Credit card fraud detection is a critical research area due to the significant financial losses and security risks associated with fraudulent activities. This study presents FraudX AI, an ensemble-based framework addressing the challenges in fraud detection, including imbalanced datasets, interpretability, and scalability. FraudX AI combines random forest and XGBoost as baseline models, integrating their results by averaging probabilities and optimizing thresholds to improve detection performance. The framework was evaluated on the European credit card dataset, maintaining its natural imbalance to reflect real-world conditions. FraudX AI achieved a recall value of 95% and an AUC-PR of 97%, effectively detecting rare fraudulent transactions and minimizing false positives. SHAP (Shapley additive explanations) was applied to interpret model predictions, providing insights into the importance of features in driving decisions. This interpretability enhances usability by offering helpful information to domain experts. Comparative evaluations of eight baseline models, including logistic regression and gradient boosting, as well as existing studies, showed that FraudX AI consistently outperformed these approaches on key metrics. By addressing technical and practical challenges, FraudX AI advances fraud detection systems with its robust performance on imbalanced datasets and its focus on interpretability, offering a scalable and trusted solution for real-world financial applications.

Urban safety critically depends on effective street lighting systems; however, rapidly expanding cities, such as Astana, face considerable challenges in maintaining these systems due to the inefficiency, high labor intensity, and error-prone nature of conventional manual inspection methods. This necessitates an urgent shift toward automated, accurate, and scalable monitoring systems capable of quickly identifying malfunctioning streetlights. In response, this study introduces an advanced computer vision-based approach for automated detection and analysis of street lighting conditions. Leveraging high-resolution dashcam footage collected under diverse nighttime weather conditions, we constructed a robust dataset of 4260 carefully annotated frames highlighting streetlight poles and lamps. To significantly enhance detection accuracy, we propose the novel YOLO-CSE model, which integrates a Channel Squeeze-and-Excitation (CSE) module into the YOLO (You Only Look Once) detection architecture. The CSE module leverages the inherent symmetry of streetlight structures, such as the bilateral symmetry of poles and the radial symmetry of lamps, to dynamically recalibrate feature channels, emphasizing spatially repetitive and geometrically uniform patterns. By modifying the bottleneck layer through the addition of an extra convolutional layer and the SE block, the model learns richer, more discriminative feature representations, particularly for small or distant lamps under partial occlusion or low illumination. A comprehensive comparative analysis demonstrates that YOLO-CSE outperforms conventional YOLO variants and state-of-the-art models, achieving a mean average precision (mAP) of 0.798, recall of 0.794, precision of 0.824, and an F1 score of 0.808. The model’s symmetry-aware design enhances robustness to urban clutter (e.g., asymmetric noise from headlights or signage) while maintaining real-time efficiency. These results validate YOLO-CSE as a scalable solution for smart cities, where symmetry principles bridge geometric priors with computational efficiency in infrastructure monitoring.

Progress in detecting payment fraud is challenged by the limited variety of publicly available datasets. Relying on one or two datasets makes it hard to compare fairly, disguises sensitivity to data changes, and limits the ability to evaluate explainable methods in depth. This article introduces FraudX SimS, a scenario-labeled synthetic dataset designed to expand the benchmark set for anomaly detection in payment transactions, particularly for fraud detection. The dataset preserves the class imbalance between legitimate and fraudulent activity and includes openly specified spatial, temporal, and behavioral features, allowing direct application of explainable artificial intelligence techniques. We establish baselines with standard machine learning models and report accuracy, precision, recall, F1-score, confusion-matrix results, and the area under the receiver operating characteristic curve (AUC-ROC) and the area under the precision-recall curve (AUC-PR), with a primary emphasis on recall given the cost of missed fraud. We further employ Shapley additive explanations to quantify feature contributions, enabling transparent error analysis and model refinement. Although synthetic, the dataset is constructed to support reproducible experimentation and cross-study comparisons, advancing the development of reliable and interpretable fraud-detection methods.

Today, information technologies are developing very quickly. They are mostly applied in all fields, including healthcare, where confidential information about patients is stored. However, the fast development of digitalization has intensified the appearance of new cyber threats. One of these kinds of threats carries a considerable risk to healthcare organizations. This threat is called malware. It covers various ways to steal or disrupt computers, installed systems, and software. While traditional protection approaches struggle to effectively counter malware, adopting and using machine learning and neural networks for data security against malware in healthcare emerges as the most useful approach. This study evaluates the effectiveness of various machine learning and deep learning algorithms, such as Random Forest, XGBoost, CatBoost, AdaBoost, Dense neural network, and Hybrid GRU-LSTM neural network in analyzing and classifying cybersecurity threats in healthcare. The utilized methodology describes the significant steps in dataset preprocessing, including data cleaning, normalization, feature selection, class balancing, and application of machine learning and neural network models for classification. The efficiency of all models was evaluated with four classification metrics: accuracy, precision, recall, and F1-score. All of them were successful, with the highest accuracy score of 0.989 for the Random Forest model. XGBoost and CatBoost had slightly lower scores of 0.982 and 0.974, respectively. DNN and GRU-LSTM reached accuracy scores of 0.960 and 0.966. The AdaBoost had the lowest accuracy score of 0.953. The experiments demonstrated the supremacy and robustness of the Random Forest model in this experiment and proved its applicability in many other tasks in the Artificial Intelligence field.

Sequence recognition is a critical task across numerous disciplines. While traditional methods utilizing Finite State Machines (FSMs) offer a structured data representation and high interpretability, their flexibility is limited. Contemporary Machine Learning (ML) algorithms exhibit high accuracy but demand substantial computational resources. Combining these paradigms can enhance the effectiveness of complex sequence recognition. This study explores the integration of FSMs with ML techniques to address sequence analysis problems. Three distinct applications are examined: text classification (spam detection), recognition of genetic sequences related to Alzheimer's disease, and image-based gesture identification. For each, hybrid models were developed and tested, combining Deterministic Finite Automata (DFA), Non-deterministic Finite Automata (NFA), and ML algorithms such as Random Forest, Gradient Boosting, and Multilayer Perceptrons (MLP). Experimental results indicate that these hybrid models achieve performance comparable to traditional ML methods, and in some instances, yield more accurate predictions. In spam classification, neural network models demonstrated the best results, with FSM-neural network combinations providing similar effectiveness. For genetic sequence analysis, gradient boosting-based models exhibited the highest accuracy, with the inclusion of FSMs maintaining performance while enhancing interpretability. In gesture recognition, neural network approaches proved most effective, but integrating FSMs with ensemble methods achieved a high level of predictive capability, surpassing conventional ML models. In conclusion, the integration of FSMs and ML presents a promising avenue in sequence analysis. Future research could focus on optimizing model architectures and applying them to other domains requiring high-precision recognition of intricate structures.

As unmanned aerial vehicles (UAVs) are increasingly employed across various industries, the demand for robust and accurate detection has become crucial. Light detection and ranging (LiDAR) has developed as a vital sensor technology due to its ability to provide rich 3D spatial information, particularly in applications such as security and airspace monitoring. This review systematically explores recent innovations in LiDAR-based drone detection, deeply focusing on the principles and components of LiDAR sensors, their classifications based on different parameters and scanning mechanisms, and the approaches for processing LiDAR data. The review briefly compares recent research works in LiDAR-based only and its fusion with other sensor modalities, the real-world applications of LiDAR with deep learning, as well as the major challenges in sensor fusion-based UAV detection.

Anomaly detection in real-world domains often involves highly imbalanced datasets, where rare but critical events are challenging to identify without distorting the natural data distribution. Many existing approaches rely on resampling techniques or prioritize accuracy-based metrics, which may not accurately reflect the model’s true effectiveness in such scenarios. This paper examines the application of the FraudX AI framework, initially designed for financial fraud detection, to both network intrusion and credit card fraud detection under natural class imbalance. The framework integrates Random Forest and XGBoost classifiers using a weighted ensemble, combined with calibrated threshold tuning and SHAP-based explainability. Experiments conducted on the CIC-IDS2017 and European credit-card datasets demonstrate that the model achieves high recall and AUC-PR without applying data balancing techniques. These results highlight the framework’s robustness and generalizability for detecting anomalies in complex, imbalanced environments.

This research addresses the critical challenge of cybersecurity in healthcare by evaluating the effectiveness of machine learning (ML) and deep learning (DL) models in identifying and mitigating five significant cybersecurity threats: distributed denial-of-service (DDoS), man-in-the-middle (MitM), malware, phishing, and SQL injection. The study integrates a secure hardware–software architecture utilizing WireGuard, a lightweight, modern VPN protocol that establishes encrypted tunnels between network nodes, ensuring robust data integrity, confidentiality, and authenticated communication. Two ML models, support vector machine and random forest, and four DL architectures, dense neural networks (DNNs), convolutional neural networks-long short-term memory (CNN-LSTM), and LSTM-gated recurrent unit (LSTM-GRU), are systematically trained and tested using publicly available datasets specific to each threat category. The experimental outcomes demonstrate exceptional detection capabilities for structured network threats, with DNN and CNN-LSTM achieving accuracies and F1-scores from 95% to 97.6% for DDoS and MitM threats. In malware classification, the performance of DNN and CNN maintains precision and recall above 94%. Phishing and SQL injection attacks have lower classification scores of around 82% for most models. Visual analytics, including accuracy, loss plots, and confusion matrices, provide valuable insights into the convergence behaviors and sensitivity of different architectures, highlighting the strong generalization of DNN and variability in recurrent models. Overall, this research highlights the substantial potential of DL, combined with secure communication technologies like WireGuard, in enhancing healthcare cybersecurity, while also identifying areas for further development and optimization.

Human Activity Recognition (HAR) involves complex, multi-output datasets that demand models balancing efficiency, scalability, and interpretability. This paper proposes a Scalable Fuzzy Neural Network (SFNN), an adaptive hierarchical deep architecture with a multi-section learning mechanism. SFNN processes smaller input windows with effective dimensionality reduction, achieving competitive performance without the heavy computational cost of traditional deep models. Leveraging the transparency of fuzzy systems, it offers a more interpretable alternative to black-box approaches. The model’s theoretical convergence guarantees and strong results on the Opportunity dataset confirm its effectiveness for diverse HAR tasks.

As Unmanned Aerial Vehicles (UAVs) become more accessible to the public, they become a common tool for malicious purposes. As a result, there is an increasing demand for Counter Unmanned Aerial Systems (CUAS) that can detect UAVs. Existing CUAS solutions often rely on high-priced radar systems or advanced technologies, primarily designed for military purposes. In this paper, a low-cost, effective, non-military CUAS that uses inexpensive smartphones' microphones and camera, along with machine learning models, is proposed to detect and track a malicious UAV (MUAV) in real-time. Our proposed CUAS is designed to be affordable and accessible to the general public, operating automatically to detect and track MUAVs in real-time.

Dynamic reconfiguration is challenging, especially for a system of systems (SoS), although it is necessary given the nature and amount of change these systems undergo. SoS has certain specifics, such as independent development and the so-called emergent behavior, which need to be handled through new approaches. We investigate the possibilities of using modeling to provide some guarantees for such complex SoS in the presence of change. Our modeling language of choice is Systems Machine Language version 2 (SysML v2) and its textual representation in particular because it leads to more formal and precise representation and promotes the utilization of tools. We focus on the automation that tools can achieve as the SoS evolves during its life cycle. Our approach uses selected sets of models for each constituent system and a higher-level integration that ensures that all combinations can work at run time. We also explore the effects on SoS when adding, removing, and updating a constituent system. A protocol is proposed for handling dynamic reconfiguration in a distributed manner.

Algorithm selection and hyperparameter tuning are critical steps in both academic and applied machine learning. On the other hand, these steps are becoming ever increasingly delicate due to the extensive rise in the number, diversity, and distributedness of machine learning resources. Multi-agent systems, when applied to the design of machine learning platforms, bring about several distinctive characteristics such as scalability, flexibility, and robustness, just to name a few. This paper proposes a fully automatic and collaborative agent-based mechanism for selecting distributedly organized machine learning algorithms and simultaneously tuning their hyperparameters. Our method builds upon an existing agent-based hierarchical machine-learning platform and augments its query structure to support the aforementioned functionalities without being limited to specific learning, selection, and tuning mechanisms. We have conducted theoretical assessments, formal verification, and analytical study to demonstrate the correctness, resource utilization, and computational efficiency of our technique. According to the results, our solution is totally correct and exhibits linear time and space complexity in relation to the size of available resources. To provide concrete examples of how the proposed methodologies can effectively adapt and perform across a range of algorithmic options and datasets, we have also conducted a series of experiments using a system comprised of 24 algorithms and 9 datasets.

The rapid development of unmanned aerial vehicles (UAVs) has intensified the need for advanced classification techniques. This paper presents a novel approach that leverages audio data transformed into visual representations through Mel-Frequency Cepstral Coefficients (MFCCs) for drone classification. Our dataset consists of 28 drone types, each with 100 five-second audio recordings, from which 30 MFCCs are extracted per file. We investigate the effectiveness of this dataset by applying various vision models to the MFCC visualizations. Our results reveal that EfficientNet achieved the highest accuracy at 96.31%, followed by ResNet50 at 94.22%, and Vision Transformer at 73.69%. These findings highlight the potential of using audio-derived visual features for robust drone classification and demonstrate the varying performance of different vision models. This study provides a comprehensive examination of the methodology, experimental setup, and results, offering valuable insights into future research directions for enhancing classification accuracy with transformed audio data.

The popularity of Unmanned Aerial Vehicles (UAVs), aka drones, has increased rapidly in recent years. UAVs are becoming easily accessible to more users. Malicious intentions can erode public safety when least expected. Current methods used for UAV detection systems include computer vision, radar, radio frequency and audio approaches. We choose the audio method for its high accuracy, low computational requirement and low cost. However, the lack of publicly available datasets is one of the main bottlenecks for developing an audio-based UAV detection and classification system. To fill this gap, we select 15 different UAVs, ranging from toy hand drones to Class I drones and record a total of 8120 s length of audio data generated from the flying UAVs. To the best of our knowledge, the proposed dataset is the largest audio dataset for UAVs so far. We further implement a Convolutional Neural Network (CNN) model for 15-class UAV classification and trained the model with the collected data. The average test accuracy of the trained model is 98.7% and the test loss is 0.076.

One of the simplest approach in Reinforcement Learning (RL) is updating Q-table using Bellman operator. While theoretical expectations hint at the potential convergence achieved by modeling the discrete Q-table with the Bellman operator, practical limitations surface in real-world scenarios. The main challenges associated with it include the exponential growth of the Q-table size with an increasing number of state dimensions and the inability to use the Q-table in continuous state spaces. Alternative approaches, such as employing neural networks to approximate the parameterized Q-function, may not necessarily result in convergence.In response to these challenges, this paper introduces an simple innovative methodology inspired by the Bellman method updating. The proposed method utilizes fuzzy rules to discretize the state space, leading to the direct use of the Bellman operator for updating the fuzzy neural network weights, effectively acting as the Fuzzy Q-table. Instead of approximating the Q-function utilizing neural network/deep neural network based on gradient approaches, the proposed method establishes a Fuzzy Q-table and updates it using the Bellman equation. This strategic decision helps to solve the convergence problem in addition to prevent entrapment in local minima problems, a common challenge faced by conventional gradient methods. The efficacy of the proposed approach is demonstrated through its application to trading in the Bitcoin Futures Market, showcasing its ability to navigate complexities and uncertainties. Beyond financial markets, this methodology presents a versatile solution applicable to a diverse range of reinforcement learning problems, addressing limitations faced by traditional Q-tables or DQN.

Architecture description languages (ADL), many of them well-established, are popular in the systems and software engineering fields. They all exhibit different strengths, weaknesses, and specializations. Many of them are targeting static architectures with some limited support for dynamic changes of the architecture during run-time. The proposed Architecture Description Language in this paper targets the definition of architectures for system of systems (SoS) that can allow for dynamic changes in each system through manipulation of components during run-time and, therefore, dynamic reconfiguration. The existing modeling tools do not target dynamic reconfiguration of systems and, especially system of systems and could be used through possible extensions or annexes. Examples of such existing languages are SysML and AADL. These capabilities allow for improved system maintenance and adaptive behavior at run-time. The main contribution of this work is achieving these goals through the development of a new specialized language. dynADL and its environment have been developed using Antlr and C++ and have been tested for generating systems in POSIX environments. It is an architecture definition language that is extensible and flexible and has an experimental git repository with C++ source code for the implementation of the tools. The proposed tools allow for code generation of skeleton code for systems of systems that have dynamic behavior.

Stampede accidents frequently occur in situations where large crowds gather. Although education on how to protect oneself and prevent accidents in dense crowds has been provided, these measures mainly focus on post-accident responses. In contrast, this study proposes a proactive approach to prevent stampede accidents by utilizing thermal cameras to detect the number of people in a space in real-time and calculate the risk of a stampede. The system collects object detection and density estimation results using thermal cameras, considering population density in a specific area, and transmits the estimated results to other devices via serial communication. Thermal imaging technology excels at detecting people with high accuracy even in challenging daytime conditions or low-light environments at night. The data collected from the thermal cameras is continuously updated through machine learning and pattern analysis to assess stampede risk, and the results are provided in real-time via a web interface. This allows safety personnel and managers to effectively monitor high-density areas and take immediate action if necessary. Additionally, the system's web interface provides users with real-time information related to stampede risk. The validity of the proposed system has been demonstrated through field applicability and performance evaluation in real-world environments, contributing to enhanced public safety by preventing stampede accidents in advance.

In recent years, the widespread use of drones in daily life and large public events has raised serious safety concerns, especially due to incidents, both intentional and accidental. One of the most important aspects to prevent these risks is the ability to detect and accurately predict the distance of UAVs (Unmanned Aerial Vehicles) from people and restricted areas. One of the pressing issues is to develop systems to monitor their flights in restricted areas and predict suspicious movements in cases where suspicious drone flights are detected, which may be launched for video reconnaissance and information theft purposes. The advancement of acoustic-based recognition systems is growing with the development of deep learning. This study explores deep learning model architectures for the task of predicting UAV distances based on their flight sounds. The objective of this study is to determine whether sound-based classification can effectively predict drone movements at different distances from acoustic sensor points as they move from one area to another. Our experimental tests tried to predict the movement of UAVs based on the classification into 3 main zones. The results showed that drone sounds could be reliably detected in the movements between zones with an average recognition accuracy of 90 % using the hybrid CNN-BiLSTM model. Moreover, the implementation of such advanced acoustic sensor systems for UAV detection can improve the accuracy of real-time prediction, especially when integrated into a multimodal system with multi-sensor fusion methods.

Time Series Classification, is one of the very challenging areas in pattern recognition problems. As the volume of time series data increases, a multitude of TSC algorithms have been developed. Notably, only a limited number of these approaches have ventured into the realm of Deep Neural Networks to tackle this task. This paper presents a new hierarchical deep neural network architecture with a multi-section learning mechanism for the classification of multi-output time series datasets. The hierarchical structure of the proposed approach uses separate segments to detect output classes of different granular levels. The presented model does not suffer from heavy computations during training found in commonly used in other deep learning models such as convolutional neural network. Additionally, our suggested method handles data dimension reduction automatically in the hidden layers of different network sections and does not require separate pre-processing units for this purpose. To demonstrate the effectiveness of the proposed model, we have utilized it in the wearable sensor-based Human Activity Recognition problem and evaluated its performance on several benchmark datasets in TSC and HAR, such as PAMAP2 and Opportunity, with several different configurations. The empirical results on the benchmark datasets show that our proposed model outperforms several other deep learning-based solutions based on measures such as F 1 -Score, accuracy, and ROC curve.

This study examines the development of a model for the threat detection system with the use of machine learning and neural network methods. The fast development of Internet technologies has led to the appearance of many digital systems and platforms. However, despite the impressive technological progress, another side also emerged in the spread of a massive number of different cyber threats. Although various ways have been created to detect and prevent them, the threats are also developing and becoming more complex each year. Therefore, new system defense and data protection methods using machine and deep learning approaches have been proposed recently. The methods based on these approaches have proved to be especially effective in the wave of new Artificial Intelligence applications. In this paper, a threat detection system has been designed to disclose different kinds of threats while maintaining the security, confidentiality, and availability of the computer system. The development of machine learning models for detecting DDoS and man-in-the-middle attacks, Structured Query Language (SQL) injections, phishing, and malware was examined. The data scaling, feature selection, feature extraction, and classification steps were also thoroughly described. Naïve Bayes, Logistic Regression, Decision Tree, Random Forest, XGBoost, CatBoost, and Deep Neural Network algorithms were utilized for training the cyber threat detection models. The experimental results evaluated all the models using accuracy, precision, recall, and F1-score metrics. The best models achieved scores in the range of 0.90 to 1.00.