Distributed Denial-of-Service (DDoS) attacks are one of the major threats in the cybersecurity domain. Lightweight feature selection is vital for improving DDoS attack detection efficiency and computational efficiency by selecting the most relevant features. Recently, numerous machine learning and deep learning models have been developed to detect various types of DDoS attacks; however, their performance is often hindered by the presence of irrelevant features, which can lead to increased false positives and longer processing times. To address this problem, we propose the WPBS-MLP Intrusion Detection System (IDS). The WPBS (Welch’s t-Test and Point Biserial Test) feature selection method is integrated with an optimized MLP classifier to detect DDoS attacks. It was evaluated on the CICDDoS2019, CICIDS2018, and CICIDS2017 publicly available intrusion detection datasets. The WPBS method selected 39 significant features from the CICDDoS2019 dataset, and these 39 features are retained and considered for experimentation w.r.t all datasets. The DenseMLP model achieved a range of 99.59% to 100% accuracy for binary classification and 84% to 100% accuracy for low-rate and high-rate attack detection across the three datasets. Further, the Wilcoxon statistical test provided validation evidence that the proposed WPBS-MLP-based IDS model showed superior performance compared to the existing research studies.
Recently, Large Language Models (LLMs) have been able to generate text that closely resembles human writing, raising concerns about academic misuse and misinformation. Existing detection approaches often depend on a single type of feature, require direct access to the underlying models, and are sensitive to variations in text length and paraphrasing. To address these issues, this paper proposes a Multi-Feature Accurate Detection (MFAD) approach that integrates handcrafted statistical and syntactic features with deep semantic features based on Global Vectors for Word Representation (GloVe) embeddings, Convolutional Neural Networks (CNNs), and Bidirectional Long Short-Term Memory (BiLSTM). The results of the experiments on Human ChatGPT Comparison Corpus (HC3) demonstrate that MFAD achieves 98% accuracy, 96.5% precision, 97.5% recall, 97% F1-score, with a minimum False Positive Rate (FPR) of 0.01 across multiple domains. Additionally, MFAD demonstrates strong cross-model generalizability across LLMs such as GPT-4, Gemini, and Claude-4, and exhibits resilience to text length variations and paraphrasing.
Metaheuristic algorithms depend on population initialization to ensure an effective balance between exploration and exploitation in complex search spaces. Randomly initialized populations result in uneven coverage and premature convergence, especially in high-dimensional or multimodal problem landscapes. In this study, we investigate traditional and seven advanced population initialization techniques, including quasi-random sequence, chaotic map, opposition-based, knowledge-based, hybrid, and machine learning-based approaches. We discuss their theoretical foundations, computational complexity, and practical effectiveness. Findings suggest quasi-random initialization improves convergence for medium-dimensioned problems, chaotic population initialization improves convergence for multimodal search landscapes, and machine learning techniques adapt to very-high-dimensional optimization problems. Advanced techniques increase computational cost but significantly improve convergence rate and algorithm robustness. Such insights can assist researchers and practitioners in selecting appropriate initialization techniques for the complexity of their problems, and also indicate the direction of future research into intelligent and adaptive initialization techniques.
The Direct Simulation Monte-Carlo (DSMC) method is based on splitting the rarefied gas process evolution into two ballistic and collision steps within a time step. The collision step is the more complicated, requiring a rigorous theoretical analysis to discover its relation to the fundamental kinetic equations. In this work, we present a systematic derivation and examination of the Bernoulli-Trial (BT) family of collision schemes. The master Kac equation describes the binary collision interactions as a stochastic process lying in the background of the DSMC probabilistic rules, and it serves as a starting point for our derivation of the BT family. This study is limited to the analysis of four BT members, including Simplified Bernoulli Trials (SBT), Generalized Bernoulli Trials (GBT), Symmetrized Simplified Bernoulli Trials (SSBT), and Symmetrized Generalized Bernoulli Trials (SGBT). These schemes are implemented in the DSMC code, and a simple relaxation problem is simulated to determine optimal values for the pair-selection parameter Nsel.
Software-defined networking allows centralized and flexible control of a network, but its performance largely depends on the controller and its location in the network. The current work presents an experimental study of the Software-Defined Networking (SDN) controller performance considering various placement methods using star, linear, and ring topologies. Placement methods, including centrality-based, greedy, clustering, metaheuristic, capacitated, and baseline approaches, are analyzed with four popular controllers: Beacon, RYU, OpenDaylight, and Open Network Operating System (ONOS). Experiments are conducted using Mininet-based emulation, and findings indicate that controller performance varies based on location within the network. Among the evaluated techniques, the partition-based approach is identified as balanced controller placement in linear and ring topologies, given stable latency and throughput values comparable to other optimized strategies. Overall, the lowest latency is produced by RYU, the maximum throughput by OpenDaylight, and the intermediate performance by ONOS and Beacon.
Supercomputers and advanced High-Performance Computing (HPC) systems are needed to address large scientific problems, AI models training or industrial/engineering tasks. One of the important steps during the installation, configuration and tuning stages, is to perform comprehensive benchmarking and obtain insights about the best ways of exploiting the systems capabilities. HEMUS is modern petascale heterogeneous HPC system. Benchmarking insights and optimization strategies from HEMUS are applicable to a broad class of contemporary HPC systems. In this paper we present results from general benchmarks, together with optimization strategies applied and the obtained conclusions. We also present benchmarks related to software simulation of quantum computing algorithms, as well as simulation using low-discrepancy sequences, as they cover a significant portion of the expected workload on the system. The results highlight the importance of careful tuning and optimization, demonstrating that significant performance gains can be achieved, with the proposed approaches broadly applicable across comparable HPC systems and workloads.
The combination of the Multiple-Input Multiple-Output technique with Orthogonal Frequency Division Multiplexing (OFDM) is a widely used technique to improve Quality of Service (QoS) in wireless communication. However, a high Peak-to-Average Power Ratio (PAPR) in OFDM leads to signal distortion when passing through a High-Power Amplifier (HPA) into its nonlinear operating region. Moreover, this distortion degrades the OFDM system performance by increasing the Bit Error Rate (BER). To address these problems, a Refracted Opposition-Based Archerfish Hunting Optimization (ROB-AHO) Algorithm is proposed to minimize PAPR in OFDM systems. The AHO Algorithm dynamically adapts to various scenarios, and the proposed ROBL helps AHO select the most suitable phase factors to minimize PAPR efficiently across operating scenarios. Experimental results demonstrate that the ROB-AHO achieved a BER of 3!!!x.7; & times;!!!x 10;-1 for Signal-to-Noise Ratio (SNR) at 10 dB that outperforms prior methods, namely the Asymmetrical Auto Encoder (AAE).
Automated travel document recognition is a key technology for digital identity verification. However, robust extraction of structured information from images captured in unconstrained conditions remains challenging due to perspective distortion, background clutter, motion blur, and heterogeneous lighting that often degrade the performance of the systems. The paper proposes a modular pipeline for automated travel document segmentation and data extraction that integrates instance segmentation, perspective rectification, optical character recognition, and rule-based field parsing. In order to avoid the use of sensitive personal data, the segmentation model is trained exclusively on a synthetic dataset generated in Blender that comprises 2500 annotated images with diverse variations in lighting, viewpoint, blur, and background. The experimental results demonstrate strong generalization from synthetic to real data with 99.50% mAP50, 99.22% mAP50-95, 90% character-level Optical Character Recognition (OCR), and 90% MRZ field extraction accuracy on synthetic data, and 88% MRZ extraction accuracy on a dataset with real documents.
The development of Artificial Intelligence (AI) and e-Government is a paradigm shift in the administration of the people. Although it supports optimization of service delivery, various challenges are also brought to the fore. This scoping review examines the challenges, threats, and regulations surrounding the implementation of AI in the government sector. Following PRISMA guidelines, fifty peer-reviewed articles published between 2021 and 2025 were reviewed. The analysis found four major groups of challenges: technical (32%), organizational (28%), ethical and social (25%), and regulatory and policy-related (15%). The most common identified obstacles included the lack of skills and talent (92%), integrating with legacy systems (85%), and algorithmic bias (75%). The authors suggest a multi-layered AI governance framework, which includes international and national regulation, organizational processes, and managerial practices. It provides strategic suggestions, which are aimed at promoting the effective and responsible use of AI in e-Government.
Advanced Encryption Standard (AES) security relies on the Substitution box (S-Box), which provides nonlinearity and confusion. Because this algebraic form is fixed, the cipher can be attacked with algebraic and structural cryptanalysis. To address this shortcoming, we proposed generating AES-compatible S-Boxes utilising chaotic Fisher-Yates initialization and metaheuristic optimization. A multi-objective fitness function, including the Average NonLinearity (AvgNL), Minimum NonLinearity (MinNL), and Differential Uniformity (DU), is used to select the optimal S-Box. The experimental results show that the constructed S-Box has a considerable average nonlinearity of about 112, and average Strict Avalanche Criterion (SAC) and Bit Independence Criterion (BIC) Non-linearity values of 0.50048 and 104.285, respectively. Also, its NPCR and UACI are approximately 99.5924 and 33.3214, respectively. Moreover, histogram, correlation, and entropy analyses of images encrypted by the proposed system indicate that the proposed S-Box provides stronger security and greater flexibility for image encryption.
The fast growth of Virtual Machine (VM) backup systems in cloud and enterprise environments has greatly led to exposures to disk-level anomalies brought about by ransomware and malicious data corruption. The currently used anomaly detectors are mainly content-based scanning or coarse-grained metadata analysis, which causes high computational complexity, slow response time, and an inability to scale to large-scale backup settings. To combat the above difficulties, a Semantic-Aware File Metadata Generation Framework (SA-FMGF) will be put forward in this paper to provide efficient and proactive file-level anomaly detection in a VM backup system. The framework proposed will use file-system and disk-level metadata only and will not require raw file content analysis. It will not lose detection ability or have them detects be interpreted. SA-FMGF represents compressed metadata, such as semantic metadata vectors, that are continuously being scored by lightweight unsupervised anomaly scoring systems to identify anomalies in the normal disk behaviour.
The rapid evolution of generative artificial intelligence has enabled the creation of highly realistic deepfake facial imagery, supporting innovative applications in filmmaking and digital media while simultaneously amplifying risks related to misinformation and public safety. As a result, deepfake detection has become a critical research priority that must combine strong predictive performance with transparent and trustworthy decision-making, since deep learning models remain largely opaque and difficult to interpret in sensitive judicial and information-critical contexts. In this work, we develop and evaluate five deepfake detection architectures, including a Vision Transformer and four Convolutional Neural Networks, trained on the 140K Real and Fake Faces dataset, with the best models achieving an accuracy of 95 percent. To address the fundamental challenge of explainability, we further integrate the LIME interpretability framework, which generates clear and visually intuitive explanations of model decisions, thereby enhancing transparency and strengthening user confidence in automated deepfake analysis.
Healthcare data is frequently fragmented over diverse organizations because of its extremely complex and confidential nature. However, the existing Federated Learning (FL) approach through a central server creates various challenges within healthcare, such as privacy vulnerabilities and regulatory compliance. Thus, this research proposes the privacy-preserving FL approach with Fully Homomorphic Encryption (FHE) for protecting the patient’s sensitive data in medical records. In the proposed framework, the Galois Automorphism-driven Linear Transformation with Brakerski-Fan-Vercauteren, named GALT-BFV, is proposed for improving the medical data privacy and security. Moreover, this research introduces the pre-trained model of XceptionNet for training the local and global models in FL. Finally, the Federated Proximal (FedProx) approach is introduced for the aggregation of local and global models. The experimental discoveries establish that the proposed GALT-BFV method reaches better accuracies of 0.98 and 0.88 on Coronavirus Disease 2019 (COVID-19) X-ray and brain tumor Magnetic Resonance Imaging (MRI) datasets, compared to previous approaches.
This work investigates the effect of Spike Timing-Dependent Plasticity (STDP) of the synapses in the randomly connected Spiking Neural Networks (SNN) on the distribution of the firing rates of the individual neurons. It was observed that STDP, as a homeostatic plasticity rule, forces SNN activity to reflect the input structure. This effect is similar but not identical to the Intrinsic Plasticity (IP) tuning of Reservoir Computing (RC) recurrent neural networks. Both IP and STDP rules allow for capturing of the input data structure into the network state. This explains why STDP-trained SNNs are good for feature extraction from multidimensional data for classification purposes.
There is an urgent need for algorithms capable of improving the selection of appropriate features that directly affect the improvement process of the algorithm’s accuracy efficiently. Therefore, in this paper, a new feature selection algorithm for binary classification is proposed, which is called binary FOX Optimization Feature Selection (BFOXFS), where BFOXFS is combined with the KNN algorithm, which is used for binary classification, and the classification error is utilized as the objective function for the proposed BFOXFS. For evaluation, BFOXFS is compared with Binary Particle Swarm Optimization Feature Selection (BPSOFS), and two versions of Binary Grey Wolf optimization algorithm for Feature Selection (BGW1FS and BGW2FS algorithms). The experimental results show the superiority of BFOXFS in feature reduction, where it chooses the minimum features with a lower fitness value comparable with others. Further, BFOXFS has better convergence capabilities, well at finding optimal values, which results in more reliable and efficient optimization algorithms. It is very suitable for implementing in real-world problems.
This paper presents an algorithm for calculating the static characteristics of a multi-coordinate electromagnetic mechatronic module. The proposed module features compact weight and dimensional parameters, enabling both linear and step movements in spatial coordinates with high accuracy and speed. A constructive calculation scheme is developed to investigate factors such as magnetic permeability along the conductivity coordinate under constant voltage supply, its rate of change, and the variation of electromagnetic force depending on the anchor displacement. For the holonomic module, which consists of electric, magnetic, and mechanical parts, the anchor is divided into non-magnetized sections to prevent flux branching in unexcited phases. The method employs a step-by-step approach for an inhomogeneous electromagnetic core, introducing a division of conductivity into central, external, and internal parts, which improves accuracy and efficiency. The developed mathematical models and algorithms allow the determination of boundary conditions, evaluation of elementary flux tubes, and validation through analytical and experimental comparison.
Phishing is a type of cyber threat that targets organizations and individuals worldwide, causing billions of dollars in losses. So far, most successful anti-phishing methods require experts to extract features from phishing sites and third-party detection systems to detect them. This paper presents a PhishFusionNet model, an effective wide-and-deep learning framework for identifying phishing URLs with a high degree of generalization and accuracy. The proposed model successfully discovers both sequential and global patterns within URLs. This is achieved by integrating character-level embeddings that represent the deep component with handcrafted URL features that capture the wide component. We have tested our proposed model on over six million real-world labelled URLs. The results of the test on such a large-scale dataset with an optimal accuracy of 98.9 % have demonstrated that our model outperforms many other tested approaches. Based on these results, we believe that our proposed model is an effective, reliable, and scalable solution for cybersecurity and real-time phishing-detection applications.
Critical failures and financial losses are two of the biggest problems with bug detection. Traditional debugging methods don’t work well as the problem gets more complicated. Large language models and deep learning have lately given promise to the automation of finding and fixing problems in software. These systems are better at finding defects and making patches than traditional methods because they learn syntactic and semantic patterns. This study presents a systematic review of benchmark datasets, detection algorithms, and repair frameworks published between 2018 and 2025. This article compares the creation of models based on graphs and tokens with that of transformer architectures and large language model -driven methodologies. It also talks about their pros and cons and how they are used in the real world. The paper also discusses unresolved challenges related to explainability, accuracy guarantees, and cross-project generalization. It also talks about scalability, validation, and evaluation metrics. It identifies research deficiencies and delineates prospective avenues for developing more reliable and robust software systems by integrating contemporary breakthroughs and offering a current summary of automated debugging research utilizing deep learning and large language model methodologies.
This research examines file protection in cloud environments using the SHA-256 algorithm. It evaluates secure hash variants in file transfer systems, aiming to improve throughput while meeting security requirements. Data analysis involved gathering and cleaning datasets, measuring encryption times, and applying statistical and graphical methods. Results show SHA-256 as an effective base for encryption, with parallel processing increasing efficiency when real-time speed is essential. The study highlights the value of transformational techniques to boost performance and recommends hybrid systems that combine SHA-256 with other algorithms for stronger security. These outcomes support further work on cryptographic methods to strengthen the safety and reliability of cloud infrastructures.
Intrusion detection is a major concern in network systems where numerous smart devices are interconnected to handle sensitive information. Such interactions expose networks to threats like weak authentication, eavesdropping, malicious payloads, and a high false alarm rate. To address these challenges, a Channelized Spatial Attention enabled Adam optimized Convolutional Network (ChSp-ACN) is developed to identify malicious activities. The proposed ChSp-ACN effectively manages data imbalance and refines input features using KNN imputation and normalization. Its spatial attention mechanism enhances detection accuracy by focusing on relevant attack features, while the Adam optimizer fine-tunes model parameters to minimize false positives. Experimental evaluation demonstrates that ChSp-ACN achieves superior results compared to existing methods, attaining an accuracy of 97.22%, specificity of 97.24%, sensitivity of 97.20%, and a False positive rate of 0.03, which attains maximal performance effectiveness under the BoT-IoT dataset.
