Blackhole and Sybil attacks represent major security threats to Wireless Sensor Networks (WSNs). Detecting these threats is essential to preserving the integrity of network operations. However, their detection is complicated by the various ways they undermine the system, causing data loss to spoofing identities, deceiving trust mechanisms, and ultimately compromising the network’s performance and reliability. Therefore, an effective model, ConvNeXtV2 with Taylor-based Mean Squared Logarithmic Exponential loss (ConTMSLE), is introduced to detect blackhole and Sybil attacks in WSNs. Initially, a WSN simulation is performed. Subsequently, Cluster Head (CH) selection and routing are executed by the Fractional Snow Leopard Beetle Optimization (FraSLepBO), considering energy, trustworthiness, distance, and selfishness. Then, at the base station blackhole and Sybil attacks are identifed. Network log data is processed using Gaussian function-based normalization, and feature selection is done using the Gini Impurity-based Weighted Random Forest (GIWRF). Finally, blackhole and sybil attack detection is done by the ConTMSLE technique. Moreover, the proposed ConTMSLE attains a better True Positive Rate (TPR), accuracy, True Negative Rate (TNR), precision, and F1-Score of 97.839
EEG-based cognitive workload assessment has attracted considerable attention in fields such as brain–computer interfaces and human–computer interaction due to its high temporal resolution and non-invasive nature. However, existing studies mainly focus on time-domain and frequency-domain features, while effective modeling of EEG spatial topological structures remains insufficient and is highly susceptible to electrode displacement and inter-subject variability. To address these issues, an EEG spatial modeling method based on Hu invariant moments is proposed for cognitive workload recognition. First, multi-band power spectral features are mapped into two-dimensional scalp topographic maps. Subsequently, Hu invariant moments with translation, rotation, and scale invariance are extracted to characterize the global spatial distribution patterns of EEG signals. The resulting geometric-invariant representation is incorporated into the original multi-domain representation of DHPL-Net, providing an explicit spatial-topological complement to conventional spectral and statistical features without modifying the network architecture. Experiments conducted on the public STEW dataset demonstrate that the incorporation of Hu invariant moments improves both accuracy and F1-score compared with the baseline method, while multi-seed experimental results further verify the robustness of the proposed approach.
This paper introduces SemChain, a semantic-guided skeleton-based human action recognition framework that integrates structural motion modeling and semantic supervision to address fundamental limitations of existing skeleton-based methods. Current approaches, particularly those based on graph convolutional networks, typically assume that action categories are well defined and separable through joint-level spatio-temporal motion patterns. This assumption is valid in controlled settings but breaks down in realistic scenarios involving fine-grained actions with overlapping or highly similar motion trajectories. This challenge is closely related to kinetic ambiguity, where distinct actions exhibit nearly identical physical dynamics but differ in semantic intent. To overcome kinetic ambiguity, SemChain enhances representation learning from both structural and semantic perspectives within a unified computational intelligence framework. Structurally, a dual-branch architecture combines joint-level modeling with explicit motion-chain modeling to capture coordinated body-part dynamics beyond isolated joints. Semantically, a novel semantic anchor scheme embeds linguistic priors into the visual feature space, enabling adaptive semantic-guided alignment and improving class separability. This design explicitly addresses the limitations of conventional one-hot supervision, which treats action categories as isolated labels and ignores their semantic relationships. Extensive experiments on both NTU RGB+D 60 and BABEL datasets have demonstrated that SemChain achieves strong performance in fine-grained real-world scenarios and competitive results on conventional benchmarks, validating its effectiveness for fine-grained skeleton-based action understanding and its robustness across diverse settings.
The detection of abnormal traffic in Wireless Sensor Networks (WSNs) is a major research problem as the complexity of the cyber-attack increases along with the limitations of available resources of the sensor nodes. In an attempt to solve the issues related to the high dimensionality of feature space and underfitting in the detection of traffic anomalies, this paper presents a unique Mutation Pelican Optimization Algorithm (MPOA) used in association with a Deep Convolutional Neural Network (DCNN). Within MPOA, the hunting behaviors of pelicans are utilized to create a population of solutions that help to determine the best feature subsets. In this proposed methodology, the concept of Mutual Information (MI) is employed to rank features based on the relationship between network characteristics and class labels. Then MPOA performs the selection of features before using the DCNN for classification. Finally, the selected features are provided to the DCNN, which recognizes the hierarchical features and classifies the network traffic either as normal or as DoS. The proposed framework combines lightweight network architecture with effective feature learning while operating on an optimally selected subset of features, thereby reducing computational complexity and storage requirements. The model was validated using three benchmark intrusion detection datasets, NSL-KDD, KDD Cup 99, and UNSW-NB15, and its performance was compared with several state-of-the-art machine learning and deep learning models, including DNN, CNN, SNN, and conventional DCNN. Experimental results demonstrate that the proposed MPOA–DCNN model achieves superior attack detection performance while maintaining a compact model size, making it well suited for deployment in resource-constrained WSN environments.
The growing requirement for secure multimedia communication has tested typical video steganography methods, which often trade off embedding capacity, visual quality, and signal-processing robustness. This study proposes an adaptive Discrete Wavelet Transform (DWT)-based video steganography architecture that uses key-based pseudo-random frame selection and adaptive coefficient embedding to hide data safely and undetected. Two layers of Haar DWT deconstruct selected video frames, and a dynamic scaling factor based on local and global statistical parameters encodes hidden information in LH and HL sub-bands. We test the proposed system on the UCF101 dataset of human action videos under Gaussian noise, JPEG compression, filtering, and frame dropping. Experimental results demonstrate that the recommended approach generates a high peak signal to noise ratio (PSNR) of 43.85 dB and a low MSE of 4.12, resulting in excellent visual clarity with reduced distortion. The suggested framework has a maximum embedding capacity of 0.58 bpp, outperforming LSB, DCT, and DWT. Robustness analysis shows a bit error rate < 0.02 with common attacks, ensuring data recovery and signal resilience. Results reveal that the adaptive DWT design balances security, embedding efficiency, and visual quality, making it suitable for secure multimedia transmission and digital content protection.
Effective spectrum use has become a key issue in the current wireless systems since the IoT and 5G/6G systems are rapidly expanding. CSS in Cognitive Radio Networks enhances spectrum-access but has low detection and high false-alarm rates in low Signal-to-Noise Ratio (SNR) environments. A new Self-supervised Transformer-enhanced Hypergraph Neural Network with Meta-optimization (ST-HGNN-MetaCSS) framework is suggested to resolve this issue. The model combines the application of hypergraphs, multi-scale temporal feature extraction, transformer-based attention, contrastive self-supervised learning, meta-optimization into adaptive learning. At RadioML2016.10b test, the model achieves an accuracy of 72
Agentic AI based on large language models (LLMs) is rapidly evolving from static chatbots to autonomous systems that plan, act, and interact with tools in open-ended environments. However, current LLM agents lack calibrated uncertainty, robust safety mechanisms, and faithful explanations, making them ill-suited for safety-critical settings such as healthcare, finance, cybersecurity, and industrial operations. Extended Belief Rule Bases (EBRB) are representative examples of interpretable, rule-based probabilistic reasoning frameworks with explicit representation of belief and ignorance, and have been successfully applied in complex decision problems without suffering from the rule explosion that affects traditional rule-based approaches. This position paper argues that EBRB could serve as a core safety and reasoning governor within LLM-based agentic pipelines, yielding hybrid systems to better align agentic AI systems with emerging regulatory and ethical requirements for Trustworthy AI (TAI) in high-risk settings. We outline: (i) a conceptual architecture integrating LLMs with EBRB in agentic workflows; (ii) the mapping from this architecture to TAI dimensions including transparency, uncertainty, safety, fairness, and auditability; (iii) concrete opportunities across healthcare, finance, cybersecurity, and industrial safety; and (iv) a research agenda highlighting open challenges in scalable rule induction, co-adaptation between LLMs and EBRB, and evaluation of hybrid agents. Our goal is not to present empirical benchmarks, but to articulate a vision and roadmap for EBRB-enhanced agentic AI that is both powerful and trustworthy.
Breast cancer detection from mammographic images remains a complex task due to variations in tissue density, imaging inconsistencies, and therapy-induced phenotypic changes. Conventional methods often struggle with tissue-intensity ambiguity in BI-RADS 4 regions, leading to inaccurate lesion segmentation. To address these challenges, this study introduces the Filtered Deep MammographNet (FD-MagNet), a novel unified deep learning framework for accurate breast cancer identification. Initially, Guided Trilateral Gabor Filter (GuGaF) enhances mammogram clarity by preserving edges and highlights directional patterns for early detection. During segmentation, alterations in dense breast tissue cause a tissue-intensity vagueness, to solve this, a 2D Masi-Kapur’s Bitterling Optimized Swin Transformer (2D MaBiS-T) is integrated, which combines entropy mapping, bio-inspired optimization, and hierarchical attention for precise tumor isolation. Further during extraction, the anatomical variations in mammographic views cause pectoralis muscle glooms, to overcome this a novel Mexican Axolotl Optimized ResNet150 (Mexi-Net) is employed to adaptively extract lesion features from sensitive regions. Finally, to resolve the loss caused by trastuzumab-induced phenotypic reversions, the Chaotic Nutcrackered Graph Rank Region-Based Attention Network (C-GRAN) is deployed to maintain diagnostic reliability through spatial awareness and targeted attention. Therefore, the FD-MagNet model achieves an accuracy of 99.60
The growing complexity of contemporary software systems has intensified the need for intelligent and dependable techniques capable of identifying defects early. Conventional machine learning models often face limitations when handling high-dimensional code metrics, class imbalance, and limited interpretability, thereby diminishing their practical utility. More critically, existing approaches treat software defect prediction as a flat-feature classification problem, failing to jointly model the temporal evolution of software metrics and the structural dependencies between modules, two complementary dimensions that empirical evidence consistently identifies as the primary drivers of defect propagation in real-world repositories. In this study, we propose a quantum-inspired hybrid deep learning framework, QLSTM-GCN, that combines the sequential modeling strength of Quantum Long Short-Term Memory (QLSTM) networks with the structural learning capacity of Graph Convolutional Networks (GCNs). Importantly, all quantum-inspired operations are numerically simulated on classical hardware using unitary matrix transformations, requiring no quantum computing infrastructure and ensuring full reproducibility on standard computing platforms. This integration enables the simultaneous capture of temporal and topological dependencies inherent in software repositories. To refine the input space, an improved Quantum-Inspired Genetic Algorithm (QIGA) is applied for optimized feature selection, reducing feature dimensionality by 37-39
Modern decision systems increasingly require uncertainty models that are both parameter-dependent and temporally adaptive. Although soft set theory provides a versatile means of working with imprecise data, classical soft sets are static and unable to reflect changes in time or context. Recent development of the dynamic soft sets has provided a route through which uncertainty in adaptive settings can be modeled, but any single algebraic matrix framework has not been developed yet. This research paper forms an extensive theory of the dynamic soft set based on time or context-indexed matrix form and operator-based aggregation ( max , min , avg ) using frequency matrix decision methods. We demonstrate closure, commutativity, and equilibrium theorems of dynamic soft matrix operations, defining strict structural stability with respect to time or context evolution. Computational experiments prove that soft matrix consistency values ( ϵ≤ 0.2 ) conserve more than 95
Software-Defined Networks (SDNs) offer programmability and ease of management, but the architectural revolution poses very serious security threats to the conventional intrusion detection system. To overcome the aforementioned problems, this paper proposes Bedbug-HMM, a flow-level anomaly detection framework for SDN controllers that processes NetFlow measurements to identify traffic anomalies causing controller overload. The framework integrates NetFlow data extraction, PCA-enhanced preprocessing, Bedbug Metaheuristic Algorithm (BMA)-optimized Hidden Markov Models with quantile-discretized observation sequences, population-based ensemble scoring, MCC-optimized thresholds, and ECA-based mitigation policies for automated OpenFlow response.The experimental analysis performed on the NSL-KDD and UNSW-NB15 datasets proves the superior performance of the proposed approach over six baselines including clustering, MLP/GNN, Baum-Welch HMM, PSO-HMM, and GA-HMM (with + 5.9
A cryptographically relevant quantum computer (CRQC) would break the public-key algorithms that protect most digital communication, but the risk does not begin when such a machine is switched on. Under the harvest-now-decrypt-later (HNDL) threat model, an adversary records encrypted traffic now and decrypts it once a CRQC exists, so any data whose confidentiality must outlast the machine’s arrival is exposed from the moment it is transmitted. This scoping review maps documented data-confidentiality lifetimes, charted by data category, against post-quantum migration readiness, documented by custodian sector, and applies Mosca’s inequality to reason about which data categories are most exposed. Following the Arksey and O’Malley framework, Joanna Briggs Institute guidance, and the PRISMA Extension for Scoping Reviews, 42 sources spanning peer-reviewed literature, standards bodies, government documents, and labeled grey literature were charted. Documented confidentiality lifetimes reach decades or are effectively perpetual for national-security records, health and genomic data, biometric identifiers, and privileged legal communications, and span years to decades for financial records. Migration-readiness evidence is uneven: most developed for government and defense and for cross-industry surveys, partial for finance and telecommunications, and largely absent for health care, genomics, and the legal sector. Reasoning through Mosca’s inequality, data categories with multi-decade or indefinite confidentiality requirements are, by that logic, already exposed to HNDL, because credible expert estimates place a meaningful probability of a CRQC within the same horizon; these exposure assessments are reasoned inferences rather than measurements. The central finding is a structural mismatch between long confidentiality obligations and immature migration, compounded by a readiness evidence gap that is itself a risk.
High detection accuracy under computing restrictions is a major concern as remote sensing imagery grows exponentially. Light-YOLOv7+, a lightweight target detection system, improves multi-scale representation and inference performance by coordinated architectural modifications. The proposed design uses a Quantized Efficient Feature Fusion Module (QEFFM) and a Spatial–Temporal Dual Attention Module (STDAM) to boost discriminative features across pyramid levels by aggregating features at low cost and reducing redundancy. STDAM adaptively reweights spatial–contextual responses. In addition, an adjustable scale feature pyramid, dynamic anchor creation, and an ultra-lightweight decoding head reduce computing cost while maintaining fine-grained localization. Testing on NWPU VHR-10, DOTA, and RSOD shows continuous performance increases under model capacity constraints. On NWPU VHR-10, Light-YOLOv7 + has a mAP of 0.878 (+ 0.062 over YOLOv7) and a DOTA recall (0.847). While improving accuracy, the model also reduces complexity, FLOPs, and inference latency, making it suitable for resource-limited deployments. The framework is suitable for large-scale remote sensing applications like environmental monitoring and catastrophe assessment because lightweight feature fusion and attention modulation balance efficiency and detection accuracy.
Human motion prediction models can be trained to provide a stable training loss with good accuracy on the short horizon, while the predictions might suffer in the autoregressive rollout stage when the learned transition function is repeatedly applied. This is especially relevant for martial arts motion since small one-step errors can be compounded by multiple steps and high acceleration rates, as well as through strong coordination between limbs and torso. This study investigates neural motion prediction from both the convergence and stability standpoint, making a distinction between optimization convergence during training and trajectory-level stability during repeated inference. We cast the problem of motion prediction as learning a parameterized transition operator and examine the impact of various components of this operator on error propagation, such as local sensitivity, spectral gain, residual prediction, and dynamics regularization. Experiments evaluate and compare the performances of various forms of predictors such as recurrent, temporal convolutional, graph-based, and attention-based ones with the same teacher-forced and autoregressive evaluation protocols. Human3.6 M is the standard benchmark with high dynamics and martial arts-oriented subsets to assess transfer and robustness. Besides mean per-joint position error, we report drift area under curve, velocity acceleration errors, bone length constraint violations, Fréchet Motion Distance, Jacobian sensitivity and spectral. The results demonstrate that in the short horizon, accuracy is not enough for stable rollouts and that stability-aware constraints not only limit the growth of the error but also enhance the physical plausibility under long-horizon prediction. The study offers a repeatable diagnostic scheme and design suggestions for the implementation of neural motion predictors in high-dynamics martial arts environments.
Class imbalance remains a critical challenge in many real-world classification tasks, particularly within the medical domain, where the accurate identification of minority class instances (e.g., disease cases) is essential. While instance selection (IS) and data re-sampling (DR) have been widely adopted as data-level techniques to address this issue, limited attention has been paid to their integration and ordering effects. This study presents a comprehensive empirical investigation into both single and ensemble-based IS and DR methods, exploring various hybrid combinations and sequencing strategies to improve classification performance on imbalanced medical datasets. We propose and evaluate multiple hybrid workflows, such as hybrid combinations including IS + DR and DR + IS, ensemble combinations including data re-sampling (EDR) and ensemble instance selection (EIS) using union (U), intersection (I) and multi-intersection (MI) strategies, as well as their combined variants (e.g., EDR + EIS, IS + EDR). A total of 22 benchmark medical datasets are examined using five standard classifiers, with performance assessed via AUC and F1 score. Results demonstrate that hybrid and ensemble strategies consistently outperform single-method baselines. In particular, combinations such as Instance Prototypes Filtering (IPF) + EDR-MI and EDR-MI + IPF achieve superior AUC, while IPF + EDR-MI and Edited Nearest Neighbor (ENN) + EDR-MI offer enhanced F1 performance. We also discuss the trade-off between selection aggressiveness, model performance, and computational efficiency. This work provides new insights into the design of effective preprocessing pipelines for imbalanced classification and offers practical guidelines for integrating IS and DR techniques in both standalone and ensemble frameworks.
Deep Learning-based Question Generation (QG) has become an essential subfield in natural language processing (NLP), with transformative implications across education, conversational AI, and information retrieval. This review critically surveys the evolution of QG techniques from early rule-based and statistical models to modern neural architecture such as sequence-to-sequence models, attention-based frameworks, and transformer-based systems like BERT, GPT, T5, and BART. Key contributions of this paper include a comparative analysis of traditional and deep learning-based approaches, a comprehensive overview of datasets and benchmarks, and a discussion of evaluation metrics encompassing lexical, semantic, and human-centered dimensions. The review further explores the integration of reinforcement learning and multimodal capabilities in state-of-the-art QG systems. It identifies pressing challenges such as semantic coherence, hallucination, bias, evaluation limitations, and domain dependence. By synthesizing recent advances and highlighting future research directions such as zero-shot learning, domain adaptation, and human-in-the-loop training this paper aims to guide researchers and practitioners in developing more robust, adaptable, and ethically sound QG systems. The work serves not only as a resource for understanding current capabilities and limitations but also as a roadmap for advancing the field toward more context-aware, customizable, and high-utility applications.
The construction sector still faces a high incidence of fatal accident and ergonomic injury risk, largely due to the inherent limitations of traditional manual safety auditing. While computer vision (CV) provides a viable solution to this problem, traditional deep learning architectures have struggled to balance the parametric requirements of micro-object detection, such as safety masks, with the stringent hardware limitations of traditional computer vision solutions. Furthermore, traditional spatial detectors are unable to assess active biomechanical safety risk and manage automated incident reporting without compromising inference latency. To address this critical gap in safety auditing, this research introduces a novel unified cyber-physical system using a Tri-State Routing Engine. By stratifying a dataset in a deterministic manner to prevent micro-object parametric starvation, the system dynamically routes inference using a hardware-aware Hybrid Ensemble of YOLO11s and YOLO11m backbones. The spatial inference of overlapping regions is mathematically fused using a custom-designed Expert-Calibrated Weighted Boxes Fusion (EC-WBF) algorithm to reduce semantic inconsistencies and deliver a peak mAP@50 of 0.910. The system also bridges passive spatial detection and active kinematic tracking using a novel routing mechanism to dynamically crop worker tensors and route them to a lightweight YOLO11n-pose network. This allows simultaneous multi-target Rapid Entire Body Assessment (REBA) ergonomic tracking and extrapolates Implicit Geometry failsafes to account for missing PPE even in severe visual occlusion. The system also introduces a novel natively coded asynchronous Robotic Process Automation (RPA) daemon. This daemon runs completely decoupled from the GPU-bound visual inference loop and sustains 14.2 FPS on severely constrained local edge hardware with 4 GB VRAM, rapidly sending Tier 1 life safety alarms for kinematic falls and compiling Tier 2 admin compliance digests, thereby successfully closing the cyber-physical loop.
Random Forest (RF) is a widely used ensemble learning method known for its high classification accuracy. However, optimizing its numerous hyperparameters remains a persistent challenge, particularly the Maximum Depth (MaxDepth) of individual trees. This study examines the boundaries of MaxDepth and Bagging Fraction (
This paper presents a lightweight and computationally efficient deep learning framework for brain tumor classification, which reinterpreted through the lens of the Neurojico era, where algorithmic mediation and cognitive sovereignty redefine medical diagnostics. The proposed model integrates MobileNetV2 with a novel Chaotic Dynamic Walrus Optimization (CDWO) algorithm to achieve neuronal precision, reflecting a paradigm shift from human-exclusive diagnostic authority to hybrid cognitive systems. Finding brain tumors early and correctly is important for improving treatment and survival rates because they are one of the most dangerous conditions. AI-driven automation is necessary because interpreting large amounts of MRI data by hand takes a long time and is likely to make mistakes. The five-step model that this paper suggests includes getting images, preprocessing them, optimizing features, transfer learning, and testing performance. Experimental validation on a dataset of 3,264 MRI images demonstrates superior performance, achieving an accuracy of 94.3
Deep transformer architectures are increasingly being used in remote sensing scene classification as they capture scene contextual dependencies on a global scale, but since they use full-resolution images this means that they require large storage requirements and high computational costs reducing their ability to scale in large scale and resource constrained systems. To meet this, this paper presents a compression-aware transformer architecture, which is called VITAE-DeiT, which collaboratively learns both semantic compression and classification through a single learning pipeline. The suggested method provides a Vision Transformer-based autoencoder to create small but semantically discriminative latent representations, which then allows performing scene classification in the compressed space, using a small transformer. In contrast to the traditional approaches, which consider compression and recognition as two distinct processes, the presented framework explicitly incorporates into the compression process semantic preservation and thus reduces as much as possible information loss, which is important in the classification process. Decades of benchmark remote sensing dataset experiments show that the suggested approach yields a test accuracy of 94.44, good precision (95.32), recall (94.50), and MCC (0.9433), which is superior to the recent CNN, hybrid, and transformer-based baselines. In addition, the compression-conscious architecture can be used to minimize the effective complexity of inference, attaining 21.3ms per image using moderate model complexity. Ablation and complexity tests also support that it is the synergistic integration of transformer-based semantic compression and model global context that contributes to the performance benefits. These findings make VITAE-DeiT a scalable and effective model of high-accuracy, remote sensing, scene classification at storage and computation constraints.
