Access control in cloud-based computing environments is in essence a security mechanism enabling decisions as to which subjects may access which protected resources and under what operations. The rapid emergence of quantum computing threatens the hardness assumptions underlying classical digital signatures and public-key infrastructures. The authors of this work recently proposed a novel quantum ice-cream model to demonstrate the basic concepts of quantum cryptography in a simple and intuitive manner. In this work we present a novel quantum ice-cream digital signature scheme by using our proposed quantum ice-cream model. Our quantum ice-cream digital signature scheme maps two non-commuting measurement bases to the fl'avor'and 'shape'attributes of the quantum ice-cream model. It achieves information-theoretic security against forgery and repudiation via threshold tests and natively supports signer-verifier-arbiter workflow that is suitable for cloud access control. Although our scheme is presented theoretically, a rigorous mathematical analysis has been carried out using Hoeffding-type inequalities. We also discuss the implementation of our proposed scheme by using weak-coherent BB84-like optics with decoy states. This article demonstrates that our proposed quantum ice-cream digital signature scheme preserves the practicality of memoryless quantum digital signature schemes while aligning with cloud authorization requirements. Furthermore, we draw analogies between the design of our proposed QIDS and robust control theory to provide an interdisciplinary perspective on system stability and security.
Wireless Sensor Networks (WSNs) are widely applied in domains such as environmental monitoring, healthcare, and defense, where accurate localization of sensor nodes is essential for reliable data interpretation. Despite extensive research, localization continues to face challenges related to energy consumption, scalability, and algorithmic efficiency. This paper presents a systematic review of position computation algorithms, emphasizing how different approaches address these challenges. The analysis highlights key trade-offs, such as accuracy versus energy cost and convergence speed versus computational overhead. It also identifies clear research trends, including a shift from traditional range-based and range-free methods toward hybrid and AI-assisted techniques that enhance adaptability in dynamic environments. Persistent open issues remain, notably handling mobility, achieving robustness under noisy conditions, and enabling deployment in large-scale real-world settings. By synthesizing these insights, the study clarifies the current state of WSN localization, outlines unresolved gaps, and points to future directions for energy-aware, scalable, and learning-driven solutions.
This bibliometric study analyzes the evolution of research on digital actors from 1995 to 2023 using data from the Scopus and Web of Science databases. Through a systematic and rigorous selection process, the study identifies publication growth patterns, leading authors and journals, geographical contributions, and the thematic evolution of the field. Unlike previous bibliometric reviews focused broadly on digital twins or virtual avatars, this study uniquely examines the intersection between digital twin technologies and performance-oriented digital actors, integrating perspectives from engineering, artificial intelligence, and digital character creation. The findings delineate an emerging and previously underexplored research subfield and provide a structured framework that highlights conceptual trends, research gaps, and directions for future studies.
The rapid integration of generative artificial intelligence into higher education has sparked contrasting reactions, ranging from technological enthusiasm to ethical concerns. This study aims to qualitatively and interpretatively examine higher education students' attitudes towards using ChatGPT in academic settings. Based on semi-structured interviews with a diverse sample of students and an in-depth semantic analysis of the transcripts, fourteen key variables were identified and grouped into three overarching dimensions: techno-cognitive, academic-pragmatic, and axiological-ethical. The findings reveal ambivalent attitudes shaped by the lack of institutional regulation, performance-driven logic, and individual moral frameworks. This research proposes a conceptual model that expands traditional technology acceptance frameworks by incorporating normative, emotional, and identity-based dimensions. Beyond its empirical contribution, the study underscores the urgency of implementing a critical and ethically grounded pedagogical framework for using AI in higher education.
Large Language Models (LLMs) have accelerated code generation, yet their security implications remain underexplored. This work evaluates the capability of eight general-purpose and code-specialized LLMs to detect weaknesses and vulnerabilities in generated code using CVE and CWE references. Stratified sampling of the CVEfixes dataset across five programming languages is assessed using precision, recall, F1-score, and relevance-oriented metrics (REINP, REINR, and REINF1). Results show limited detection of specific vulnerabilities (CVE), but better performance for general weaknesses (CWE), especially in Python and Ruby. We discuss limitations in model training and prompt conditioning, and outline improvements through dataset diversification, prompt engineering and hybrid human-in-the-loop approaches. The study highlights the current potential and limitations of LLMs for practical vulnerability detection in generated code.
This work explores the intersection of artificial intelligence (AI) and the sustainable development goals (SDGs) in Africa, focusing on public interest in AI and SDGs in addressing the continent's unique we analyse AI and SDGs search interest across African countries over a nine-year period from September 2015 to August 2024. We investigate relative AI adoption indices for solving the SDGs, the current adoption readiness rankings, and the differences in regional trends across Africa. We employ Croston's intermittent time series trend forecast estimation method, agglomerative hierarchical clustering on principal components, and visual analytics to extract empirical insights from monthly AI- and SDGs-related relative GTE search interest across African regions. The study highlights regional disparities in search trends, with the Southern, Western, Eastern, Northern, and Central African regions ranking first through fifth on their current adoption readiness rankings for solving SDGs with AI. The current differences in regional trends across the African regions are multi-dimensional. They can be better simplified by supplementing the GTE search data with additional indicators and economic drivers such as GDP per capita, digital literacy rates, internet penetration, and national policy frameworks on AI and sustainable development in future analyses.
Cybersecurity is an essential topic in society due to the increasing application of technology across different application industries that are connected to the Internet. There are several methods to address cybersecurity challenges, such as Intrusion Detection Systems (IDS). This article presents new methods based on Federated Learning that can be applied to IDS to improve cybersecurity attack detection. Researchers use different approaches to create IDS solutions, including machine learning algorithms and techniques such as Federated Learning (FL) to classify network traffic as normal or malicious. FL is an emerging technology and is expected to benefit cybersecurity by improving attack detection and threat identification. It consists of a set of clients, each one with local data, whose goal is to collaboratively train one or more global models without centralizing the data, through several iterations known as rounds. An important characteristic to consider in cybersecurity data for machine learning algorithms is the presence of non-Independent and Identically Distributed (non-IID) and imbalanced data. Non-IID data describes FL datasets in which data is not evenly distributed between clients. Unlike previously proposed methods in other works, this article proposes approaches that focus on non-IID data while using FL to create models without requiring multiple training rounds. The FLENV and FLEWNV methods are proposed, implementing a federated learning framework that uses a single training round and aggregates clients through ensemble learning with normalized and weighted voting, focusing on non-IID and imbalanced data. The well-known Ton-IoT and BotIoT datasets were analyzed to evaluate the proposed methods and other possibilities commonly used in the state of the art. The approaches outperformed other methods from the literature in the experiments, indicating that leveraging knowledge of non-IID and imbalanced data in cybersecurity can lead to improved attack classification frameworks.
Tracking sea turtle migration is hindered by noisy and incomplete geolocation data, as well as irregular sensor transmission. These limitations make it challenging to model trajectories and accurately interpret ecological patterns. This study presents a predictive framework for modelling the trajectories of green turtles (Chelonia mydas) using satellite telemetry and artificial intelligence techniques. Georeferenced data from SPOT-375B tags were pre-processed to address noise, data gaps, and spatial anomalies. A Long Short-Term Memory (LSTM) neural network was trained with normalized time series data to forecast future positions, capturing the temporal dependencies of turtle movement. A Kalman filter was applied post-prediction to enhance trajectory continuity and reduce uncertainty through recursive state estimation. Experimental results show that the approach yields an average MAE of 0.0986, MSE of 0.0307, and RMSE of 0.1288, and reduces mean prediction error by 43.75 % relative to a recurrent neural network (RNN) baseline while requiring similar to 36 % of its CPU time. This integrated pipeline enhances the reliability of wildlife trajectory forecasting and provides a scalable solution for ecological tracking under uncertain detection conditions, facilitating a deeper understanding of species behavior and more effective conservation strategies.
This study presents an intelligent system for dynamic urban traffic management, focused on the automatic adjustment of speed limits and the activation of lanes as parking areas to optimize urban traffic. Using realistic scenarios simulated in the SUMO (Simulation of Urban MObility) platform and controlled by a neural network implemented in Python, the system effectively responds to variable conditions such as vehicle density, traffic flow, and other factors. To evaluate the achieved optimization, the results are compared using various traffic performance indicators, including congestion levels, average speed, travel time, among others. Through adaptive real-time behavior, the system achieves greater traffic fluidity, reduced congestion, and better use of available infrastructure. The designed control system significantly improves performance compared to uncontrolled scenarios: on average, it reduces travel time by 16 %, CO2 emissions by 13 %, waiting time by 12 %, and accident probability by 15 %, while traffic flow increases by 11 %. The results show that this approach can effectively complement traditional traffic light control and adapt to diverse urban contexts, positioning itself as a promising solution to improve mobility in modern cities.
Initial population in a swarm intelligence-based optimization algorithm plays an important role in avoiding local optima, early convergence, and global space exploration. If initial values provide proper coverage of the search space, then it may quickly lead to the optimal solution, whereas limited coverage may stick an algorithm into local optima. Several techniques are presented by researchers to properly initialize the population instead of random numbers. Chaos theory is a well-known concept utilized in designing an improved variant of various existing nature-inspired optimization algorithms. However, its impact on the initial population of a metaheuristic was never thoroughly studied before. This study investigates the impact of ten chaotic maps on the initial population of a sailfish optimizer algorithm (SFO). SFO is a recent metaheuristic algorithm which mimics the hunting behavior of sailfish. Ten chaotic maps have been used to initialize sailfish and sardine populations. All techniques (with or without chaotic maps) are tested with twenty-three classical benchmark functions. The numerical results, ranking and statistical analysis show that chaotic maps are useful in avoiding local optima and are effective for improving the global search capabilities of a metaheuristic.
Nowadays, deep learning models are quickly increasing in complexity and size, making the training or inference phases unsuitable for most office computers or even high-performance computing servers. Examples of such models include, but are not limited to, large language models and image diffusion models, where the number of parameters to adjust can reach hundreds of billions. Although deep learning models are extremely useful tools, the need for powerful computers to train and use them might compromise the obtained results and also reduce accessibility for many researchers. In this work, we propose a general method to reduce the training time and computing resources (central processing unit and memory usage) by the division of a large-size model into several smaller submodels that are easier to handle but do not affect the final performance. This division depends on the original model and architecture, and hence we propose specific strategies for regression and classification problems. The main result of this work is the development of a public Python library called Skynnet to offer this alternative to deep learning computations.
Mental health disorders present a growing global concern, yet predictive models often overlook age-specific variations in risk factors. This study introduces an explainable artificial intelligence (XAI) framework for age-stratified mental health risk prediction using Shapley Additive Explanations (SHAP). Using the Open Sourcing Mental Illness (OSMI) dataset, the study stratifies individuals into five different age groups (18-55+ years) and applies machine learning models-Random Forest, extreme gradient boosting, and support vector machine-enhanced with SHAP to identify key predictors across life stages. The findings reveal significant variations in risk factors: younger adults (18-24) are influenced by social support, while familial history and past mental health disorders gain prominence in middle-aged groups (25-54). For older adults (55+), social networks and environmental stressors become critical. Unlike traditional << black box >> AI models, SHAP provides interpretable insights, ensuring transparency in predictive decision-making. This study contributes to the literature by demonstrating that mental health risks are not static but evolve with age, necessitating tailored interventions. The framework advances age-specific predictive modelling and offers actionable insights for policymakers and clinicians, particularly in resource-constrained settings. By addressing the limitations of conventional AI approaches, this research establishes a foundation for personalised, explainable, and effective mental health risk assessment across diverse populations. The integration of XAI with age stratification sets a new benchmark for mental health research, highlighting the transformative potential of AI-driven, context-sensitive solutions in addressing the global burden of mental health disorders.
Big data analytics encounters scalability, latency, and privacy challenges, especially within real-time streaming contexts. We propose the Privacy-Aware Quantum Stream (PAQS), a distributed framework inspired by quantum principles, to overcome these obstacles. PAQS utilizes quantum superposition to effectively represent high-dimensional data, quantum entanglement for sophisticated correlation analysis and anomaly detection, and federated learning combined with homomorphic encryption to maintain privacy without compromising performance. The adaptive switching mechanism balances quantum-inspired and classical processing according to sensitivity and dimensionality criteria. Experiments are conducted on three datasets—OpenStreetMap, MIMIC-III, and KITTI, which show significant improvements: a throughput of 2.53 TB/sec, a 60 % reduction in latency, an anomaly detection accuracy of 92.3 %, and an 85.4 % decrease in privacy violations when compared to baselines. These findings validate that PAQS provides consistent, secure, and scalable real-time analytics, positioning it as a strong solution for smart cities, healthcare, and autonomous transportation applications.
The rising prevalence of diabetes has made Diabetic Retinopathy (DR) a major cause of blindness, underscoring the necessity for a computer-aided diagnostic system that can support clinical diagnoses without requiring extensive human effort. Many researchers have turned to deep learning to create automated screening and diagnostic tools for DR. However, for such systems to be truly effective in clinical practice, they must provide highly accurate assessments and well-calibrated estimates of uncertainty. Unfortunately, deep neural networks often tend to be overconfident in their predictions and are not easily amenable to probabilistic approaches. In our study, we introduce a novel approach for evaluating diagnostic uncertainty in DR predictions by employing ensemble-based calibration techniques. What sets our approach apart from cutting-edge convolutional neural network models is our use of the EfficientNet architecture, which offers superior accuracy through transfer learning. We then apply a set of post-calibration techniques to transform the model's probabilistic output into a confidence level. To gauge the uncertainty of our forecasts, we compute the entropy of the calibrated confidence value. This approach greatly assists users in determining whether it is necessary to seek a second opinion. Our model achieves an impressive accuracy score of 96 %, and our ensemble technique exhibits a notable reduction in Expected Calibration Error (ECE), in addition to providing a reassuring uncertainty score.
Crowdsourcing is the most effective means of obtaining labelled data for supervised machine learning. However, the varying expertise of crowd workers often results in noisy annotations. While traditional label aggregation methods attempt to handle label noise, they typically overlook the relationships between different data instances. Moreover, crowdsourced datasets often experience class imbalance, wherein predominant classes eclipse minority classes, hence exacerbating label accuracy issues. This paper proposed a Reliability-Weighted Bayesian Label Aggregation (RWBLA) to overcome the above challenges. First, the K-nearest neighbours (KNN) method is used to improve the label set for each instance by augmenting labels from its closest neighbours, resulting in multiple noisy label sets. Next, it improves the aggregation by assigning weights to the neighbouring labels based on worker reliability, adaptive distance, and label similarity for each instance. In addition, worker reliability is determined dynamically based on the neighbourhood information, and to handle the imbalance issue, label similarity is modified. In the end, a weighted Bayesian inference method is used to infer the correct label for each instance. The performance of the proposed approach is evaluated on 20 synthetics and three real-world crowdsourcing datasets. It shows that RWBLA consistently surpasses eight baseline label aggregations, improving aggregation accuracy by 3 % to 13 %. Moreover, in analyses utilizing imbalanced real-world crowdsourcing data, RWBLA surpassed state-of-art aggregation algorithms by 2 % to 6 %, underscoring its efficacy in situations with minor class imbalances.
Background: This study presents a systematic bibliometric mapping and analysis on the acceptance of technology in the field of artificial intelligence (AI), machine learning (ML) and neuronal networks (NN), evaluating the evolution and research trends within this interdisciplinary field. Methods: Using data from the Web of Science (WoS) and Scopus databases, we identify important authors, institutions, and geographic distributions, highlighting key research areas and emerging themes. The analysis was performed using VOSviewer (v.1.6.20) and RStudio (v.4.1.3) with the Bibliometrix package. Results: Our analysis reveals a steady increase in scientific output between 1999 and 2023, with a notable acceleration in recent years, indicating a growing interest in how AI technologies are accepted in various domains. The research illuminates the central role of technology and AI acceptance models, as demonstrated by thematic and keyword analyses. The study reveals a pronounced focus on the technological facets of AI acceptance, alongside discernible gaps in research linking energy, climate mitigation, and sustainability. Differences in findings underline the characteristics of the WoS and Scopus databases. Conclusions: The findings argue for a diversified research agenda to overcome these identified gaps, fostering a more comprehensive understanding of technology acceptance in the age of AI. This research charts a course for future explorations within this critical interdisciplinary field.
Cryptography plays a pivotal role in safeguarding data from unauthorized access. Various encryption techniques have been developed and implemented to secure data during transmission through robust encoding measures across diverse systems, contributing to standardizing cryptographic practices. Previous studies have introduced several algorithms that focused on augmenting encryption security, such as the Key Scheduling Algorithm (KSA), S-Boxes customized to incorporate key, and PT dependencies within the RC4 algorithm. These S-Boxes were later expanded by integrating Henon Chaotic Maps and Logistic Chaotic Maps. The main objective was to create distinct S-Boxes with high-strengthening security features of the Advanced Encryption Standard (AES) algorithm. Different evaluations, such as tests for nonlinearity, Avalanche Effect (AE), the Strict Avalanche Effect (SAE), and Execution Time Performance Efficiency in percentage (ETPE %) were employed to assess the robustness of these S-Boxes. This paper is dedicated to evaluating the robustness of different S-Boxes using Bit-Independent Criteria. We incorporated five distinct S-Boxes into the AES algorithm, presenting an innovative approach. The novelty of our work lies in applying Bit-Independent Criteria tests, quantitatively measuring the strength of each S-Box individually. Tests show that the S-Boxes produced through chaotic maps noticeably strengthen encryption. The Henon based S-Box, using secret key "0123456789ABCDEF", achieved an average BIC value of-0.009, while the standard AES S-Box recorded-0.04, indicting a relative improvement of approximately 77.5 % in the output bit independence. Similar gains appeared with Logistic map-based S-Boxes under various key settings. By utilizing the Bit-Independent Criteria, our research aims to offer a comprehensive assessment of the S-Boxes' effectiveness in enhancing AES encryption security. This endeavor ultimately contributes to advancements in secure data transmission practices. Overall, our results demonstrate that chaotic S-Box design boosts AES security and guides future work on tougher cryptographic schemes.
Humans’ mental conditions are often revealed through their social media activity, facilitated by the anonymity of the internet. Early detection of psy- chiatric issues through these activities can lead to timely interventions, po- tentially preventing severe mental health disorders such as depression and anxiety. However, the complexity of state-of-the-art machine learning (ML) models has led to challenges in interpretability, often resulting in these models being viewed as «black boxes». This paper provides a comprehensive analysis of explainable AI (XAI) within the framework of Natural Language Processing (NLP) and ML. Thus, NLP techniques improve the performance of learning-based methods by incorporating the semantic and syntactic features of the text. The application of ML in healthcare is gaining traction, particularly in extracting novel scientific insights from observational or simulated data. Domain knowledge is crucial for achieving scientific consistency and explainability. In our study, we implemented Naïve Bayes and Random Forest algorithms, achieving accuracies of 92 % and 99 %, respectively. To further explore transparency, interpretability, and explainability, we applied explainable ML techniques, with LIME emerging as a popular tool. Our findings underscore the importance of integrating XAI methods to better understand and interpret the decisions made by complex ML models.
Several tasks must be performed for a software development project to be completed successfully. Task allocation is a complex task that significantly impacts the project's success. Techniques have been proposed over the years to support this step, aiming to minimize cost and development time and to reduce the negative impact of team members leaving the project. In this context, the Truck Factor (TF) is a metric that can determine the risk to a project and can be used when distributing tasks among team members. The TF concerns the distribution of knowledge about the project among the development team members, ensuring that knowledge is not concentrated in only one part of the team. This theme is relevant nowadays since team rotation has become frequent due to the increasing demand for software in recent years. Member allocation in software teams does not have an exact solution since it is an NP-hard problem. Thus, Search-Based Software Engineering (SBSE) techniques, which can apply optimization algorithms such as genetic algorithms, have been used in several types of research to solve this class of problem over the years. In multi-agent environments, a certain number of agents have perception and communicate to achieve their goals. Researchers in the multi-agent field use simulated environments to validate their research since the modeling and simulation must consider the main variables of a real environment. Therefore, this work proposes to build a multi-agent simulation using SBSE techniques to minimize the impacts caused by the Truck Factor in a software development team. In the simulations, we model different configurations for the software development teams. Through statistical analysis and hypothesis testing, our results show that the proposed approach minimizes the impacts caused by task allocation in software development teams by around 25 % when considering the TF metric during task allocation.
Cardiovascular diseases are among the leading causes of death globally, emphasizing the critical need for machine learning models that are both accurate and fair in clinical decision-making. This study introduces the Enhanced Regularized Polynomial XGBoost (ERP-XGB) model, which integrates polynomial feature expansion with L1, L2, and gamma regularization terms to improve classification accuracy, address class imbalance, and reduce algorithmic bias. ERP-XGB was evaluated on four benchmark datasets: Heart Failure (299 samples), Heart Attack (1,319 samples), Heart Disease (917 samples), and BRFSS (253679 samples). On the Heart Attack dataset, ERP-XGB achieved a ROC AUC of 99.59 ± 0.21 %, accuracy of 96.97 ± 0.49 %, F1 score of 97.73 ± 0.43 %, precision of 96.30 ± 0.73 %, and recall of 98.87 ± 0.47 %, with an average run time of 30.63 seconds. In terms of fairness, ERP-XGB reported an Equalized Odds (EO) score of 0.02 ± 0.01, Disparate Impact (DI) of 0.96 ± 0.02, and Demographic Parity (DP) values of 0.61 ± 0.01 for the unprivileged group and 0.64 ± 0.01 for the privileged group. On the Heart Disease dataset, ERP-XGB demonstrated even stronger performance, achieving a perfect ROC AUC of 100.00 ± 0.00 %, accuracy of 98.60 ± 0.43 %, F1 score of 98.58 ± 0.37 %, precision of 100.00 ± 0.00 %, and recall of 97.29 ± 0.48 %, with a run time of 41.45 seconds. Fairness evaluation showed EO at 0.03 ± 0.01, DI at 1.78 ± 0.03, and DP values of 0.69 ± 0.01 for the unprivileged group and 0.38 ± 0.01 for the privileged group. For Heart Failure, ERP-XGB achieved 89.82±0.02 % ROC AUC, 82.93±0.03 % accuracy, and strong fairness (DI=0.91±0.31). On BRFSS, it attained 90.57±0.000 % accuracy but showed lower recall (11.89±0.004 %) and fairness challenges (DI=0.38±0.03). These results confirm that ERP-XGB offers an effective balance between high predictive performance and robust fairness in clinical datasets, making it a promising tool for equitable cardiovascular disease diagnosis.
