgment and thus still require a significant amount of time and effort. Do such potential issues reported by automated tools affect the results of guideline-based web accessibility evaluation? This paper attempts to provide an answer to this question in the context of web accessibility evaluation of Greek public hospitals. To this end, 112 Greek public hospital websites were evaluated against WCAG 2.1 using the TAW automated tool. Results found a significant effect of the manual inspection of potential issues on the accessibility results obtained. The only exception was when the data were analyzed per WCAG priority level where no significant effect was found for level AA. We also build linear regression models to estimate the number of actual accessibility issues based only on the ones that are automatically identified.
The explosive increase in sensory equipment for the gathering of data in recent years has pushed computing resources to their limits and has created the need for a more decentralized approach in Big Data Analytics. The term Edge Computing refers to this new paradigm where computation is performed away from the traditional cloud, in a more distributed manner and physically closer to the actual gathering of data. However, in large scale applications like an energy smart grid system spanning entire cities, there are many challenges that need to be faced for such a system to perform effectively, namely the continuous communication and exchange of information between heterogeneous edge devices, sensors and actuators, and the division of devices in spatial clusters that need to perform special operations depending on the characteristics of the area. The aim of this paper is to provide a small-scale conceptual architecture for the communication of clusters of edge devices in a smart grid system, using the actor model and the AKKA framework as a communication backbone.
This paper compares three of the most commonly used applicationlayer communication protocols, namely HTTP3, MQTT, and CoAP, for IoT applications. While HTTP3 is used as a baseline, MQTT and CoAP are optimized for resource-constrained environments and have thus found widespread use in IoT. This research work investigates the mechanisms, design features, and performance characteristics of each protocol and outlines their respective strengths and weaknesses. It will also carry out practical simulations to assess the utilization of processing resources under different communication models, such as publish-subscribe and request-response. The results of this work will be very useful in selecting an appropriate protocol for different IoT use cases with a trade-off between efficiency, security, and ease of implementation.
This paper addresses the challenge of producing fuel-optimal speed profiles for trains by transforming time-optimal profiles. In a timeoptimal speed profile, the train travels over a specific path with various speed limits in the shortest time. Our approach aims to adjust the time-optimal speed profile by extending its total duration to produce a fuel-optimal speed profile within that extended time frame. The proposed method greedily modifies parts of the given speed profile in a gradient descent manner, trying to maximize the ratio of fuel savings to time increase. Our analysis concluded that it is most effective to adjust either "Acceleration and Deceleration"or "Acceleration, Cruising and Deceleration" throttle sequences, with the former always being the best option.
This paper presents a novel approach to test case prioritization using Large Language Models (LLMs) for complex software systems. Traditional prioritization methods often struggle with the dynamic nature of modern software development and the large amounts of unstructured data generated during the software lifecycle. Our method leverages LLMs to analyze diverse data sources, including code changes, user feedback, and system documentation, creating a more adaptive and context-aware prioritization strategy. We applied our approach to an Internet of Things (IoT) based system for motion tracking in ten-pin bowling. The experimental results show significant improvements over a baseline Additional Statement Coverage method. Our LLM-enhanced approach achieved a 12.12% higher Average Percentage of Faults Detected (APFD) score and reduced test suite execution time by 26 %. These findings demonstrate the potential of LLMs to enhance software testing practices, particularly in early fault detection and efficient resource utilization. The paper discusses implementation details, evaluation metrics, and future directions for integrating this approach into continuous integration and deployment pipelines.
This work presents a novel classification model aimed at automatically categorizing user inquiries submitted to the helpdesk unit of the Greek School Network (GSN). Addressing the need for robust and accurate classification systems, a model architecture that combines a pre-trained XLM-Roberta (XLM-R) with Temporal Convolutional Network (TCN) and Transformer layers is proposed. The proposed model operates over binary-class structures and is evaluated on the GSN dataset, a dataset that includes labeled user queries. This is a first set of experimental results in an ongoing investigation on GSN data, where our proposed scheme is being compared to a total of seventeen classification schemes that include both traditional machine learning models and other transformer-based architectures. Results indicate that the LM-R TCN Transformer outperforms the alternatives, achieving an accuracy of 0.94 with similarly high precision, recall, and F1 scores, while also demonstrating robust inter-annotator agreement via Kappa and MCC scores of 0.82. Our findings suggest that the proposed architecture offers a strong and reliable solution for text classification tasks in the specific GSN context, particularly for our low-resource Greek language framework. The latter also highlights the broader applicability of transformer-based models in educational service domains.
Open data has become an important player in promoting government transparency and accountability around the globe. Governments have established policies to provide citizens or users with access to data. One way to obtain such data is through Freedom of Information Act (FOIA) requests. Considering users' requests for data and responding with the required information makes open data ecosystems user-driven. This paper explores freedom of information (FOI) requests received by Germany's Ministry of Health as a user-driven approach to open government data (OGD). By applying advanced natural language processing (NLP) techniques, specifically Latent Dirichlet Allocation (LDA) and BERTopic, we analyze a large corpus of FOI requests to identify key topics and trends in public inquiries. The findings reveal valuable insights into citizens' information demands and demonstrate the strengths of these NLP methods in extracting actionable patterns. In this way, governments can easily understand user needs in health-related datasets (this study, in particular, focuses on requests made by citizens related to COVID-19 data). This study lays the groundwork in two ways: first, by understanding and thematically categorizing user needs related to open data, and second, by improving government data transparency, informing the prioritization of dataset releases, and supporting evidence-based policymaking.
This paper provides a systematic review of hybrid algorithms combining Evolutionary Algorithms (EAs) with Machine Learning (ML) techniques, focusing on routing-related problems. It explores the integration of various ML methods, such as Reinforcement Learning (RL), Supervised Learning (SL), and Unsupervised Learning (UL), with EAs, aiming to address complex, multi-objective optimization challenges often encountered in logistics, robotics, and network routing. The review identifies RL-based methods, particularly Q-Learning (QL) and Deep RL (DRL), as the most prominent approaches due to their adaptability and capability to dynamically adjust solutions. SL techniques, including Decision Trees (DTs) and Artificial Neural Networks (ANNs), and UL approaches, such as clustering, also play significant roles in enhancing EAs' performance.
Artificially generated media are winning ground in the digital landscape due to their potential for dissemination. The speed at which synthetic media is created and disseminated in an ever-expanding digital landscape is alarming. Today almost everyone has access to digital tools for creating synthetic media, with which they can produce and disseminate their own version of reality, which may influence public opinion. After all, demographics are not the only factors that influence the rapid distribution of fake media, as it happens through the digital world. The latest developments in AI technology make it imperative to provide additional education in the field of digital media, as well as in the development of critical thinking. Given the social, ethical, and economic implications of digitally created media, their successful detection has introduced new areas of interest, such as the scientific research of digital media authentication. The distinction between authentic and artificial content is now of fundamental importance, even as a forensic analysis, since the authentication of digital data can require complex actions and the use of modern cutting-edge technologies.
Since fruit plays an important role in the human diet, pomiculture has a significant economic impact on agricultural countries worldwide. However, the lack of manual labor makes it necessary to explore alternative solutions for the harvesting process. Emerging technologies that are modernizing agriculture, including robotics, machine learning, and drones, provide a promising direction to address these types of challenges. More specifically, the use of automated harvesting robotic systems could mitigate labor shortages by operating more hours than humans and reducing the physical demands of traditional manual harvesting. The widely available, cost-effective electronics and sensors, along with the easily developed machine learning models, make it feasible to construct robotic systems for autonomous detection and picking of fruits. On the other hand, the deployment of robotic systems in orchards remains a challenge due to the dynamic and unpredictable nature of these environments, leaving room for further research. This paper aims to develop a machine learning architecture to identify oranges and orange trees in orchard environments, using both ground and UAV imagery. The proposed model is an integral part of a ground-based autonomous robotic harvesting system. This paper describes the methodology followed, discusses the preliminary results and proposes future work to contribute to the field of automated agricultural systems. Despite the medium-size dataset and the contribution of a user-friendly platform for developing models for edge devices, the results of the models are very encouraging, and through the evaluation the most suitable algorithms were selected for detecting trees upon entering an orchard and oranges at close range.
The Internet of Cloud (IoC) which is the synthesis of Cloud Computing (CC) and the Internet of Things (IoT), introduces significant security challenges and vulnerabilities. To address these challenges requires focused, gamified, and tailored cybersecurity training, as traditional methods lack the adaptability needed for IoC's unique vulnerabilities. In addition, this paper examines the role of AI Learning Management Systems (LMS) and gamification play in the cybersecurity training process. This paper presents how the combination of AI-based LMS and gamification can increase engagement, raise awareness, and support up skilling, potentially enabling users to address IoC-specific threats more effectively in the IoC. Furthermore, it seeks to demonstrate how the combination of existing and established frameworks can support and assist cybersecurity training for users at all levels, thereby contributing to the management of ongoing challenges within the IoC ecosystem.
In this paper, we study the problem of top-.. butterfly discovery in uncertain bipartite networks. Butterflies are important motifs on bipartite graphs, and their existence gives valuable information regarding the network's connectivity and robustness. We focus on uncertain (or probabilistic) graphs, where each edge is annotated with an existential probability. In addition to counting, the discovery of the top-.. most probable butterfly motifs in the.. top is an interesting and computationally expensive problem. Our approach serves as a foundation for future studies on uncertain bipartite networks, with possible extensions to the detection of other important network motifs beyond butterflies. Future work could focus on the adaptation of the algorithms to handle dynamic networks, the application to different types of graphs, the development of real-time analysis methods for continuous data streams and more.
Recent research in the field of mobile text entry aims to address the need for faster and more generalisable evaluation by introducing computational models that simulate human behavior. In this paper, we examine a state-of-the-art open-source simulator of text entry with virtual keyboards (RoboType), and present a comparison between empirical data of human text entry performance and the data generated under the same conditions by RoboType. We find that while the simulator does indeed generate realistic behaviours, metrics derived from it deviate from real data on a like-for-like comparison basis. We conclude by reflecting on potential corrective actions to better approach true human typing behavior.
Fire, a destructive and often life-threatening force of nature, has been a constant companion of humanity throughout its history. The ability to detect fires swiftly and accurately is paramount in mitigating their catastrophic impact. In this paper, we've designed and implemented a smart forest fire monitoring and detection system using microservices and container-based virtualization, having as its primary aim the ability to detect the forest fire as early as possible. The efficient integration of the proposed approach in large-scale distributed environments, exploiting the advantages offered by the above cutting-edge technologies as well as the use of modern orchestrators, is also discussed and documented.
In the fast-paced domain of cybersecurity, specifically regarding the area of Windows authentication frameworks, we introduce new research on vulnerabilities in the protection of Windows Hello PIN and Trusted Platform Module executions. Based on a modest but rich record of previously known information and specialized analysis, this instrument is a significant step towards comprehension and testing of Windows-related authentication security. The primary aims, consequently, are to develop a tool that detects the occurrence of TPM on devices in the network and gains the status of TPM or to extract and pre-hash Windows Hello PINs to forcefully test them, particularly in environments where the TPM is absent or disabled. This represents an important continued scholarly line of work in the pursuit of building upon existing literature and offering on-the-ground, real-world chances that have concrete threats to their security presumptions surrounding Windows Hello PIN and TPM. Furthermore, this paper intends to furnish a foundational basis for future work, such as considering threat actors' possible courses of action and optimizing protective cybersecurity abilities.
The optimization of complex real-world systems often presents a challenge when explicit derivatives of the objective function are unavailable. In this paper, we address a real-world black-box optimization problem arising from the design and operation of Concentrated Solar Power (CSP) systems. CSP systems present a unique challenge due to their nonlinear, multi-modal nature, which complicates optimization using traditional gradient-based methods. Derivative-free optimization (DFO) techniques are well-suited to tackle such blackbox problems, but even these methods can become impractical when the number of function evaluations required is too large since the evaluation can be expensive. To overcome this limitation, we utilize an adaptive sampling DFO approach that requires a smaller number of function evaluations by intelligently selecting informative points based on surrogate models. The surrogate models are then optimized using derivative-based optimization algorithms to find new sampling points that may be (near-) optimal. We apply this methodology to optimize a CSP problem from the SOLAR benchmark tool, specifically focusing on minimizing the cost of thermal storage, and the total investment cost. The results demonstrate that our adaptive sampling method, ADASNOBFIT, outperforms the well-known SNOBFIT algorithm in terms of solution quality.
Shifting from traditional desktop 3D environments to extended reality (XR) presents several challenges. These challenges are more intense in applications that require audio, visual graphics and navigation controls, such as cultural heritage and virtual museum applications. While many studies focus on the general features of XR, they often fail to address the specific difficulties involved in adapting desktop applications for XR experiences. This paper explores the challenges of transitioning a virtual museum application developed for a desktop 3D environment to an XR platform. We performed a thematic analysis to identify and classify the main challenges into eight categories: development, content adaptation, accessibility, user experience, cognitive load and fatigue, technical challenges, cross-platform compatibility, and data privacy.
The joining of hardware speed-up for machine learning (ML) in small devices shows a very important step in technology growth. This ability to run complicated ML models on low-power systems opens up many uses, like Internet of Things (IoT) gadgets, robots and different edge computing tasks. This paper studies the present condition of hardware speed-up for ML in small devices, clarifies difficulties and looks at future paths for progress.
