Introduction: The increasing complexity of networked systems and the rapid evolution of cyber threats have created significant challenges for intrusion detection systems (IDS), particularly in environments with highly unbalanced network traffic. This study proposes the IML-UTAD framework to improve intrusion detection performance and enhance detection of minority-class attacks. Methods and Materials: The proposed framework incorporates comprehensive data preprocessing, including data cleaning, transformation, and normalization. Adaptive Synthetic Sampling (ADASYN) is employed to address class imbalance by generating synthetic samples for minority attack classes. Hybrid Secretary Bird Optimization (HSBO), integrated with an information-theoretic feature selection approach, is used to identify relevant features, reduce redundancy, and improve feature representation. The selected features are classified using an ensemble model that combines Random Forest, XGBoost, and Support Vector Machine (SVM) via weighted soft voting. Results: Experimental evaluation on benchmark intrusion detection datasets demonstrates that IML-UTAD achieves improved accuracy, precision, recall, F1-score, and Matthews Correlation Coefficient (MCC). The framework particularly enhances detection performance for minority-class attacks compared with conventional individual classifiers and other existing approaches. Discussion: The integration of ADASYN effectively mitigates the adverse effects of class imbalance, while HSBO improves feature relevance and reduces redundant information. The ensemble model combines the complementary strengths of multiple classifiers, thereby enhancing robustness, classification stability, and generalization across different attack categories. Conclusion: The proposed IML-UTAD framework provides an efficient, scalable, and reliable solution for intrusion detection in unbalanced network traffic. By integrating imbalance handling, optimized feature selection, and ensemble learning, the framework demonstrates improved reliable detection performance, supporting its potential application in real-world network security environments.
Introduction: While laboratory-trained models are highly accurate, several factors can affect the accuracy of models in the field, such as changes in lighting, background noise, plant growth stage, and imaging conditions. In this review, emphasis will be placed on robust and efficient edge-intelligence approaches for detecting pomegranate diseases. Methods: Based on PRISMA, eight aspects were covered: lightweight CNN/YOLO models, Kolmogorov-Arnold networks, Mamba-based models, domain and test-time adaptation, uncertainty quantification, foundation models, model compression, and multimodal sensor fusion. Of these studies, 108 were published within the last six years (2020-2026), and three were fundamental studies. Results: Emphasis is placed on the trade-offs between accuracy, computational efficiency, adaptation stability, and uncertainty calibration. While lightweight models fail to capture the small lesions, heavy models are computationally expensive; state-space models and test-time adaptation either lose local information or suffer from compounding errors under distribution shift. Running and managing foundation models on resource-limited edge devices is also challenging. Discussion: While most studies have investigated one challenge, there are a couple of studies that have combined visual features with bioimpedance and/or temporal (weather) data to detect presymptomatic features. This shows the need for integrated architectures that involve efficient representation learning, model compression, model adaptation, multimodal sensing, and riskaware prediction. Conclusion: PomStage-Net, a conceptual framework (not yet experimentally validated), is proposed for lightweight, temporally ordinal, and uncertainty-calibrated detection of pomegranate diseases. Future research should focus on causally inspired multimodal sensor fusion, adaptive learning, and uncertainty characterization to achieve reliable, real-time, fielddeployable agricultural diagnosis.
Introduction: The rapid growth of Internet of Things (IoT) networks has significantly increased the attack surface, making the need for intrusion detection methods a priority, especially in resource-limited systems with limited computational power. Objective: We introduce CSO-LIDS, an intrusion detection framework that uses Cat Swarm Optimisation (CSO) to select informative network-traffic features for binary attack detection. Methods: CSO identifies feature subsets by combining both the seeking mode for local refinement and the tracing mode for global exploration. The selected features were tested on a knearest neighbors classifier using the CICIDS2017 dataset (80%- 20% split) for training and testing. Results: The optimal configuration achieved a classification accuracy of 99.88% on the test split, with 0 false negatives and 679 false positives, and an execution time of 21.74 s. The numbers used here are configuration-specific and pertain to the particular data split this work examines; we make no assumption that they remain unchanged under alternative data partitions or across independent repeated runs. Classification performance was relatively similar across the CSO configs tested overall, but execution time varied greatly between configurations. Discussion: The parameter study analyses whether better classification performance can generally be achieved by increasing the CSO population size or by performing more search iterations, following the experimental protocol established in this work. In this study, we present results of a binary classification on the CICIDS2017 dataset using a workstation-based implementation. Therefore, these results should not be used as an indication of the performance that would be achieved on real IoT or edge devices. Conclusion: CSO could be a useful feature-selection method for intrusion detection. Its generalizability and deployment efficiency require further validation using repeated independent runs, new IoT-specific datasets, external validation, and resource-constrained hardware.
Introduction: Morphologically rich languages such as Kazakh present major challenges for embedding-based question-answering systems because extensive affixation increases lexical variation and fragments semantic information across surface forms. This study examines whether Relational CSE (Complete Set of Endings), a rule-based morphology-aware segmentation method, can improve semantic matching and answer quality in Kazakh educational QA without changing the underlying neural architecture. Methods: A corpus of approximately 15,000 Kazakh question-answer pairs from school programming education was used. Clean and Relational CSE-segmented corpus representations were evaluated under complementary within-dataset and strict-fair protocols. The main model was paraphrase-multilingual-MiniLM-L12-v2, with additional experiments using BAAI/bge-m3 and intfloat/multilingual-e5-large-instruct. The evaluation combined intrinsic analysis of tokenization efficiency with extrinsic testing in an embedding-based top-1 QA setting, including within- dataset, strict-fair, fine-tuning, and expert-based protocols. Results: Relational CSE segmentation reduced token-level entropy and vocabulary dispersion, lowering actual Bits per Token from 8.9980 to 6.8260 and effective vocabulary size from 4376 to 3333. In the within-dataset setting, where each system was evaluated in its own representation space, segmentation was associated with higher values for several answer-level metrics; for MiniLM, BERTScoreF1@1 increased from 0.812120 to 0.831933 and Semantic@1 from 0.252667 to 0.320667. In contrast, under the strict-fair paired protocol with a fixed clean answer- reference space, segmentation significantly increased question-space similarity across all three encoders but did not produce consistent answer-level gains. MiniLM answer-level differences were not statistically significant after Holm correction, while BGE-M3 and E5-large showed significant decreases in selected answer-level metrics. Under a common strict-fair finetuning protocol using the same 676 test items for all three encoders, Baseline–Segmented differences were no longer significant for BGE-M3 or E5-large after Holm correction, whereas MiniLM retained significantly higher QSim but lower TokenF1@1 in the Segmented condition. Discussion: The effect of morphology-aware segmentation depends on the evaluation regime. Gains observed when questions and answers share the same segmented representation do not necessarily transfer to answer selection when the answer-reference space is held fixed. Conclusion: Relational CSE is a useful representation-level preprocessing strategy for Kazakh educational QA, but its benefits are not uniform across evaluation settings. It reduces tokenlevel entropy and effective vocabulary size while increasing sequence/token load, and consistently strengthens question-space similarity in the non-fine-tuned strict-fair comparison. Answerlevel effects depend on the encoder, metric, and whether evaluation is performed within a selfconsistent representation space or under a fixed answer-reference protocol.
Introduction: Motor vehicle insurance claims processing in Zambia is predominantly manual, involving multiple independent stakeholders, fragmented information systems, lengthy verification procedures, and limited transparency, all of which contribute to processing delays and increased opportunities for fraudulent claims. This study aimed to design, implement, and evaluate a permissioned blockchain-based smart contract framework for automating motor vehicle insurance claims and billing while improving transparency, traceability, and transaction integrity within a multi-stakeholder insurance ecosystem. Methods: The study adopted a Design Science Research (DSR) methodology and was conducted as a proof-of-concept, deployment-oriented case study. A permissioned blockchain network was developed using Hyperledger Fabric, with smart contracts implemented in Java to automate vehicle registration, policy management, claim submission, evidence verification, billing, and settlement workflows. The prototype network comprised five participating organizations: the Zambia State Insurance Corporation (ZSIC), the Road Transport and Safety Agency (RTSA), the Pensions and Insurance Authority (PIA), the Zambia Police Service, and accredited motor vehicle garages. Role-based access control, Membership Service Providers (MSPs), Transport Layer Security (TLS), and multi-party endorsement policies were implemented to enforce secure transaction validation and institutional accountability. The prototype was evaluated through functional testing and performance measurements of transaction execution, endorsement, ordering, and commitment operations. Results: The prototype successfully automated the principal stages of the motor vehicle insurance claims workflow while enforcing predefined business rules through smart contracts. Experimental evaluation demonstrated that blockchain transaction processing, including endorsement, ordering, validation, and commitment, was completed within 2–5 seconds under the prototype deployment environment. The system also provided immutable audit trails, distributed data integrity, role-based authorization, and multi-party transaction validation, demonstrating the feasibility of improving transparency, traceability, and accountability in insurance claims processing. Discussion: This research represents one of the first practical implementations in Zambia of a customized Hyperledger-based blockchain framework designed specifically for the local insurance environment. The prototype demonstrates that blockchain technology is technically feasible and can significantly improve efficiency, accountability, and fraud prevention in the insurance sector. Conclusion: Overall, the findings suggest that blockchain technology strengthened efficiency, accountability, and trust within Zambia’s insurance sector. That said, putting the system into large-scale operation successfully will depend on having the right technical expertise, appropriate regulatory backing, and sufficient infrastructure.
