		FeedBack \| Contact Us \| Links \| Site Map

Welcome to Journal of Theoretical and Applied Information Technology

Submit Paper / Call for Papers

Journal receives papers in continuous flow and we will consider articles from a wide range of Information Technology disciplines encompassing the most basic research to the most innovative technologies. Please submit your papers electronically to our submission system at http://jatit.org/submit_paper.php in an MSWord, Pdf or compatible format so that they may be evaluated for publication in the upcoming issue. This journal uses a blinded review process; please remember to include all your personal identifiable information in the manuscript before submitting it for review, we will edit the necessary information at our side. Submissions to JATIT should be full research / review papers (properly indicated below main title).

Journal of Theoretical and Applied Information Technology
May 2026 | Vol. 104 No.9

Title:	EFFECTIVENESS OF THE USE OF CLOUD PLATFORMS AND NETWORK SIMULATORS IN HIGHER MILITARY EDUCATION IN CBRN TRAINING
Author:	PETRO DZIUBA, SERHII BURBELA, LESIA BALAHUR, SERGII STEPANOV
Abstract:	The article addresses the knowledge gap caused by the absence of an integrated quantitative framework for assessing how cloud platforms and network simulators affect CBRN training in higher military education. Existing studies describe virtual simulation, specialized modelling software for radioactive material spill scenarios and institutional mechanisms for medical and biological emergency response, but they do not provide a unified comparative model that combines technical readiness, pedagogical integration, organizational adaptability, security resilience and communication interoperability. The need for this study is determined by the growing role of digitally supported CBRN training and by the requirement to evaluate not only the availability of technologies, but also their measurable contribution to practical training outcomes. The aim of the study is to develop and test a Cloud Efficiency Index (CEI) for assessing the effectiveness of cloud platforms and network simulators in the CBRN training of students of higher military educational institutions. The research employed Delphi-AHP weighting, cluster analysis, correlation and regression analysis, bootstrap modelling (n = 500), and visualization in Python, Power BI and Tableau. The study compared Ukraine, Italy, Germany and Austria and identified three digital integration models: highly consolidated, balanced and buildable. The main new knowledge created by the study is the CEI-based Digital Readiness Matrix, which links digital maturity with practical training effectiveness and allows cross-country comparison of CBRN training systems. The results showed that Italy (CEI = 0.81) and Germany (CEI = 0.78) have reached a high level of digital maturity, Austria (CEI = 0.69) demonstrates a balanced model, and Ukraine (CEI = 0.54) remains at the stage of building digital infrastructure. Correlation and regression analysis confirmed a positive relationship between digital readiness and practical training effectiveness (r = 0.77; R2 = 0.68). The scientific contribution consists in transforming fragmented literature on simulation-based CBRN training into a measurable model for evaluating digital transformation, security resilience and resilient infrastructure in military education.
Keywords:	Digital Transformation; Cloud Platforms; Network Simulators; Higher Military Education; CBRN Threats; Cloud Efficiency Index (CEI); Security Resilience; Resilient Infrastructure.
DOI:	https://doi.org/10.5281/zenodo.20252199
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	AERP: AGGREGATED ENCRYPTING ROUTING PROTOCOL FOR WIRELESS SENSOR NETWORKS (WSN)
Author:	OMAR KHALID SALIH ALHAFIDH , YOUNIS SAMIR YOUNIS , SADOON HUSSEIN ABDULLAH
Abstract:	The data harvested of the sensor nodes in Wireless Sensor Networks (WSNs) have to be secured. However, the power consumption constraint of these networks makes the normal encryption technique an issue. Moreover, the routing process and clustering in WANs create data aggregation in some nodes. This aggregation increases the complexity of encryption. To mitigate the encryption issue, this work addresses the critical challenge of securing data aggregation in WANs under strict energy constraints. The main contribution of this paper is proposing an aggregated encryption routing protocol (AERP), which introduces a lightweight encryption and routing mechanism that minimizes power and memory consumption while maintaining data security. The proposed protocol covers three main folds. In the first fold, a semi-clustering routing protocol can be implemented easily without the complex setup and cluster formatting phase. Second, the protocol proposes a new key distribution and exchanging method for the encryption process. This method depends on regenerating the encryption key from the data collected from different number of nodes. The third fold, the encryption algorithm, Playfair, has been adopted to be modified to encrypt the data and to round over the encrypted data in the aggregating nodes. This work integrates lightweight encryption, dynamic key generation, and encrypted data aggregation without requiring decryption at intermediate nodes. To evaluate AERP, a sensor node has been constructed and Advanced Encryption Standard (AES) and the modified Playfair have been written to show the power usage of these algorithms. In addition, AERP has been written in simulation environment to show the power usage of data routing. Our results showed that the protocol reduced the encryption power usage of AES with 50% and memory usage with 50% of the microcontroller. On the other hand, the encryption power consumption has reduced the network life time with less than 4%.
Keywords:	Three Diminution (3D), Wireless sensor Networks (WSNs), Advance Encryption Standard (AES), Multi-hops Routing, Semi-clustering Internet of Things (IoT)
DOI:	https://doi.org/10.5281/zenodo.20252208
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	OPTIMIZED FINGERPRINT IMAGE DENOISING AND SEGMENTATION USING SPECTRAL-RESIDUAL ATTENTION ZERO-SHOT CONVOLUTIONAL NEURAL NETWORK
Author:	JAINY JACOB M., D. SHANMUGAPRIYA
Abstract:	Fingerprint recognition is extensively applied in biometrics security and forensic because it is unique and permanent. In reality, however the real fingerprint images are usually corrupted with noise and low contrast, and fractured ridges, which greatly decreases the reliability of the recognition. The objective of this work is to enhance the quality of fingerprint images and the segmentation by using a unified deep learning system. The main aim is to maintain the ridge continuity and minutiae data and deal with low-quality fingerprints and partially corrupted ones. So, this research proposed to use a hybrid method that involves the use of Decompress Zero-Shot Convolution Neural Network (DZOC) to implement fingerprint denoising and Spectral-Residual Attention Zero-Shot CNN (SRC-NET) to perform ridge segmentation. DZOC is based on compression reconstruction approach in which a residual is learned to avoid noise without destroying fine ridge information. SRC-NET is then applied to the denoised output and uses spectral-residual analysis and attention to appropriately identify the ridges versus background noise without using labeled training data. The effectiveness of the proposed framework is also shown in an experimental analysis, in which the SRC-NET attains a segmentation accuracy of 95.07, which is higher than existing models of segmentation. The findings validate the hypothesis that the proposed DZOC and SRC-NET framework can greatly increase the level of fingerprint clarity and reliability in segmentation. In general, the method offers a powerful and scalable approach to the real-world fingerprint recognition systems to be used with noisy and low-quality acquisition conditions.
Keywords:	Biometric Security, Convolution Neural Networks, Denoising, Fingerprint Recognition, Image Enhancement, Zero-Shot CNN
DOI:	https://doi.org/10.5281/zenodo.20252221
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	REAL-TIME URBAN TRAFFIC OPTIMIZATION USING EDGE-AI ASSISTED FEDERATED DRL AND GRAPH NEURAL NETWORKS
Author:	DR RAYADU CHINNARAO, VISHAL NAMIREDDY, DR. SUNIL L. BANGARE, DESIDI NARSIMHA REDDY, R S S RAJU BATTULA, S N LAKSHMI MALLUVALASA
Abstract:	The effects of urban congestion are well documented and include decreased mobility, higher emissions and decreased productivity. Current urban traffic optimisation systems need to gather data centrally, reveal mobility traces, and do not cope with the spatial nonstationarity of the intersections well. In this work, we propose EdgeFedGNN-DRL, a federated deep reinforcement learning (FDRL) framework based on graph neural networks (GNNs) for real-time urban traffic optimisation, which is augmented by edge-AI. EdgeFedGNN-DRL deployed DRL agents at intersection edge nodes and input local states into the DRL agents using a compact GNN encoder. Model updates were aggregated at a coordinator under explicit bandwidth and privacy budget, which was done periodically based on the graph. A policy optimisation method was used, which was based on federated proximal policy optimisation (Fed-PPO): Quantised model updates and optional differential privacy were applied to satisfy the realistic communication constraints. The framework was tested in a microscopic simulation with the high-resolution traffic flow data from Glasgow (470 sensor files; 1,461 days; 16,480,080 hourly records) on SUMO for repeatable control experiments. We evaluated EdgeFedGNN-DRL against a fixed-time controller based on the Webster rules and a centralised hybrid baseline consisting of GNN+PPO. Average travel time, per-vehicle delay distribution, throughput, queue length, CO₂ emissions, model convergence and edge inference latency were measured in the experiments. EdgeFedGNN-DRL achieved a reduction in average travel time of 15.3%, 7.6%, and an increase in throughput of 14.4%, 6.2% compared to Webster, respectively, and a reduction in per-client per-round uploads to around 48 KB, per-vehicle per-round delay of 24.2% and 10.0% compared to Webster, respectively, and a reduction in CO₂ emission of 11.6% and 5.7% compared to Webster, respectively, while keeping edge inference latency around 85 ms. The obtained results show that EdgeFedGNN-DRL can facilitate a privacy-preserving, scalable and latency-aware traffic signal control strategy, applicable to Edge-AI implementation in urban areas.
Keywords:	Federated Learning, Reinforcement Learning (RL), Graph Neural Networks, Edge-AI, Traffic Signal Control.
DOI:	https://doi.org/10.5281/zenodo.20252239
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	TREE TOPOLOGY-BASED RESOURCE ALLOCATION AND DISCOVERY ALGORITHM FOR ONE HOP
Author:	MUSA DIMA GENEMO, DR. YALAMANCHILI SAROJA, KISTAM GOPI, VEERAMANI MUTYAM, DR BHARGAVI PEDDI REDDY, DR.RAMIREDDY NAVATEJAREDDY, SRINIVASA RAO MADALA, NIMMAGADDA YASWANTH SAI, K. KOTESWARARAO
Abstract:	The components of a distributed system include multiple computing nodes which may use different types of hardware and connect to each other through a communication network that operates without shared memory or clock systems. The primary advantage of distributed systems enables organizations to share resources which results in faster computing times together with increased data availability and system reliability. The process of resource sharing requires organizations to first identify their resources and then assign those resources before anyone can access them. Computer systems implement caching to store frequently used data and program segments. The new caching technique called virtual caching enables a host node to delegate its cache authority for specific pages to nearby nodes who can use its partial cache space. The virtual caching protocol specifies neither the client node's process for accessing virtual cache nor its method for retrieving virtual cache from distant servers. The resource discovery and allocation problem needs us to develop a solution for this particular issue. The resource donation process requires us to find donor nodes who have extra resources while we need to assess which resources will be given to deficient nodes at each connected network point. Our goal is to achieve complete fulfillment of requests which deficient nodes make for resources. The importance of virtual caching lies in the fact that it doesn't require physical movements of cache. The requirements of the system determine when virtual cache allocation can be modified. The proposed heuristics provide efficient performance because they require minimal time and perform limited communication operations. We assess distribution quality through a comparison between heuristic distribution results and the distribution outcomes from the ILP problem solution. We present and assess several heuristics which aim to reduce unfulfilled resource requests that arise from resource-deficient nodes during their resource search process which has a maximum distance of two hops. The paper restricts its analysis to single hop operations only. The paper demonstrates through its research that this algorithm reaches optimal solution status based on our established benchmark criteria. The resource distribution for bounded hops restricts its operations to one-hop resource distribution. Every resource-surplus node transfers additional nodes to other nodes while maintaining its own resource balance without reaching resource-deficient status. Load distribution cannot be divided into infinite parts. Our research focuses on measuring the fulfillment level for each resource-deficient node request.
Keywords:	Donor, Deficient, Caching, Optimal, Topology, Resource Allocation
DOI:	https://doi.org/10.5281/zenodo.20252257
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	ENHANCED TEXT SUMMARIZATION USING HIERARCHICAL TRANSFORMER AND KNOWLEDGE GRAPH INTEGRATION FOR LEGAL DOCUMENTS
Author:	NINEESHA P , DR P DEEPALAKSHMI
Abstract:	Indian legal judgmental are substantially difficult to summarize due to their intense clauses and complex structure. Existing summarization techniques often fail to represent subtle legal reasoning and contextual dependencies embedded in the documents. This work presents a novel approach to Indian legal judgment summarization called Hierarchical Transformer with Knowledge Graph Integration (HT-KGI). It consists of two main components: hierarchical transformer encoding and knowledge graph integration. First, a hierarchical transformer encoder is employed to model contextual dependencies at both the sentence level and the document level using a dual-attention framework. Secondly a knowledge graph grounded in legal ontologies is incorporated to clarify the semantics of legal concepts and reveal their underlying relational structure. To assess the effectiveness of the proposed framework, we carried out an extensive evaluation using a large collection of Indian Supreme Court orders. The result strongly provided a marking improvements over the baseline systems, not only in ROUGE metrics but also in the semantic flow and legal soundness of the generated summaries. According to experimental results, the HT-KGI approach outperforms typical transformer models by 17% in ROUGE-1 scores, 22% in semantic coherence, and 19% in legal accuracy while retaining a 35% reduction in computing cost. For specialized document summarizing tasks, the suggested approach shows how well domain-specific knowledge graphs and hierarchical transformer structures work together.
Keywords:	Hierarchical Transformer, Knowledge Graph Integration, Legal Document Summarization, Extractive Summarization, Semantic Coherence, Legal Text Processing.
DOI:	https://doi.org/10.5281/zenodo.20252511
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	AUTONOMOUS DRONE-AI COLLABORATION FOR ADAPTIVE CRACK MAPPING AND PROGRESSIVE DAMAGE QUANTIFICATION DRONE–AI COLLABORATION FOR NON-DESTRUCTIVE ASSESSMENT OF CIVIL INFRASTRUCTURE
Author:	MADHAVI KATAMANENI , NAGIREDDI SURYA KALA , P RAVI PRAKASH , MELAM NAGARAJU, Dr.T. MURALIDHARA RAO , M. V. B. T. SANTHI , Dr. DIVVELA SRINIVASA RAO , Dr. HYMAVATHI THOTTATHYL
Abstract:	The deterioration of aging concrete structures requires inspection systems that are autonomous, adaptive, and highly accurate. Conventional UAV-based crack detection pipelines often separate navigation, segmentation, and damage progression analysis, limiting their ability to respond dynamically during missions. This study proposes an Autonomous Drone–AI Collaboration Framework that unifies UAV flight control, deep learning–based crack perception, and temporal damage modeling into a single adaptive workflow. The framework employs a Transformer-enhanced Swin-UNet for robust crack segmentation, EfficientNet-B3 for severity classification, and a TD3 reinforcement learning agent that continuously adjusts flight paths based on real-time crack feedback. A temporal convolutional network (TCN) further models crack growth, enabling predictive monitoring across inspection cycles.
Keywords:	Autonomous Drones, Computer Vision, Crack Detection, Damage Quantification, Deep Learning.
DOI:	https://doi.org/10.5281/zenodo.20252543
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	ENHANCING CLOUD SECURITY AND PERFORMANCE USING INTELLIGENT LOAD BALANCING WITH BLOCKCHAIN APPROACH
Author:	KRISHNA SOWJANYA K , MOULEESWARAN S K
Abstract:	Cloud computing has transformed how organizations store, process, and access data, but security and performance concerns remain. This research presents a novel approach that combines intelligent load balancing with blockchain technology to address these challenges. By integrating the Golden Eagle Optimizer (GEO) algorithm for load balancing, the system optimizes performance and resource allocation in cloud environments. Blockchain enhances security, transparency, and trust, addressing issues like security vulnerabilities and performance inconsistencies. The GEO algorithm achieves a success rate of 0.9493 and a security score of 0.9535, making it an effective optimization solution. Using Python Jupyter for optimization, the system ensures high availability and considers security factors when distributing workloads across servers. The combination of blockchain and GEO results in a secure, efficient, and resilient cloud system with improved Quality of Service (QoS). This approach has significant implications for the future development of secure cloud computing systems.
Keywords:	Golden Eagle Optimizer (GEO), Quality of Service (QoS), Blockchain, Cloud Security, Load Balancing.
DOI:	https://doi.org/10.5281/zenodo.20252571
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	SUPERPIXEL BASED CLUSTERING OPTIMIZATION FOR REAL TIME ESPRESSO CREMA ANALYSIS
Author:	HENDY CHRISTIAN , MARIA SERAPHINA ASTRIANI
Abstract:	Coffee crema is a key visual indicator of espresso extraction quality, yet practical, real-time automated analysis of crema remains challenging. Existing crema analysis methods remain limited because they rely on dense pixel-level processing, RGB-sensitive analysis, and semi-manual segmentation, which increase computational overhead and reduce robustness under practical imaging conditions. This study proposes a lightweight real-time espresso crema analysis pipeline that integrates automatic cup detection using the Circular Hough Transform, illumination-robust color processing in the CIELAB color space, SLIC superpixel segmentation, and K-Means++ clustering. By replacing dense pixel-wise clustering with region-level superpixel abstraction, the proposed approach reduces the number of analysis units while preserving crema area extraction and dominant color characterization. Experiments conducted on round-glass espresso images derived from the dataset of Choi et al. show that the proposed pipeline reduces the number of analysis units from 121,789 pixels to 28,979 superpixels, corresponding to a reduction in data complexity of more than 75%. The resulting crema area remains high across samples, ranging from 84% to 95%. In addition, the proposed method achieves a processing time of 4.07 s per image, compared with 6.1 s reported for the reference GrabCut-based pipeline, indicating a 33% runtime reduction. These findings demonstrate that region-level visual abstraction can preserve extraction-relevant crema information while improving computational efficiency. The main contribution of this work is the demonstration that automatic geometric ROI localization, perceptually robust color representation, and superpixel-based clustering can form a practical and scalable alternative to prior pixel-wise crema analysis pipelines for real-time espresso monitoring systems.
Keywords:	Machine Learning, Color Clustering, Hough Circle Transform, Superpixel Segmentation, Coffee Crema, Espresso
DOI:	https://doi.org/10.5281/zenodo.20252579
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	EDGE AI-DRIVEN AIR QUALITY FORECASTING USING KALMAN FILTERING: A LOW-LATENCY IOT FRAMEWORK FOR REAL-TIME ENVIRONMENTAL MONITORING
Author:	AYONIJA PATHRE , PARASA RAJYA LAKSHMI , KATAM NAGA LAKSHMAN, EERLA RAJESH , T.MURALIDHARA RAO 5, M. V. B. T.SANTHI , N.DHARANI KUMAR , Y.SUMANTH
Abstract:	Air quality is a critical factor influencing public health and environmental sustainability. Machine learning (ML) has become a potent instrument in the field of air quality monitoring and prediction in recent years. A network of sensors is often used by machine learning-based air quality systems to gather data in real-time on a variety of air pollutants, including particulate matter, nitrogen dioxide, sulphur dioxide, ozone, and more. These sensors are strategically placed in urban areas to provide comprehensive coverage. Existing cloud-based air quality monitoring systems often suffer from latency, bandwidth dependency, and delayed decision-making, limiting their real-time effectiveness. The collected data, often characterized by its multidimensional nature, undergoes a process of feature extraction. ML algorithms are employed to train on historical datasets. Once trained, the ML models are capable of making real-time predictions of air quality based on current input data. These predictions can be continuously monitored and updated, providing valuable insights into the dynamic nature of air quality within a given area. This article presents a novel approach for air quality monitoring and forecast system using edge computing and IoT in environmental application, using the Raspberry Pi and the Kalman Filter technique. Using the machine learning method, this work improves real-time decision-making by minimizing data transmission lags brought on by network and bandwidth constraints. In contrast to standard cloud-based monitoring and forecasting of air quality systems. Air pollutants like SO2, PM10, PM2.5, and NO2 are immediately predicted using the RPi, as an edge device with significant computational capacity. Additionally, the KF algorithm boosts sensor accuracy by 30% in compared with to sensor observation data and improved projected value accuracy. The proposed framework demonstrates the feasibility of low-latency edge-based air quality forecasting with improved prediction accuracy and reduced communication overhead. The results indicate its practical applicability for scalable smart city environmental monitoring systems.
Keywords:	AI, Air Quality, Forecasting, Framework, Kalman Filtering
DOI:	https://doi.org/10.5281/zenodo.20252595
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	A LAYER-WISE ATTENTION CALIBRATION FRAMEWORK FOR DEEP NEURAL NETWORKS IN RESOURCE-CONSTRAINED ENVIRONMENTS
Author:	M. NAGABHUSHANA RAO, RAMESH BABU PITTALA, DR. K. VENU, Dr P N V V L PRAMILA RANI, GUNDALA VENKATA RAMA LAKSHMI, Dr. BOBY K GEORGE, Dr. GRK PRASAD
Abstract:	More frequent use of deep neural networks in low-resource settings has revealed that conventional structures often cannot maintain both accuracy and computation efficiency at the same time. Solutions such as model pruning, quantization and knowledge distillation have been tried to reduce how complicated a model is, but they are not dynamic solutions. Still, most of these methods make errors, need to be retrained frequently and cannot adjust at run time to new challenges or system restrictions. Instead of similar strategies, this paper presents LWAC which regulates attention in different parts of the network when it’s being used for inference purposes. Not like fixed compression which must be retrained, LWAC adds lightweight calibration units to the model. These units check relevance and demands of each layer and enable correct pathways, while blocking those that do not help. All of the experiments were validated on CIFAR-10, Tiny ImageNet, UCI HAR and MHEALTH, with MobileNetV2, ResNet-34 and CNN-LSTM hybrids serving as the architectures. Latency in inferencing and energy use drop by 15% to 22% in LWAC, even as its accuracy can improve by as much as 1.2%. Moreover, different types of analysis underscore that LWAC can make important predictions and clarify the decisions it reaches. By supporting flexible use of layers based on needs, LWAC brings strong deep learning capabilities to places where computers are under pressure.
Keywords:	Deep Neural Networks, Attention Calibration, Layer-Wise Modulation, Resource-Constrained Inference, Adaptive Computation, Edge AI, Energy-Efficient Learning, Model Optimization, Real-Time AI Systems, Context-Aware Attention Control
DOI:	https://doi.org/10.5281/zenodo.20252642
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	FINDING AND ALLOCATING RESOURCES IN THE BEST WAY FOR A SINGLE HOP USING NON-ANONYMOUS ARBITRARY TOPOLOGY
Author:	M. VENKATA RAMANA, G. LAXMI DEEPTHI, NARASIMHA RAO THOTA, D. NAGA PURNIMA, V.V.RAMA KRISHNA, SIVANANDAM K, SANDEEP KONE, SHAIK JILANI BASHA, K. KOTESWARARAO
Abstract:	The distributed system comprises multiple computing nodes which may have different hardware and software components and these nodes connect through a communication network which does not permit them to share their memory or clock signals. The main advantage of distributed systems allows multiple users to access resources which results in faster processing and increased data accessibility and system dependability. The process of resource sharing requires two steps which include resource discovery and resource allocation. In computer systems caching allows users to access frequently needed data and programs which they have stored in their system. The new virtual caching system enables a host node to permit nearby nodes to use part of its cache memory for caching their data. The virtual caching protocol fails to explain the process through which a client node receives virtual cache from a remote host. We develop a resource discovery and allocation problem as a solution to this issue. Our research examines methods to find resource surplus donor nodes while assessing their resource needs for operational systems which require resource sharing in a network that connects all its parts but limits resource deficient nodes to a distance of one hop. Our goal is to reduce the number of requests that remain unmet by resource deficient nodes. Users can modify their virtual cache allocation whenever they need to do so. The proposed heuristics provide efficient performance because they require minimal processing time and restrict communication to essential operations. The distribution quality estimation process we use involves comparing actual distribution results obtained from heuristics with distribution results obtained from ILP solution. The study presents multiple resource request fulfillment heuristics which researchers will evaluate to determine their effectiveness in solving node resource request problems that occur within defined hop limits. The study restricts its examination to single hop connections between nodes. The paper demonstrates that the algorithm produces optimal results according to the established benchmark. The resource distribution for bounded hops operates with a single hop restriction. Resource surplus nodes transfer their additional nodes to other nodes in a manner that preserves their own resource capacity. The system does not allow for fractional load distribution. The study specifically examines how much each resource-deficient node will receive from its resource request.
Keywords:	One Hop, Resources, Topology, Non Anonymous.
DOI:	https://doi.org/10.5281/zenodo.20252665
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	HYBRID INTELLIGENT METHODS OF COMPUTER VISION AND DEEP LEARNING FOR HIGH-PERFORMANCE PATTERN RECOGNITION IN COMPLEX INFORMATION SYSTEMS
Author:	ANDRII KYSIL, KOSTIANTYN MINKOV , ANDRII SYROTENKO , ILNARA SHARIPOVA, SERHII ZAICHENKO
Abstract:	Modern information systems operate in complex conditions with high data variability and destabilizing factors. This creates a need for models that can simultaneously capture local features and take into account global spatial-semantic dependencies with high robustness and efficiency. The aim of the study was to develop and experimentally verify a hybrid intelligent model of computer vision (CV). The proposed approach combines the advantages of convolutional neural networks (CNNs) and transformer architectures to increase accuracy, stability, and generalization ability in pattern recognition tasks. The research employed the following methods: construction of a hybrid neural architecture (CNN + Transformer), the use of reprocessing CV methods, and an adaptive model router. Special attention is paid to uncertainty analysis (entropy estimation and MC Dropout), as well as comparative testing with the base models: pure heavy, pure edge, preproc+heavy, and ensemble. The experimental results showed that the proposed model achieves the highest accuracy (Accuracy = 0.941), F1-score (0.930), and significantly reduces the calibration error (ECE = 0.030) compared to alternative approaches. The model demonstrates improved robustness to data variability and more uniform behaviour in terms of robustness metrics (mCE_norm and RS), which indicates an effective combination of local and global feature processing mechanisms. The hybrid architecture provides an optimal balance between accuracy, stability, and computational cost, outperforming most modern hybrid and traditional approaches. Further research prospects include the introduction of Neural Architecture Search, optimization for edge environments, extension of the model to multimodal data, and deepening the methods for interpreting decisions.
Keywords:	Computer Vision, Deep Learning, Hybrid Models, Convolutional Neural Networks, Transformers, Pattern Recognition, Robustness
DOI:	https://doi.org/10.5281/zenodo.20252688
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	IMPROVING THE HIGHER EDUCATION ENVIRONMENT THROUGH DETERMINISTIC ASSESSMENT: A CLOUD-INTEGRATED DECISION SUPPORT SYSTEM FOR ACADEMIC ACCREDITATION
Author:	MOHAMMED HUSSAIN ALHARBI, HAFEDH MAHMOUD ZAYANI
Abstract:	Traditional academic accreditation processes often suffer from subjective bias and non-reproducible outcomes due to a heavy reliance on holistic human judgment, which creates a critical data-integrity gap in quality assurance. This study introduces the Accreditation Standards Evaluation Platform (ASEP), a cloud-integrated Decision Support System (DSS) designed to resolve this problem by transitioning institutional evaluation into a deterministic, rule-based computational framework. At the core of the ASEP architecture is the decomposition of accreditation standards into seven measurable performance dimensions (E1 through E7), which are processed via a unique 'Sum-and-Compare' algorithmic logic. By utilizing a Binary Evaluation Vector (Vin) and a sequential matrix validation process—incorporating a dedicated consistency layer (E6 to cross-validate application and evidence)—the system eliminates the qualitative "black-box" nature of traditional assessments. The pilot implementation results demonstrate that the ASEP system achieved a 60% reduction in assessment processing time compared to manual methods. More importantly, the system's deterministic logic resulted in a 66% decrease in missing evidence and a 30% reduction in inflated compliance ratings, effectively bridging the 62% average discrepancy previously observed between self-assessment and independent audit scores. This research confirms that algorithmic determinism is a necessary evolution for achieving scalable, audit-ready quality assurance in higher education.
Keywords:	Decision Support Systems; Rule-Based Algorithms; Automated Assessment; Quality Assurance; Educational Data Mining; System Architecture
DOI:	https://doi.org/10.5281/zenodo.20252707
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	SECURE ENHANCED DUAL AUTHENTICATION IN MPSO FOR WSNS
Author:	Dr.M.SUPRIYA, Dr.T. ADILAKSHMI
Abstract:	Wireless Sensor Networks (WSNs) are essential in fields like military surveillance, industrial monitoring, and environmental observation. However, they are particularly vulnerable to different security attacks during Cluster Head (CH) selection and data aggregation due to their distributed nature and limited computational capacity. This work offers an improved and secure method for CH selection combined with the Low-Energy Adaptive Clustering Hierarchy (LEACH) protocol in order to overcome these difficulties. The suggested framework includes sophisticated security features as well as an effective way to identify malicious network nodes. The enhanced scheme eliminates nodes that exhibit unusual or malevolent packet-dropping tendencies during the CH selection stage by combining the traditional LEACH protocol with two evaluation metrics: Dropped Packet Ratio (DPR) and Residual Energy (RE). Particle Swarm Optimization (PSO) is used to further optimize routing efficiency, allowing for dependable path formation and better energy utilization. An Enhanced Dual Authentication (EnDA) method with a strongkey management system adds an extra layer of security that ensures data integrity and guards against unauthorized access. Simulations show notable improvements in network performance, security, and dependability.. Consequently, the suggested system improves the general security and dependability of WSN operations.
Keywords:	Residual Energy. Dropped Packet Ratio (DPR), Enhanced dual authentication (EnDA) , K-LionER ,PSO,Energy Consumed
DOI:	https://doi.org/10.5281/zenodo.20252741
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	IDENTIFICATION OF BRAIN TUMOR LESIONS THROUGH AN ADVANCED MULTI-MODAL METHOD: A TRANSFORMER-BASED FRAMEWORK AUGMENTED BY GAN FOR USE IN REAL-TIME CLINICAL SETTINGS
Author:	RAMESH BABU VURE, LALITHA KUMARI PAPPALA
Abstract:	The imperative for prompt and precise diagnosis directly influences patient outcomes, but medical imaging presents significant challenges in the detection and segmentation of brain tumor lesions. The existing methodologies exhibit several issues, including an imbalanced dataset, inadequate multi-modal data fusion, and an absence of real-time capabilities. An exhaustive framework for classification-deployment and comprehensive preprocessing, including segmentation, will be employed to achieve extremely robust and interpretable Tumor lesion identification. With this, we can surmount these challenges. To address data sparsity and class imbalance, CLAHE will initially enhance the images by GAN-based augmentation, followed by the application of adaptive histogram equalization with a local focus to enhance contrast. Our objective is to provide varied, authentic samples that yield favorable results for lesions. Model accuracy can be enhanced through simplification. A MMTN will provide the seamless integration of imaging and clinical data. By employing a self-attention mechanism, we will accomplish hierarchical feature fusion and establish accurate lesion categorization. Utilizing ensemble learning with cross-validation to amalgamate the predictions of many models enhances the model's robustness and applicability across diverse contexts. This results in an area under the curve (AUC) of 0.97 and a Dice coefficient of 0.89. The solution is now prepared for actual deployment with Grad-CAM visualization and Tensor Flow Lite. With an inference time of under half a second, we can generate predictions for clinical validation processes in a comprehensible manner. The proposed approach rectifies discrepancies in datasets, ensures practical applicability, and establishes a new standard for multi-modal integration. The domain of Tumor segmentation and diagnosis has seen significant advancements, with a remarkable 90% sensitivity and 93% accuracy rate.
Keywords:	Case Studies in Real-Time Clinical AI with Dense Net: Multi-Modal Transformers, Augmenting GANs, Tumor Detection the U-Net.
DOI:	https://doi.org/10.5281/zenodo.20252752
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	FEDERATED LEARNING IN LUNG CANCER RADIOMICS AND MEDICAL IMAGING: A META-SURVEY OF COLLABORATIVE MODELS
Author:	SRIVIDYA.CH , DR. RAMA SUBRAMANIAN K , MADHUBALA.M
Abstract:	Lung cancer is a major cause of cancer morbidity and death worldwide; new methods of testing are urgently needed but are not always available and accurate. Federated learning (FL) is an emerging paradigm in the machine learning field that shows significant potential in the context of medical data analysis -- namely, it allows for model training over a large number of distributed data sources where patient data can remain private. We note that this overview is focused on more recent works on federated learning methods for lung cancer-related imaging and screening tasks, with references chosen primarily based on a directly comparable methodology. Specifically, we review the research works with respect to different FL architectures, the most commonly used data preprocessing methods as well as the metrics for FL performance assessment in healthcare applications. Furthermore, we analyze the advantages of federated learning such as improved model generalization on heterogeneous data, and enhanced privacy protection. An overview of the FL strategies in terms of their centralized, decentralized, and hybrid architectures is discussed, along with the advantages and disadvantages for LC screening. It further explores a broader range of challenges such as non-IID (non-Independent and Identically Distributed) and IID data distributions, commutation costs, and stability of training, etc. We provide an extensive evaluation of the performance of various FL methods on random field studies compared with other experimental results reported in literature to benchmark the current state-of-the-art methods against each other. It also mentions the integration of federated learning with complementary techniques like (deep learning), (blockchain) to facilitate big scale cooperation learning in reducing overfitting and simultaneously mitigate the security problems due to deep learning, block chain respectively. Conclusion The findings indicate that federated learning could simultaneously improve performance and adapt populations for lung cancer screening under the constraints of stringent data privacy. Many critical issues remain, including those for non-IID data handling, communication efficiency, and scalability performance. In this paper, we clarify future directions (such as secure aggregation mechanisms, personalized federated learning models, and secure multiparty computation) to mitigate these problems. Thus, this work should provide an excellent reference for researchers and practitioners interested in applying federated learning to privacy-friendly lung cancer screening and diagnostic support systems. It covers the systematic review on 37 peer-reviewed studies, (2013–2025) and controlled benchmarking for 5 FL algorithms that is FedAvg(Federated Averaging), FedSGD(Federated Stochastic Gradient Descent), FedProx(Federated Proximal), FedAtt(Federated Attention) and FedEnsemble with respect to the Chest CT-Scan dataset (publicly available, 9,500 images, Kaggle). FedEnsemble achieved the highest optimally weighted score × class performance across multiple performance measurements; 92.0% accuracy, 91.9% F1-score, 91.5% precision and 92.2% recall suggesting it produced the most ideal overall performance, whereas, FedSGD had the lowest communication bandwidth requirement (i.e., 45 round and 180 MB of data), with an optimally weighted score × class of 80.0% accuracy.
Keywords:	Federated Learning, Lung Cancer Detection, Machine Learning, Privacy-Preserving, Decentralized Data, Deep Learning, Blockchain, Data Heterogeneity
DOI:	https://doi.org/10.5281/zenodo.20252765
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	DESIGN OF AN ITERATIVE MODEL FOR SELF-EVOLVING NEURO-FUZZY INTRUSION DETECTION WITH CONTEXTUAL REINFORCEMENT AND ADVERSARIALLY VALIDATED RISK CALIBRATION OPERATIONS
Author:	YAMARTHI NARASIMHA RAO, SATISH KUMAR PATNALA, Y PADMA, KUNDA SURESH BABU, LAKHMIKANTH PALETI, B VARAPRASADA RAO
Abstract:	The speed with which cyber threats are evolving has kept introducing new features for their counteraction intrusion detection systems (IDS) that traced intrusion detection in real time as well as changing conditions of network contexts and attack behaviors. Existing neuro-fuzzy IDS models are clear and adaptable, but they are characterized by static rules, low context sensitivity, and limited ability to withstand adversarial shifts. Those using reinforcement learning as the basis of their structures are quite adaptive but rely mostly on a clear explainability argument during the process of risk validation in process. This is based on Self Evolving Neuro-Fuzzy Intrusion Detection System with Contextual Reinforcement Adaptation, which is integrated with five very new analytical components in a closed feedback loop in the process. The Context Stamped Multiple View Telemetry Encoder will create uncertainty-aware latent representations which will be anchored to operational contexts for the most precise mode differentiation of attacks. The Differentiable Evidence-Weighted Fuzzy Rule Induction (DEW-FRI) makes use of these embeddings for generating sparse, uncertainty-weighted fuzzy rules for easier interpretability of the decision process. Contextual Reinforcement Rule Evolution (CoRRE) will adjust its structure and dynamic modification of threshold parameters useful for maximizing utility of detection under operational constraints. Adversarial Context Stress-Tester with Self-Play (ACSS) generates hard, context-consistent intrusion scenarios to expose rule vulnerabilities and to strengthen boundary robustness. Finally, the Conformal Risk-Calibrated Deployment Layer (CRiCDL) will give per-alert risk bounds and selective decision policies with guaranteed coverage, feeding accepted feedback for further evolution and evolving necessary. This architecture will handle better robustness, calibrated risk control, and fewer false positives while maintaining interpretability sets. Some preliminary evaluations show improvement in AUROC from 0.947 to 0.973, a 28% reduction in analyst alerts, and a decline of 41% in adversarial evasion rates. By coupling context-aware representation, interpretable fuzzy reasoning, reinforcement-driven adaptation, adversarial resilience, and statistically validated deployment, this work establishes a new paradigm for self-evolving, trustworthy intrusion detection sets.
Keywords:	Neuro-Fuzzy Systems, Contextual Reinforcement Learning, Intrusion Detection, Adversarial Robustness, Risk Calibration, Analysis.
DOI:	https://doi.org/10.5281/zenodo.20252782
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	ENHANCING RELIABILITY AND DATA RATE IN MOLECULAR COMMUNICATION FOR IOBNT USING MOLCOMMLSTM
Author:	OMAR KHALID SALIH ALHAFIDH, MAWADA MUHAMMAD SULIMAN , SADOON HUSSEIN ABDULLAH, SOHA MOHAMED HASSAN
Abstract:	Molecular communication (MC) in bio-nano things (IoBNT) is one of the promising paradigm for nanoscale networks, where information is encoded in the concentration of molecules. However, reliability in the MC system and data rate mapping remain challenges in achieving high efficiency and accurate symbol detection during communication between transceivers. In order to solve these challenges, a novel approach called MolCommLSTM is proposed, which combines an adaptive modulation technique and a deep long-short-term memory (LSTM) detector for symbol detection. The proposed MolCommLSTM system model dynamically adjusts molecular concentration levels based on the estimated distance between the transmitter (TN) and receiver (RN) nano-machines to optimize system performance. In addition, a deep LSTM directly maps the received signals to the transmitted symbols, facilitating efficient and accurate symbol detection. For theoretical comparison, a traditional detection algorithm called the maximum likelihood detector (MLE) is used. The MLE is implemented as a threshold-based decision rule where the detected signal strength is compared against a predefined threshold to infer the transmitted symbol. This approach efficiently demodulates the original signal by identifying the slot with the highest likelihood of a reaction, corresponding to the highest concentration of molecules. Extensive experimental results show that the proposed MolCommLSTM system model significantly improves the symbol error rate (SER) and achievable data rate (ADR) over the concentration position shift keying (CPSK) modulation scheme. Furthermore, the proposed MolCommLSTM system model effectively addresses the bias effect and minimizes inter-symbol interference (ISI) in the diffusion based molecular channel, resulting in superior performance compared to the MLE detector.
Keywords:	Reliability, Data Rate, Communication, IOBNT, MOLCOMMLSTM
DOI:	https://doi.org/10.5281/zenodo.20252815
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	INTELLIGENT POWER QUALITY IMPROVEMENT IN HYBRID RENEWABLE ENERGY SYSTEMS VIA STATCOM AND GREY WOLF OPTIMIZATION
Author:	S. V. R. LAKSHMI KUMARI , M. UMA VANI
Abstract:	This study presents an intelligent control approach to enhance power quality in a grid-connected hybrid renewable energy system integrating solar photovoltaic and wind sources. Such systems are highly susceptible to environmental variations, particularly wind speed fluctuations, which can reduce operational efficiency and stability. In addition, disturbances including three-phase faults and voltage deviations at the point of common coupling (PCC) may negatively influence system performance and reliability. To overcome these challenges, a Static Synchronous Compensator (STATCOM) is employed to provide dynamic reactive power support, thereby strengthening renewable energy integration and improving voltage regulation. Owing to the nonlinear and complex characteristics of hybrid systems, an advanced multi-objective Grey Wolf Optimization (GWO) algorithm is adopted to optimally tune controller parameters, enhancing robustness and overall system reliability. Simulation analyses under diverse operating conditions demonstrate that the system maintains voltage and current levels close to 1 pu during swell and sag events, ensures effective reactive power compensation during high renewable penetration, improves power quality under unbalanced nonlinear load conditions, and sustains PCC voltage within the range of 0.93–0.98 pu during three-phase faults. The results confirm notable improvements in voltage profile, current waveform quality, and Total Harmonic Distortion (THD), along with faster dynamic response, thereby validating the effectiveness of the proposed GWO-based STATCOM control strategy for hybrid renewable energy applications.
Keywords:	Static Synchronous Compensator (STATCOM), Grey Wolf Optimization (GWO), Reactive Power Regulation, Voltage Stability, Power Quality (PQ), Hybrid Energy Resource Systems (HRES).
DOI:	https://doi.org/10.5281/zenodo.20252884
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	DESIGN OF AN IMPROVED MODEL INTEGRATING FIREFLY-ANT COLONY OPTIMIZATION AND DEEP FEATURE FUSION FOR BRAIN TUMOR & LUNG CANCER DIAGNOSIS
Author:	P KIRAN KUMAR , B RAMA
Abstract:	Accurate diagnosis of brain tumours from the scans via magnetic resonance imaging for the rising trends of neurological conditions. Current methods frequently suffer from improper segmentation, ineffective feature description and insufficient robust classifiers. This resulted in a suboptimal approach towards the improved discrimination between benign or malignant cases, in general. We offer a new system that integrates the strengths of both Improved Hierarchical Macqueen's Firefly K Means and the KNN classifier, with methodologies based on nature and deep learning techniques. We further elaborate that the proposed pipeline is multi-stage and: (1) Bio-inspired Preprocessing and Segmentation utilizing Hybrid Firefly-Ant Colony Optimization (FFA-ACO), with high-precision tumor boundary extraction at 95% of Dice Similarity Coefficient. (2) A Multi-Level Feature Fusion Framework uses the firefly optimized fusion strategy integrating hand-crafted features including GLCM, LBP along with deep feature: ResNet-50; reaching an accuracy rate of 97%. 3. Bio-inspired Hyperparameter Tuning Firefly-bee Colony Optimizes for improving the efficiency and process in terms of achieving CNN. 2% -4% improved classification accuracies could be reached for some process. (4) An Integrated ML-DL-Bioinspired Ensemble Classifier (AWEC) is based on adaptive weighting of KNN, SVM, and DenseNet-121 outputs, achieving 98% classification accuracy. (5) Explainable AI with Glowworm Heatmap Visualization: It gives a 90% overlap on transparency with annotated versions. (6) Genetic Algorithm-Tuned GAN: This cross-modal synthetic data generation augments the dataset up to 50%. (7) Cross Modality Fusion with Transformer-Based Fusion: This results in a fusion of MRI, CT, and PET scans and an improvement in multimodal accuracy of 5%. This work moves forward the horizon of precision diagnosis by providing a holistic and interpretable framework to achieve high-reliability and robust outcomes, with transparent and improved clinical practices.
Keywords:	Bioinspired Optimization, Deep Feature Fusion, MRI Segmentation, Brain Tumor & Lung Cancer Classification, Explainable AI, Process.
DOI:	https://doi.org/10.5281/zenodo.20252901
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	QUASI OPPOSITIONAL LEARNING BASED AFRICAN BUFFALO OPTIMIZATION FOR ROBUST FEATURE SELECTION IN NETWORK INTRUSION DETECTION SYSTEM
Author:	PRASANNA KUMAR KRISHANAMOORTHY, SUBRAMANIYAN RADHAKRISHNAN
Abstract:	Detecting the anomaly network behavior is complex to establish the secure communication in network or system. The anomaly activities in networks seriously threaten the privacy of data, functions and the whole network infrastructure. The redundant, irrelevant and high-dimension features cause the overfitting issue in the learning process, resulting in high False Positive Rate (FPR) and less classification performance. To eliminate the redundant and high-dimension features, this article developed the Quasi Oppositional Based Learning (QOBL) strategy – African Buffalo Optimization (ABO) algorithm. The QOBL strategy is included in conventional ABO algorithm which improves the search ability and convergence rate for enhancing performance of feature selection process. In the classification phase, Recurrent Neural Network – Long Short-Term Memory (RNN-LSTM) technique is utilized to find intrusions in networks with high accuracy and less FPR. QOBL-ABO and RNN-LSTM based classifier obtained 99.99% accuracy on NSL-KDD and 99.94% on UNSW-NB15 dataset when compared to previous algorithms like Deep Neural Network (DNN).
Keywords:	African Buffalo Optimization, Long Short-Term Memory, Network Intrusion Detection System, Quasi Oppositional Based Learning and Recurrent Neural Network.
DOI:	https://doi.org/10.5281/zenodo.20253024
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	DIGITAL PLATFORMS FOR THE CAPTURE AND LONG-TERM PRESERVATION OF EVIDENCE OF WAR CRIMES IN COMBAT ZONES
Author:	VIKTORIIA ZARUBEI, YELZAVETA KUZMICHOVA-KYSLENKO, OLHA RYMARCHUK1, BOHDANA TYCHNA, EDUARD SOLOVIOV
Abstract:	Introduction: The relevance of the research was determined by the need to ensure the evidentiary admissibility, integrity, and long-term preservation of digital evidence of war crimes in combat zones in the context of fragmented platform architectures and increasing algorithmization of data collection. Aim: The aim of the study was to empirically substantiate and metrically verify the regulatory and technical framework of digital platforms for capturing and long-term preservation of evidence of war crimes in combat zones. Methods: The study was based on an interdisciplinary methodology that combined normative analysis and controlled technical experiments to close the gap between admissibility requirements and the architectural behaviour of digital platforms in combat zones for the empirical identification of configurations that ensure long-term evidentiary admissibility. Results: The study showed that the effectiveness of digital platforms for capturing and long-term preservation of war crimes evidence was determined by the architectural ability to maintain a continuous evidentiary loop. Class-specific quality losses (packet loss 3–17%, timestamp deviation ±2–9 s), provenance gaps with high cryptographic stability (HSI up to 100%), uneven detection of manipulations (MDR 61–94%), and a compromise between preservation and availability (DRR 92–99%) were identified. The synthesized framework demonstrated the highest consistency of technical and legal parameters (LCI ≈ 0.93; AS ≈ 0.92; PPR > 0.95), ensuring stable evidentiary admissibility. Academic novelty: The academic novelty consisted in the platform-centric metric linking of digital evidence capture and preservation with admissibility criteria through the LCI and AS indices and framework verification with the measured metrics HSI, PPR, MDR, and DRR. Conclusion: Further research may focus on in-situ validation of the framework in real judicial and investigative processes, expansion of metrics through explainability for AI components, and development of standardized benchmark sets for cross-jurisdictional comparison of admissibility. Open issues include: (i) how to ensure longitudinal stability of provenance under multi-year storage conditions; (ii) how to calibrate LCI/AS thresholds across jurisdictions with heterogeneous evidentiary standards; (iii) how to integrate XAI metrics without degrading operational efficiency; (iv) how to minimize ALM growth while maintaining DRR ≥95 %; (v) how to prevent provenance fragmentation in hybrid and decentralized architectures; (vi) how to standardize admissibility benchmarks for conflict-zone data variability.
Keywords:	Digital Evidence, War Crimes, Digital Platforms, Provenance, Admissibility of Evidence, Long-Term Preservation, International Criminal Law
DOI:	https://doi.org/10.5281/zenodo.20253080
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	AUTHENTICATING IOT DEVICES WITHOUT REVEALING THEIR RF FINGERPRINTS: A ZERO-KNOWLEDGE FRAMEWORK ON BLOCKCHAIN
Author:	YASSINE LKHALIDI , MOHAMED LKHALIDI , HATIM KHARRAZ AROUSSI , ACHRAF TIFERNINE
Abstract:	IoT device authentication must resist impersonation and credential theft while respecting the computational constraints of edge devices. Existing frameworks rely on static cryptographic keys that, once extracted, enable full impersonation, whereas RF fingerprinting schemes that bind identity to hardware imperfections transmit and store device templates in plaintext, exposing them to template theft and linkability attacks. ZK-RFAuth is a three-phase authentication framework that integrates Siamese neural network-based RF fingerprinting, Groth16 zero-knowledge proof (ZKP) verification, and proof-of-authority blockchain logging. During registration, a Siamese convolutional network extracts a compact embedding from raw I/Q samples and commits a Poseidon hash of the quantized mean template on-chain. During verification, the prover generates a Groth16 proof demonstrating that the L1 distance between a fresh embedding and the registered template falls below a per-device threshold without revealing either vector. The proof and authentication outcome are recorded on-chain for tamper-evident auditing. Evaluated on the WiSig dataset (28 WiFi transmitters, 224,000 frames), ZK-RFAuth achieves 91.4% closed-set accuracy and 2.25% equal error rate at embedding dimension d = 64, with 88.4% genuine acceptance rate and 70.8% open-set rogue rejection using per-device P95 thresholds. The ZKP circuit requires only 972 rank-1 constraint system (R1CS) constraints over 100× fewer than an equivalent SHA-256 circuit producing 144-byte proofs verifiable in approximately 3 ms.
Keywords:	IoT authentication, Zero-Knowledge Proofs, RF fingerprinting, Siamese network, Blockchain.
DOI:	https://doi.org/10.5281/zenodo.20253095
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	INTEGRATION OF ISLAMIC FINANCE, DIGITAL MARKETING, AND ARTIFICIAL INTELLIGENCE: FINANCIAL RISK AND INVESTMENT MANAGEMENT
Author:	HASSAN ALI AL-ABABNEH , IBRAHIM RADWAN ALNSOUR
Abstract:	This research addresses the issue of the unintegrated system between the digital marketing, artificial intelligence, and Islamic finance with financial risk management. Although financial services are being digitized rapidly, the current methods are not holistic and lack the explanations of the marketing tools in enhancing the reduction of risks within the framework of Shariah. The purpose of the research is to construct and empirically estimate a model that will assess the effect of digital marketing and AI on financial performance and investment decision-making in Islamic banks. The approach is a mixture of PESTEL analysis and panel regression model, and the Digital Marketing Performance Index (DMPI) is a compound measure of digital activity. The findings indicate that technological (β=0.45), legal (β=0.25), and DMPI (β=0.50) variables positively affect the return on investment and customer engagement the most. Digital marketing incorporation and AI result in an 8-12% ROI increase and enhances customer interaction. The research finds that digital marketing can be regarded as a strategic risk management tool, and not a mere communication tool, which helps to increase transparency, trust, and sustainability in Islamic financial institutions.
Keywords:	Islamic Finance, Digital Marketing, Financial Risk Management, PESTEL Analysis
DOI:	https://doi.org/10.5281/zenodo.20253111
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	ENHANCING COFFEE LEAF DISEASE DETECTION WITH RMFA-CNN: A REAL-TIME MULTI-FEATURE DEEP LEARNING FRAMEWORK
Author:	P.GOBINATH ,Dr.M.RAMASWAMI
Abstract:	Disease prediction in coffee plants has been widely investigated with several approaches utilizing diverse features and measures. However, existing methods often fail to achieve precise classification and are affected by high false prediction rates. The research problem has more impact on the crop of plants and achieving higher yields. The research problem is contributed with a novel approach which incorporates multiple features like intensity and texture of the leaf image towards prediction. Also, the model is designed with three levels of convolution layers to reduce the feature size and supports maximizing classification accuracy. To overcome these limitations, we propose a Real-Time Multi-level Intensity Feature Analysis based Convolutional Neural Network (RMFA-CNN) for efficient disease prediction in coffee plants. The model focussed on handcraft features to be extracted with dedicated schemes of preprocessing, segmentatitaon and feature extraction where the classification is performed with CNN. In the proposed framework, plant images are first pre-processed using a region centric diagonal normalization algorithm which traverses the entire image and enhances visual quality based on intensity features. Subsequently, a gray covariance segmentation algorithm is applied to partition the image into regions according to gray level characteristics. From the segmented regions, colour and texture features are extracted and transformed into a unified one dimensional feature vector for deep CNN training. During testing, the model estimates Intensity Disease Support (IDS) and Texture Disease Support (TDS) which are further combined to compute the Disease Class Support (DCS). Based on the DCS values, the system accurately predicts the disease class. The method is evaluated with two publicly available Arabica coffee leaf datasets, namely JMuBEN and JMuBEN2, which were acquired under real-world conditions at the Mutira coffee plantation in Kirinyaga County, Kenya. Experimental results demonstrate that the proposed RMFA-CNN significantly improves classification accuracy up to 98.6 and reduces false predictions thereby enhancing the reliability of coffee plant disease prediction.
Keywords:	Disease Prediction, Coffee Plant, RMFA-CNN, DCS, Intensity Disease Support (IDS) and Texture Disease Support (TDS)
DOI:	https://doi.org/10.5281/zenodo.20253119
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	VALIDITY AND RELIABILITY ANALYSIS OF A MECHANICAL EXECUTIVE COMPETENCY INSTRUMENT FOR THE HVAC INDUSTRY USING THE RASCH MEASUREMENT MODEL
Author:	MUHAMMAD NUR HANAFI HAROLANUAR, FAIZAL AMIN NUR YUNUS, SUHAIZAL HASHIM, ARASINAH KAMIS, AZMAN HUSSIN, DEDY IRFAN, RONAL WATRIANTHOS
Abstract:	The heating, ventilation, and air conditioning (HVAC) industry needs a workforce with technical skills, knowledge of the newest technologies, and problem-solving abilities. Without a clear competency model, graduates' or employees' ability levels are difficult to align with industry standards, which affects the effectiveness of HVAC systems that are critical in various sectors and the quality of services they provide. As a result, this model is the primary guide when designing curricula, training, and assessments to ensure that the HVAC workforce can meet industry demands consistently and competitively. Thus, this research is developed to examine the empirical evidence on validity and reliability of the survey item of the mechanical executive competency by using Rasch Measurement Model Analysis to make the constructed item applicable to the formal large scale research setting. A total of 30 students with HVAC background from University Tun Hussein Onn Malaysia involved in this study. Beforehand, this research item has been validated by three competency and research experts from various faculties. WINSTEP software version 3.69.1.1 was used to analyse the data, and the results showed that the survey item had great reliability across four constructs. Based on the Person Reliability of 0.91 and Item Reliability of 0.88, it was concluded that the constructed instrument is reliable and can be used on a large scale. It is anticipated that by assuring the instrument's strong validity and reliability, researchers would be able to use or adapt this excellent instrument in their research. Several aspects of students' technical and interpersonal skills are at a moderate level and the finding obtained are significant to provide useful input for policymakers, educational institutions, and the sewerage industry, and help address the issue of unemployment and skill mismatch in Malaysia as a whole
Keywords:	Competency, HVAC industry, Rasch Measurement, Validity, Reliability
DOI:	https://doi.org/10.5281/zenodo.20253130
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	A SEMANTIC IOT FRAMEWORK FOR PRECISION AGRICULTURE: INTEGRATING HETEROGENEOUS SENSOR DATA INTO AN ACTIONABLE KNOWLEDGE GRAPH
Author:	MANJU SADASIVAN , ASHOK KUMAR T A
Abstract:	The exponential growth of IoT sensors in agriculture generates vast heterogeneous data. However, the full potential of this data remains underutilized due to scattered sources and a lack of semantic interoperability. This work proposes a novel semantic IoT framework capable of converting unprocessed sensor data into a harmonized, queryable Knowledge Graph. The architecture integrates the data from various data sources, including vision, soil and environment sensors, into a cloud database. A Semantic Gateway will build a semantic model based on existing agricultural ontologies, thus enabling smart data integration and reasoning. The framework offers a user-friendly web interface that enables stakeholders such as farmers to ask questions in natural language. These questions are further translated into SPARQL to query the Knowledge Graph. Using the dataset from Kaggle as a simulated output of a vision sensor, we present the ability of our system to provide contextual answers for disease diagnosis and treatment recommendations. However, the deployment and validation of the proposed system in a real-world scenario is certainly reckoned to be the critical next steps. The framework was able to process more than 10,000 RDF triples from two weeks of simulated farm data. The system has achieved 94.2% precision in disease identification (Early Blight, Late Blight) of tomatoes by correlating visual symptoms to data from soil sensors. Nonetheless, a deep study and evaluation using wider range of crops and diseases is recommended in future to further improve the system. In complex querying, a semantic approach showed 38% quicker response times than conventional siloed approaches, and was able to make semantically pertinent treatment suggestions with 91.8% relevance. Results show that the proposed semantic-based approach enhances the usability of data and decision-support compared to traditional siloed systems.
Keywords:	Precision Agriculture, Semantic Web, Ontology, Knowledge Graph, IoT, SPARQL, Sensor Data Integration.
DOI:	https://doi.org/10.5281/zenodo.20253147
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	FEDERATED LEARNING BASED TRAFFIC INTELLIGENCE FOR PRIVACY PRESERVING DISTRIBUTED OPTIMIZATION OF WIRELESS NETWORK PERFORMANCE
Author:	J. RAVINDRA BABU, K. PRASUNA, Y. V. K. D. BHAVANI, PRASAD DEVARASETTY, R. SUDHA KISHORE, S. SINDHURA
Abstract:	The intensive development of wireless communication systems and data-intensive applications has created a very dynamic and unpredictable network traffic patterns, which are very much challenging to maintain Quality of Service (QoS). The classic network optimization methods that are based on the traditional and centralized methodologies tend to be inefficient to adapt to the real-time changes in traffic and brings forth the concern of scalability, privacy, and communication overhead. In solving these challenges, the present study suggests a federated learning-based traffic intelligence to privacy-preserving distributed optimization of wireless network performance. Various distributed nodes can cooperatively learn a global model in the proposed approach without providing raw data, thus preserving data privacy and minimizing communication cost. The framework combines the traffic prediction with adaptive network optimization techniques such as dynamically allocating bandwidth, load balancing, and congestion control. Experimental findings indicate that the proposed model can greatly enhance the performance of a network in comparison to traditional and centralized machine learning methods. Significant increases in accuracy of prediction, throughput, latency, packet loss ratio, and spectral efficiency have been noted. In addition, the framework ensures high utilization of the network to different traffic conditions, which underscores its scalability and flexibility. Generally, the suggested federated learning-based solution offers a scalable and efficient privacy-conscious solution to next-generation wireless networks, such as 5G and IoT networks.
Keywords:	Federated Learning, Wireless Networks, Traffic Prediction, Network Optimization, Privacy Preservation
DOI:	https://doi.org/10.5281/zenodo.20253164
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	EARLY DETECTION OF THROAT AND OROPHARYNGEAL CANCER USING YOLO-DRIVEN FEATURES AND RF–XGBOOST
Author:	SR AJITHA, DIPANWITA DEBNATH , CHALUMURU SURESH , DR M SURYA BHUPAL RAO , P.V.S MURALI KRISHNA , D.NIROSHA , J.SWETHA PRIYANKA , JOHN T MESIA DHAS , Dr SIVA KUMAR PATHURI
Abstract:	Oropharyngeal cancer, or throat cancer, is a significant health problem in the world, as it is usually diagnosed in the later stages when it is less treatable and the chances of survival of a patient is greatly reduced. This research offers a novel method of early diagnosis of throat cancer, through the combination of machine learning algorithms with medical data processing. Predictive oncology is a highly promising field of artificial intelligence use, which is difficult because of its application in various areas, including better patient outcomes and early diagnosis. This paper presents a new hybrid architecture named YOLO-RFXGB. The model is a combination of real-time object detection of YOLO (You Only Look Once) and the strong ensemble learning of RFXGB, which incorporates both a Random Forest and XGBoost Gradient Boosting algorithms. The proposed system utilizes a three-stage pipeline: (i) data acquisition and preprocessing, (ii) feature extraction and lesion localization with the help of YOLO, and (iii) classification and prediction of prognosis with the help of RFXGB. Its results, tested on multimodal medical images, are compared to the existing deep learning models like ResNet-101 and RegNetY. Experimental findings indicate that the proposed YOLO-RFXGB model greatly outperforms these baseline methods, with the high prediction accuracy of up to 96%. These results indicate that it has a great potential of clinical implementation in early-stage cancer screening systems as it is a good tool that can be used to enhance the quality of diagnosis and patient survival rate.
Keywords:	Throat Cancer, YOLO, RF+XGB, ResNet-101, RegNetY
DOI:	https://doi.org/10.5281/zenodo.20253174
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	FACT-CHECKER: A MULTIMODAL ARCHITECTURE FOR FACT CHECKING OF SOCIAL MEDIA ARTICLES IN INDIA USING A HYBRID ATTENTION-CNN AND LSTM
Author:	C. VISHNU MOHAN , R. CHENNAPPAN
Abstract:	The rapid rise of misinformation within India’s multilingual and multimedia digital environment has intensified the demand for advanced automated fake-news detection systems. Most existing solutions primarily rely on text-only analysis, failing to capture the critical relationship between textual statements and their accompanying images. Even current multimodal models that integrate visual information struggle with domain generalization and often lack robust attention mechanisms capable of identifying subtle yet meaningful visual inconsistencies. To overcome these limitations, this research introduces Fact-Checker, a hybrid multimodal architecture designed to enhance misinformation detection by jointly analyzing visual and textual cues. The proposed framework employs an attention-empowered Convolutional Neural Network (CNN) to identify distortions, manipulations, and salient patterns in images, while a Bidirectional Long Short-Term Memory (BiLSTM) network is used to extract contextual semantics from textual content. Both modalities are processed in parallel, transformed into discriminative feature representations, and subsequently fused at the feature level to support a unified classification process. The model is evaluated using the Indian Fake News Dataset (IFND), a culturally and linguistically relevant resource, ensuring strong applicability to Indian social-media contexts. Experimental results demonstrate that Fact-Checker achieves an overall accuracy of 81%, significantly outperforming existing unimodal and multimodal baselines. These findings underscore the effectiveness of integrating attention-based visual analysis with deep contextual text modeling for detecting misinformation. Furthermore, the proposed architecture establishes a scalable foundation for future enhancements, including multilingual adaptation, incorporation of additional modalities, and deployment across varied social-media platforms to strengthen the fight against digital misinformation.
Keywords:	Fake News Detection, Convolutional Neural Network, Attention Mechanisms, Residual Learning, Bidirectional LSTM.
DOI:	https://doi.org/10.5281/zenodo.20253252
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	CONDITIONAL MALWARE IMAGE GENERATION USING WGAN-GP FOR DATA AUGMENTATION
Author:	NGUYEN HOA CUONG , TRAN THI VINH , NGUYEN NGOC QUY
Abstract:	The efficacy of vision-based deep learning models in malware classification is frequently hindered by data scarcity and severe class imbalance. To address this critical challenge, this paper proposes the implementation of a Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) to synthesize high-fidelity, two-dimensional malware representations. We conditioned the network on five distinct classes: benign, spyware, trojan, virus, and worm. Designed to generate 224x224 RGB images, the model was trained on a structured subset of malware datasets. Empirical results over 60 training epochs demonstrate highly stable convergence and the effective elimination of mode collapse, a common flaw in standard GANs. Following the robust training phase, the model successfully generated a completely balanced dataset comprising 10,000 synthetic images (2,000 samples per class). This reliable data augmentation strategy provides a vital foundation for mitigating class imbalance, thereby improving the predictive accuracy and generalization capability of downstream deep learning-based malware detection systems.
Keywords:	Malware Detection, Generative Adversarial Networks, CWGAN-GP, Data Augmentation, Deep Learning, Vision-based Classification.
DOI:	https://doi.org/10.5281/zenodo.20253357
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	ADAPTIVE ERRP-GUIDED DEEP REINFORCEMENT LEARNING FOR BRAIN COMPUTER INTERFACES IN ASSISTIVE TECHNOLOGIES
Author:	DR. DIVYA NIMMA, DR GAURAV VISHNU LONDHE, DR. EDIGA CHANDRAMOHAN GOUD, R S S RAJU BATTULA, ELANGOVAN MUNIYANDY, VUYYURU LAKSHMA REDDY
Abstract:	Brain–computer interfaces (BCIs) are relying more on reinforcement learning; however, the current systems have challenges in unreliable error-related potentials (ErrPs), inability to combine multimodally, and in adapting to the drift during a session and new human users, which limit the safety and usability of the technology in real-life scenarios. Addressing these gaps is critical for advancing intelligent assistive technologies. The purpose of this study was to come up with a powerful, uncertainty-conscious, and dynamically flexible reinforcement-learning BCI system which optimally utilizes the ErrP feedback to enhance the control reliability. Our experimental analysis was based on heterogeneous multimodal data and we evaluated three models, namely Baseline (EEGNet and SAC), Hybrid (EEG, Transformer & Decision Transformer), and our proposed Deep Adaptive ErrP-Aware Reinforcement Learning (DAERL) system, which combines evidential uncertainty, multimodal fusion, and world-model imagination to optimize policies. Data consisted of EEG, EOG, EMG, gaze and short video frames which underwent graph, transformer and latent-dynamics processing. DAERL performed better than all comparison models with an ErrP AUC of 0.93, task-success rate of 90.3% and with normalized information-transfer rate of 17.6 bits/min, and critical-error rate back to 2.4 per hour, a 79 % improvement over the baseline. Further, DAERL had a high 12 trial adaptation half-life that exceeded the 120-trial baseline by far. The results indicate that uncertainty modelling, multimodal fusion, and world-model reasoning significantly improve the reliability and personalization of BCI, making DAERL a perspective framework to implement in the next-generation assistive neurotechnology.
Keywords:	Brain–Computer Interface, Reinforcement Learning, Errp Detection, Multimodal Fusion, Uncertainty Estimation
DOI:	https://doi.org/10.5281/zenodo.20253461
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	HYBRID CCO–SWO OPTIMIZATION FRAMEWORK FOR NET-ZERO SMART BUILDING ENERGY CONTROL
Author:	M. RAJKUMAR, Dr. V. GOKULA KRISHNAN, Dr. P. JESU JAYARIN, R. SENTHILKUMAR, Dr. KARNAM SREENU, Dr. S. KAVIARAAN
Abstract:	It is hard to get smart buildings to operate on net-zero energy mainly because energy consumption, occupant comfort, and renewable energy integration are nonlinear, multi-objective, and dynamic. Traditional optimization and rule-based control methods do not solve the problem effectively as they are less adaptive to changing environmental conditions and complex system interactions. As a result, they lead to suboptimal performance and increased reliance on the grid. In that regard, the present paper introduces an innovative hybrid metaheuristic optimization tool that combines the Centered Collision Optimizer (CCO) and Spider Wasp Optimization (SWO) for adaptive and real-time energy management. This new tool uses a variance-driven switching method to constantly adjust the balance between global exploration and local exploitation, thus preventing early convergence and improving the quality of the solution. Among other methods, the proposed CCO-SWO model has been shown to significantly outperform standard methods such as PSO, GWO, and HHO in EnergyPlus and MATLAB simulations. The report reveals a 31.5% drop in total energy consumption, a 41.3% less reliance on the grid, and an enhanced level of thermal comfort with a deviation of only 0.94C, whereas the high Net-Zero Index (NZI) of 0.96 was also maintained. These results prove the proposed framework as a highly efficient, reliable, and versatile approach for intelligent energy controlling in smart buildings strongly supporting sustainable and net-zero energy system deployment.
Keywords:	Smart Buildings, Net-Zero Energy, Hybrid Metaheuristic Optimization, Centered Collision Optimizer, Spider Wasp Optimization, Energy Management Systems, Sustainable Buildings.
DOI:	https://doi.org/10.5281/zenodo.20253533
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text

Title:	FORTIFYING FIREWALLS AGAINST EVOLVING DDOS ATTACKS WITH CONTRASTIVE AI AND LLMS
Author:	SRINIVASARAO DHARMIREDDI, DESIDI NARSIMHA REDDY, MRS. MODUGULA SIVAJYOTHI, MARISETTI KALYAN RAMUDU, ELANGOVAN MUNIYANDY, UDAY KIRAN KASI
Abstract:	Distributed Denial of Service (DDoS) attacks continue to threaten network availability and challenge firewall decision logic. Existing firewall-centric defences often fail to detect low-rate, multi-vector, and carpet-bombing DDoS patterns while meeting inline latency constraints. This paper proposes Contrastive-LLM Firewall (C-LLM-FW), a hybrid defence that integrates self-supervised contrastive representation learning with a distilled Large Language Model (LLM) encoder to produce context-aware flow representations for real-time classification and mitigation. The proposed method pretrained a contrastive encoder on benign flows to form a stable latent manifold and then used an LLM encoder as a contextualizer over short flow-sequence tokens; a compact cross-attentive fusion and a lightweight classifier issued firewall decisions while an enforcement agent applied actions in the Data Plane. Experiments were performed on the publicly available BCCC-cPacket-Cloud-DDoS-2024 corpus and on a simulated urban IoT emulation dataset for large-scale botnet scenarios. The method was compared to XGBoost as a classical baseline and to DoLLM as an advanced LLM baseline. C-LLM-FW improved F1 to 93.5% on the primary dataset, an absolute gain of 2.8 points versus DoLLM and 9.4 points versus XGBoost; inference latency reduced by ~28 ms on average and throughput resilience doubled at 1 Gbps attack injection. The results demonstrate that by leveraging the capabilities of contrastive AI alongside LLM encoders, practical gains in firewall detection accuracy and robustness can be achieved while maintaining production-grade inference latency that is ideal for deployment on the edge.
Keywords:	DDoS Detection, Contrastive Learning, Large Language Models (LLM), Firewall, Hybrid Model.
DOI:	https://doi.org/10.5281/zenodo.20253558
Source:	Journal of Theoretical and Applied Information Technology 15^th May 2026 -- Vol. 104. No. 9-- 2026
Full Text