This paper presents a controlled comparative evaluation of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer models for short-term photovoltaic power forecasting in an IoT-oriented monitoring scenario. Environmental variables acquired by IoT sensing devices, together with recent photovoltaic power history, were used as inputs for near-future prediction. The experimental study was performed on the SOLETE dataset using multiple input window lengths and direct forecasting horizons under a unified preprocessing and evaluation pipeline. The preprocessing procedure included chronological ordering, physically motivated feature engineering, cyclical temporal encoding, leakage-free normalization, and sliding-window sequence generation. The results showed that recurrent architectures outperformed the Transformer-based model in most evaluated settings. GRU achieved the best performance in 17 out of 20 configurations, with the overall best result obtained for W = 48 and h = 1 , reaching RMSE = 0.4777, MAE = 0.2297, and R^2 = 0.9155 . These findings indicate that compact recurrent architectures remain highly suitable for short-term photovoltaic forecasting from IoT-oriented environmental measurements and provide a strong basis for deployment-aware intelligent energy monitoring.
The growing adoption of large language model (LLM)-based agents introduces new security risks, including prompt injection and jailbreak attacks. This paper presents a unified empirical evaluation of gatekeeper-level defenses under controlled experimental conditions. Two pre-inference input filtering approaches are compared: an LLM-based classifier (ShieldGemma) and a rule-based framework (Guardrails AI), using standard benchmarks (AdvBench and TruthfulQA). The results show that both approaches can substantially reduce attack success rates, from 85
The 5G mobile security architecture introduced a new scheme for subscriber identity privacy. This “Protection schemes for concealing the subscription permanent identifier” is based on using the Elliptic Curve Integrated Encryption System (ECIES) cryptographic mechanism. This provides an elegant method for allowing the subscriber device to conceal the Subscriber Permanent Identifier (SUPI) while still being able to present itself to the network. The Subscriber Concealed Identifier is known as SUCI. However, ECIES, is not quantum-safe, and this makes the SUCI scheme vulnerable to a Cryptographically Relevant Quantum Computer (CRQC). This means that the present scheme is not future proof. In this paper we investigates this problem, and present a new SUCI protection profile which could be accommodated within the existing scheme, thus making it future proof and eligible for the coming 6G system architecture.
Small water utilities are increasingly digitizing their operations to gain efficiency, but this development also exposes them to growing cyber-physical security risks, including hybrid threats potentially conducted by nation-state actors. Existing research and frameworks are often inaccessible to small utilities with limited resources and technical literacy. This study applies an Integrated Action Research (IAR) approach to co-develop and evaluate a sociotechnical framework aimed at enhancing operational resilience for a small Finnish water utility. By utilizing a set of custom workshop canvases to understand daily realities, the findings indicate that the most acute risks relate to human capital, technical continuity, and unclear communication chains. To provide immediate, actionable guidance, an A7-sized physical handbook was introduced to assist operators and the utility manager in navigating various cyber and hybrid threat scenarios. This proof-of-concept demonstrates a scalable model for other critical infrastructure sectors globally.
Accurately predicting short-term traffic flow is one of the key issues in smart city management. However, relying solely on deep neural networks is still insufficient to fully overcome the high variability inherent in traffic data. If the noise in the data is not effectively addressed, the model may misinterpret the data structure, thereby affecting prediction accuracy and stability. Variational Mode Decomposition (VMD) can extract stable periodic signals from time series data, reducing the impact of noise. Although a few existing traffic flow prediction studies have utilized VMD, they have decomposed the entire dataset, including test data, during the preprocessing phase, leading to data leakage issue. To address this issue, this study proposes a rolling VMD technique and integrates the technique along with multi-dimensional temporal features to a deep learning-based prediction model. Experimental results on a publicly available traffic dataset show that the proposed integrated framework exhibits robust error suppression capabilities, significantly enhancing the overall prediction quality and stability of the model. More importantly, the proposed rolling VMD successfully addresses the data leakage issue. This effectively improves the practical applicability of the model in real-world scenarios.
In this paper, we study long-term stability of a distributed transaction processing system under adversarial transaction arrivals and node failures. We model a system of n nodes and m shared objects where an adversary injects transactions at a constrained rate, and up to f≤ n/3 nodes may experience crash-faults. We define stability as the guarantee that every transaction is executed within a finite time bound, preventing unbounded backlog growth. We first establish a fundamental limit, proving that no scheduler can maintain stability if the injection rate exceeds min{1/η _m,k, k(n-f)/m} . Below this bound, we design a fault-tolerant global-aware scheduler that ensures both stability and atomicity. By providing a formal bound on the adversarial injection rate, we demonstrate that our scheduler remains resilient against resource-exhaustion attacks that seek to destabilize the network. Experimental evaluations validate our theoretical bounds and demonstrate the effectiveness of the scheduler in handling adversarial bursts and node crashes.
Age-related macular degeneration (AMD) is the main cause of severe central vision loss in adults over 55, impairing daily activities and quality of life. Advanced forms include neovascular (wet) and atrophic (dry) AMD, the latter lacking effective treatments. Risk factors include age, smoking, genetic predispositions (CFH, ARMS2), nutritional deficiencies, and higher prevalence among women due to longevity, biological, hormonal, and sociocultural factors. Diagnosis relies on clinical examination and advanced imaging, while intravitreal anti-VEGF injections remain the primary treatment for neovascular AMD, reducing disability without offering a cure. In Spain and Europe, where over 17
The widespread use of mobile devices has made them increasingly attractive targets for cyberattacks, considering the sensitive and private information that we currently store everday on our devices. Considering the weaknesses of the current antimalware signature-based, we propose a novel privacy-preserving and explainable deep learning approach for the detection and the localisation of malicious payload of Android malware. By considering Federated Learning, the proposed method transforms opcode sequences extracted from Android applications into images, enabling the adoption of image classification without compromising user data. Model explainability is achieved through Gradient-weighted Class Activation Mapping algorithm, aimed to highlight the opcode regions most influential in the predictions, thus providing candidate payload localisation. Experiments conducted on real-world Android applications demonstrate interesting detection performance, with explainability showing the list of opcode that are related to the malicious payload from the model point of view.
Digital marketplaces in the gig economy require sophisticated human-centered design strategies to support usability, trust, and inclusive interaction across diverse user groups. This study uses a freelancing platform as a teaching case study through which university students applied mixed-methods UX/HCI research techniques, including a human-centered design process that integrates analyses of people, activity, context, and technology; heuristic evaluation; task-based usability testing; the System Usability Scale; and the User Experience Questionnaire. It also provided university students with hands-on UX/HCI pedagogy in research methodologies. Twelve university students learning UX/HCI methods completed controlled interaction sessions with original and redesigned prototypes of a freelancing platform across service search, gig creation, messaging, and profile management tasks. The redesigned interface improved overall usability and experience quality, with System Usability Scale scores increasing from 58.33 to 66.25 and attractiveness reaching 1.86 on the User Experience Questionnaire. Task performance also improved, particularly in AI-assisted search and basic navigation, while expert review identified a 77
We study a compact response-diagnosis task for short algebra answers in digital practice settings. Instead of modeling a complete tutoring or feedback pipeline, we focus on assigning a learner response to one of four categories—Correct, Partially Correct, Variable Slip, and Incorrect. Our approach combines simple string-based and symbolic features with lightweight classifiers. Using a manually annotated dataset of student responses, we evaluate whether these inexpensive and interpretable signals are sufficient for a practical first-pass diagnosis layer. The results show that standard models such as logistic regression and random forests perform well on this task while remaining transparent and easy to deploy. We frame the paper as a compact empirical study and discuss both its practical usefulness and its limits for broader educational deployment.
Capture-the-Flag (CTF) games are widely used as experiential and challenge-based learning approaches in cybersecurity education. Despite their growing adoption, limited empirical evidence exists on how students’ expectations of educational game features align with their actual experiences in CTF learning environments. This paper investigates students’ perceptions of instructional and game-based features by comparing their general expectations for educational games with their experiences in a specific CTF designed for cybersecurity learning. Using pre- and post-intervention survey design, data were collected from 51 students to analyze differences between expectations and experiences. Results indicate that the real-world application of theoretical knowledge was consistently rated as the most valuable feature in both expectation and post-intervention evaluations in the engagement and learning effectiveness domains. In contrast, entertainment value, progress tracking, and interactive elements were rated lower than initially expected. Challenges and puzzles were perceived as more effective than points or rewards in supporting conceptual understanding of cybersecurity knowledge. This study identifies key mismatches between expected and experienced CTF features, highlighting the importance of meaningful challenges and authentic task design over traditional gamification elements. The findings provide practical insights for designing more effective cybersecurity learning environments and should be interpreted as evidence of perceived usefulness and mismatch rather than direct measures of instructional effectiveness.
Healthcare processes involve complex interactions and temporal constraints that are not adequately captured by traditional process mining techniques based on low-level event logs. In this paper, we propose a semantic-temporal framework that integrates ontology-based event enrichment with formal temporal verification. Events may include RDF representations grounded in ontologies such as SNOMED CT, enabling semantic conditions to be evaluated as propositions within a temporal logic framework. The approach supports automated verification of data-aware temporal constraints and detects clinically significant violations that remain unnoticed by conventional methods. A healthcare case study illustrates its applicability.
Machine learning models can predict cardiovascular disease with high accuracy, yet their clinical adoption remains limited by a lack of interpretability. Most existing studies apply a single explainability technique to one model, making it difficult to assess whether the resulting explanations are method-dependent or genuinely reflective of learned patterns. This paper addresses that gap through a structured multi-model, multi-method explainability comparison. Four classifiers (Random Forest, XGBoost, SVM, and Multi-Layer Perceptron) are trained on the Cleveland Heart Disease dataset and explained using both SHAP and LIME at global and local granularities. To assess explanation reliability, SHAP-based feature importance is statistically compared with traditional Gini importance via Spearman rank correlation (r = 0.819, p < 0.001). The key finding is that the top predictors, number of major vessels (ca), chest pain type (cp), and thalassemia status (thal), are consistent across models and explainability methods, suggesting these rankings reflect genuine data patterns rather than artifacts of any single technique. This study constitutes a proof-of-method investigation for explanation consistency rather than a clinically deployable system. The results provide an empirical basis for selecting and cross-validating explainability methods in tabular clinical prediction tasks.
Infrared (IR) thermography is a challenging method in brain surgery. The main difference between benign and malignant tumours depends mainly on the metabolic rate and vascularisation of tumours. Malignant tumours, such as medulloblastomas, have elevated metabolic rates that lead to high blood flow. The increase in blood flow causes an increase in the temperatures of tumour cells. On the other hand, benign tumours, like adenomas, have depressed metabolic rates, which lead to low temperatures. In this study, real-time IR thermography using a Uncooled FPA microbolometer IR camera was used to differentiate amongst different types of brain tumours on the basis of their intraoperative temperatures. Ten random patients with different ages and benign or malignant tumours were tested intraoperatively. The IR images of tumour sites after removing the cortex and biopsies after extraction from the brain were acquired. These images were compared with the magnetic resonance imaging or positron emission tomography images of each case to confirm whether IR thermography provided the same results or not. Three neurology surgeons confirmed the results. Malignant tumours had an average temperature of 37.4 ℃ (36 ℃–40 ℃), whereas benign tumours had an average temperature of 33.8 ℃ (32 ℃–35 ℃). Necrotic cells were proven to have lower temperatures than active tumours because of the loss of blood vessels and reduced metabolic rates. IR cameras can be used intraoperatively as noninvasive, portable and cheap technique to help surgeons identify the margins of different types of brain tumours in real time. They can help surgeons make conclusive decisions during surgeries if an extracted tumour is benign or malignant without the need to wait for postoperative biopsy cultures.
Developing robust machine learning models for rare clinical conditions is challenging due to limited sample sizes, heterogeneous data modalities, and strict privacy regulations that restrict cross-institutional data sharing. This study presents a federated multimodal learning framework that enables multiple hospitals to collaboratively perform disease classification without sharing raw patient data. Each institution, represented as a client, trains a local model using a single modality (structural MRI, SNP-based genomics, or mRNA transcriptomic profiles). The resulting predictive signals are then used to train a global model under federated coordination, enabling cross-institutional collaboration while preserving data privacy. The proposed approach is evaluated on the aligned ADNI cohort and achieves balanced performance across all modalities (accuracy ≈ 0.83 ; F1-score ≈ 0.82 ), matching or outperforming baseline models. In particular, transcriptomic accuracy improves from approximately 0.60 under centralized training to 0.83 in the federated multimodal setting, representing a relative improvement of about 25
This paper presents an AI-driven immersive Virtual Reality (VR) system for Olympic shooting disciplines (pistol, rifle, and shotgun), designed to support dual-task cognitive-motor training in individuals with disabilities. The system integrates static precision tasks and dynamic target-tracking scenarios within a unified and adaptive immersive environment, supported by a virtual coach based on marker-less motion capture (Captury Live). The system was evaluated through a two-week feasibility study involving 66 participants with cognitive disabilities in inclusive educational programs. The primary objective was to assess usability, accessibility, and tolerability. Results showed excellent usability (SUS: 79.4 ± 8.6), high engagement, and low incidence of cybersickness symptoms. Observed short-term trends suggested potential improvements in postural stability, coordination, and attentional engagement. However, these findings are exploratory and not derived from controlled experimental comparisons. Overall, the results support the feasibility of integrating immersive VR, AI-driven feedback, and marker-less motion capture into an inclusive training framework, motivating future longitudinal and controlled studies.
The teaching of chemistry often relies on two-dimensional representations that limit students’ spatial understanding of atomic and molecular structures. This study proposes and validates an augmented reality system with tangible interaction designed to support the comprehension of abstract chemistry concepts such as atomic structure, the periodic table, and compound formation, targeted at high school students. To address this challenge, an educational system integrating augmented reality and physical manipulation of objects was developed. The project followed the DIPF (Design, Implementation, Testing, and Operation) methodology. As part of the design, 3D-printed cubes with raised patterns were created to function as recognition markers, while chemical elements were modeled in three dimensions. A mobile augmented reality application was implemented to enable visualization and interaction with these elements. The system was evaluated in real educational environments with high school students and chemistry teachers, focusing on both technical functionality and user acceptance. The proposed approach introduces a tangible interaction model based on augmented reality for chemistry education, combining physical relief markers, three-dimensional visualization, and molecular fusion. This integration expands the technological possibilities for learning abstract scientific concepts in secondary education. Results demonstrate stable visualization of atomic models, reliable recognition of physical markers, and clear representation of simple molecular combinations. Additionally, increased student motivation and participation were observed during practical activities. The integration of immersive and interactive technologies in science education can enhance the quality of learning, stimulate interest in STEM disciplines, and contribute to reducing gaps in access to innovative educational resources.
The reliability of healthcare decisions derived from large-scale medical studies depends strictly on the structural integrity of the underlying N-size datasets. Recently, a prominent cohort study reported statistical associations suggesting increased clinical risks; however, a data science audit using fundamental demographic and incidence metrics reveals a pronounced external validity discrepancy. Specifically, the studied cohort exhibited a 26
Objective: This study aims to critically evaluate the current state of machine learning (ML) and deep learning (DL) approaches for automated malaria diagnosis using microscopy images, with a particular focus on methodological rigor, validation practices, and clinical applicability. Results: A total of 173 studies were analyzed. Convolutional neural networks (CNNs) dominated the field, appearing in 65.3
Heart disease remains one of the leading causes of death worldwide, placing a significant burden on healthcare systems globally. Early and accurate identification of patients at risk is critical, yet access to specialised diagnostic tools is not always available across all clinical settings. This study focuses on developing and evaluating a machine learning pipeline to predict the presence of heart disease from routine clinical measurements. A merged version of the UCI Heart Disease dataset, combining records from the Cleveland, Statlog, Switzerland, Long Beach VA, and Hungarian cohorts, was used for model development. Four classification algorithms were evaluated under identical preprocessing conditions: logistic regression, random forest, gradient boosting, and support vector classification. All models were assessed using an 80/20 stratified train-test split and 5-fold stratified cross-validation to ensure reliable and unbiased performance estimates. The support vector classifier achieved the highest test ROC-AUC of 0.9412 and a recall of 0.9118, which demonstrate a strong ability to correctly identify patients with heart disease. The final trained model was deployed as a publicly accessible web application that accepts patient inputs and returns a real-time risk prediction. This tool is intended to support early screening efforts, particularly in settings where specialist resources are limited, and demonstrates that machine learning can provide meaningful predictive performance from basic clinical data. It should be noted that this study is limited to internal validation on a single merged dataset; clinical readiness is outside the scope of the current work.
