		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
October 2025 | Vol. 103 No.19

Title:	AN IMPLEMENTATION OF ENHANCED INCEPTION-RESIDUAL CONVOLUTIONAL NEURAL NETWORK IN LUNG CANCER PREDICTION
Author:	Dr.S.PANDIKUMAR, Dr.S.PREMKUMAR, Dr.T.GUHAN, T.MARGARET MARY, Dr.M.KANNAN, Dr.V.PREAM SUDHA
Abstract:	Recent developments in Deep Learning are assisting in pattern recognition, classification, and quantification. Additionally, DL offers optimal performance across all industries, particularly in the medical field. For diseases to be treated and for humans to survive, precise diagnosis is essential. The primary goal of this study is to develop a useful tool for Computer Aided Diagnosis (CAD) systems to identify a suspicious lung nodule for cancer early detection automatically. Deep Learning is used to carry out this task. Using a filter, image acquisition and de-noising have been done at the first step. Segmentation and feature extraction have been accomplished in stage two. Finally, classification is accomplished using CNN methods in Deep Learning dubbed Enhanced Inception of Residual Convolutional Neural Network. Performance Evaluation metrics like accuracy, precision, recall, and f-measure are examined, and the results are also optimised. This method is contrasted with ML-E-RF and LR module 1, and it is ultimately determined that DL has a quick and precise diagnostic system.
Keywords:	CT-scan, CNN, Adaptive Wiener Filter, MV-CNN, EIRCNN
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	INTEGRATING COMPETENCY QUESTION-DRIVEN ONTOLOGY AUTHORING WITH ONTOLOGY VALIDATION FOR LEARNING ANALYTICS FROM HETEROGENEOUS PLATFORMS
Author:	SYAHMIE SHABARUDIN, SAZILAH SALAM, IBRAHIM AHMAD, MOHD HAFIZAN MUSA, NURFADHLINA MOHD SHAREF, DALBIR SINGH, WENDY HALL, ROHANA MAHMUD
Abstract:	Ontology authoring for learning analytics from heterogeneous platforms is a complex task, particularly for authors who may lack proficiency in logic. This paper introduces a novel approach that leverages competency questions (CQs) and test-driven development principles to streamline ontology, authoring and validation. We analyse common questions from stakeholders at 13 public universities to create competency questions, identify patterns, and utilise linguistic presuppositions to define ontology requirements. Our methodology ensures that these requirements are testable and can be validated, facilitating an integrated ontology for learning analytics. Additionally, we present a detailed ontology validation report, demonstrating the effectiveness of our approach through consistency checks, property validations, and individual test cases. This integrated method aims to enhance the accuracy and reliability of ontologies in representing learning analytics data from diverse platforms.
Keywords:	Competency Questions, Ontology Validation, Learning Analytics, Heterogeneous Platforms, Test-Driven Ontology Development
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	ENHANCING AIOT WITH COMMUNICATION-EFFICIENT FEDERATED LEARNING: A BLOCKCHAIN-ENABLED APPROACH FOR GREEN AND SECURE IOT SYSTEMS
Author:	MUNDLAGIRI PRAVEEN KUMAR AND, DR. C. DASTAGIRAIAH
Abstract:	The integration of Artificial Intelligence of Things (AIoT) with Federated Learning (FL) provides transformative capabilities for distributed intelligent systems. However, challenges such as excessive communication overhead, energy inefficiency, and security vulnerabilities limit the scalability and sustainability of AIoT deployments. This research proposes an innovative framework combining communication-efficient Federated Learning with blockchain-supported secure aggregation. The approach integrates gradient quantization, sketching techniques, and periodic averaging with lightweight blockchain consensus algorithms. Large-scale simulation experiments on benchmark datasets demonstrated up to 62% bandwidth savings, 55% reduction in communication rounds, 40% decrease in energy consumption, and improved model accuracy compared to existing FL approaches. The framework successfully enables green, secure, and scalable AIoT systems while conforming to sustainable AI principles and ensuring resilient collaborative learning in resource-constrained edge environments.
Keywords:	Artificial Intelligence of Things (AIoT), Federated Learning (FL), Blockchain Technology, Communication-Efficient Federated Learning, Green IoT.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	IMPROVING THE SECURITY OF THE IOT UTILIZING THE APPLICATION OF ROBUST CRYPTOGRAPHIC ALGORITHMS
Author:	DR. ANIMESH SRIVASTAVA, DR P PRABAKARAN, PANJAGARI KAVITHA, DR KARAKA RAMAKRISHNA REDDY, DR. PARMOD KUMAR, DR.S.SUMA CHRISTAL MARY
Abstract:	Securing Internet of Things (IoT) communications in real time has become a huge challenge, as these systems are widespread and vulnerable to cyber and quantum threats. Conventional cryptography techniques can place excessive computational penalties in place or not provide countermeasures to threats that are emerging in the post quantum scenario. Utilising algorithms such as, the post-quantum CRYSTALS-Dilithium technique for authentication, Advanced Encryption Standard (AES) for lightweight symmetric encryption, and Elliptic Curve Cryptography (ECC) for effective key exchange, this study offers a hybrid cryptographic framework. The solution is designed for IoT devices with limited resources that operate in real-time settings. To evaluate the model under several attack vectors, such as brute-force, MITM, and side-channel assaults, a simulation-based assessment was carried out using Contiki-NG and Cooja, together with real-world sensor datasets. Results indicate that the framework achieves over 95% detection accuracy, encryption latency under 1 ms, and energy usage below 5 mW per node, while maintaining low memory consumption. These findings show that post-quantum secure communication methods can be implemented in a scalable manner for constrained situations. The proposed solution provides an empirical trade-off between the performance and the cryptographic power skills, and develops a suitable candidate in next-generation management of IoT security that needs to perform and operate within the bounds of both traditional and quantum-era risks.
Keywords:	IoT Security, Cryptography, Lightweight Encryption, Post-Quantum Cryptography, Hybrid Algorithms
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	ENHANCING MALWARE DETECTION USING HYBRID FEATURE SELECTION TECHNIQUES IN PREDICTIVE MODELS
Author:	GAYATHRI DEVI N, V. KRISHNA, NEELIMA GURRAPU, RAJESH BANALA, SHAIK JILANI BASHA, KOMATI SATHISH
Abstract:	Malware detection remains a critical challenge in cybersecurity due to the growing complexity and diversity of malicious threats. Existing literature largely focuses on either filter-based or wrapper-based feature selection methods, often limited to specific malware categories or datasets. This study addresses this gap by introducing a hybrid feature selection approach that integrates statistical filter metrics with model-specific wrapper refinement, aiming to optimize feature subsets for enhanced malware detection. Publicly available datasets, including the Microsoft Malware Classification Challenge dataset and the Kaggle Malware Dataset, were used to evaluate multiple models such as Gradient Boosting and Neural Networks. Experimental results show that the hybrid approach consistently outperforms individual methods in terms of accuracy, recall, and computational efficiency, achieving up to 95% accuracy and 96% ROC-AUC. The findings contribute new knowledge by presenting a generalizable, scalable framework that improves malware detection performance across heterogeneous features and datasets.
Keywords:	Malware Detection, Hybrid Feature Selection, Machine Learning, Filter And Wrapper Techniques, Cyber Security
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	DESIGN AND IMPLEMENTATION OF ASSISTIVE TECHNOLOGIES UTILIZING BRAIN-COMPUTER INTERFACES
Author:	DR. V. V. R. MAHESWARA RAO , PRIYANKA R. RAVAL ,DR.M.V.RAJESH , DR. NIDHI MISHRA , DR GANTA JACOB VICTOR , DR. G. N. R. PRASAD
Abstract:	Current technologies using BCI-assistance tend to be confined within simulation settings or are not very usable or responsive to allow real-time practical usages of them. This paper fills this gap by introducing a non-invasive brain-computer interface (BCI) system based on the electroencephalogram (EEG) that allows individuals with motor deficits to operate the wheelchair, as well as smart home appliances, using motor imagery-based signals. The machine employs the Emotive EPOC++ + (14-channel) electroencephalogram signal collection headset. Preprocessing consists of bandpass filtering (8-30 Hz), 50 Hz notch filtering, and the correction of artefacts with the help of the Independent Component Analysis (ICA). Power Spectral Density (PSD) and Common Spatial Pattern (CSP) feature extraction, and a Convolutional Neural Network (CNN) as an implementation on PyTorch are used when classifying these features. The UART protocol is used to send the classified mental commands to an Arduino microcontroller that triggers the devices used. The system was proven to have large accuracy in classification and small latency, only proving the effectiveness of the system in real-time operations. Usability tests with participants yielded highly positive feedback on comfort, ease of use, and minimal training required. The novelty of this work lies in the fact that the EEG classification based on the deep neural network is incorporated into the functional assistive hardware integrated within the natural environment and can be deployed. The study adds a validated, real-time, end-to-end BCI system that links intent recognition through EEG to physical device use with a higher level of usability, performance, and practical feasibility than current state-of-the-art solutions.
Keywords:	Assistive Technology, EEG, Motor Imagery, Neural Signal Processing, Brain-Computer Interface
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	INTELLIGENT LAND SUITABILITY ANALYSIS UTILIZING MULTILAYER PERCEPTRON AND IOT SENSORS
Author:	PRIYANKA R. RAVAL , DR R SHALINI. , NARESH E , MOAZZAM HAIDARI , DR.HARIKRUSHNA GANTAYAT , HYMAVATHI THOTTATHYL
Abstract:	In contemporary agriculture, intelligent computational models are becoming increasingly popular in analysing land suitability to achieve greater precision and scale. The paper is a comparison of hybrid and advanced Multilayer Perceptron (MLP) architectures with rich Internet of Things (IoT) sensor data on a large, real-world data set (10,000+ samples, 50. The purpose is to rigorously test the robustness, accuracy, and computational speed of both MLP-based IoT systems in a practical agricultural environment. In high-dimensional, information-rich settings, the advanced MLP coupled with advanced IoT sensors (which comprises drone-acquired Normalised Difference Vegetation Index (NDVI)) achieves 92.4% accuracy, a 0.91 F1 score, and a 0.88 Matthews Correlation Coefficient (MCC), surpassing the hybrid model. The Hybrid MLP + Hybrid IoT Sensor, on the other hand, has a robustness score of 0.9 and operates well in handling noisy circumstances, real-time inference, and quick deployment. Both models facilitate practical and context-sensitive benchmarking so that the appropriate system can be chosen by the stakeholders. The study contributes to methodological practices in land suitability assessment and provides recommendations to scale to additional crops, areas, and sensors to promote data-informed and robust agricultural decision support.
Keywords:	Land suitability, Multilayer Perceptron, IoT sensors, Agriculture, Normalised Difference Vegetation Index (NDVI)
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	TELUGU NLP CHALLENGES AND METHODS: A SURVEY OF FILTERING, STEMMING, AND TRANSFORMER-BASED HATE SPEECH DETECTION
Author:	SANDEEP KUMAR MUDE , K YOGESWARA RAO
Abstract:	Telugu is a major Dravidian language with complex grammar, deep morphological structures, and flexible syntax. These linguistic features, while rich and expressive, pose significant challenges for natural language processing. This paper surveys the current landscape of Telugu NLP and presents a hybrid approach that integrates rule-based grammar modeling with machine learning techniques. The focus is on building efficient systems for text categorization, clause segmentation, grammar verification, and hate speech detection. One of the main challenges addressed is clause segmentation in compound and complex Telugu sentences. Due to implicit subjects and overlapping structures, traditional parsing methods struggle. The proposed solution uses syntactic pattern recognition and partial parsing based on subject-predicate matching, allowing for efficient segmentation without full syntactic trees. The system handles clauses with shared subjects and ensures syntactic agreement across dependent and independent components using a POS-based verification model. In addition to grammar checking, this study emphasizes the role of stemming in processing inflection-heavy languages. Telugu words often contain layers of suffixes that must be stripped to find the root. The paper compares linguistic rule-based methods with data-driven approaches and finds that hybrid models—combining affix rules with statistical frequency patterns—yield superior performance for classification and retrieval tasks. Hate speech detection and text classification are explored using multilingual transformer architectures. These include mBERT, XLM-Roberta, IndicBERT, and MuRIL. The models are fine-tuned using Telugu-specific datasets and adapted with regional tokenization strategies. The paper highlights how models trained on regional corpora offer more contextual understanding and semantic precision for detecting implicit or culturally nuanced hate speech. This survey consolidates computational models, linguistic frameworks, and evaluation benchmarks to demonstrate how rule-based strategies can effectively complement machine learning. It encourages more cross-lingual NLP research, especially for Indian languages lacking extensive digital resources. By presenting a linguistically grounded yet scalable system, the paper contributes both theoretical and practical tools for Telugu NLP research and applications.
Keywords:	Telugu Nlp, Filtering, Stemmig, Transformer, Hate Speech
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	EVALUATING THE ROLE OF BODY BIASING IN ENHANCING DOMINO LOGIC CIRCUITS FOR ADVANCED VLSI TECHNOLOGIES
Author:	Dr. SRINIVAS AMBALA , K. PURUSHOTHAM , VAKITI SREELATHA REDDY, J.MANGA , RAVINDRANADH JAMMALAMADUGU, KOMATIGUNTA NAGARAJU
Abstract:	Traditional complementary metal-oxide semiconductor (CMOS) technologies face scaling limitations, which have driven the search for innovative circuit-level optimisation methods to meet the rising demand for energy-efficient and high-performance microprocessors. Body biasing in domino logic is one such technique that holds promise for adjusting threshold voltage (Vth) and enhancing delay versus power modes under variable operating conditions. High-speed microprocessor datapath components utilise domino logic circuits, and this research examines how forward and reverse body biasing approaches influence them. The main aim is to investigate how body biasing optimises performance metrics such as energy efficiency, leakage power, dynamic power, and delay, while maintaining circuit reliability and noise immunity. Various body-biased domino logic configurations were modelled and compared under different supply voltages (VDD), operating frequencies, and process variations, using both HSPICE simulations and MATLAB/Python-based data visualisation. The experimental setup included logic units like adders, multipliers, and decoders. The comparison shows that RBB effectively reduces leakage but slows performance, whereas FBB significantly decreases delay at the cost of increased leakage power. A balanced trade-off between delay and power consumption exists within a tunable voltage window, which is scalable. Dynamic power consumption drops sharply at lower VDD levels; however, latency increases, which FBB can mitigate. This dual modulation enables adaptive energy control, making it suitable for modern microprocessors where workload scaling is crucial. The results demonstrate that domino logic design can effectively balance performance and efficiency in future technology nodes with the integration of body bias control schemes, making it suitable for low-power and high-speed VLSI applications.
Keywords:	Domino Logic Circuits, Microprocessor, Body Biasing, Dynamic Power Consumption, VLSI Application
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	EMPLOYING QUERY PERFORMANCE IN NOSQL DATABASES FOR APPLICATIONS UTILIZING BIG DATA
Author:	DR. RATNA RAJU MUKIRI, ASESH KUMAR TRIPATHY, DR.M.V.RAJESH, B RAMANA REDDY, ELANGOVAN MUNIYANDY, DR.S.SUMA CHRISTAL MARY
Abstract:	Traditional relational databases face challenges with scalability, schema flexibility, and real-time performance in the era of large data output from various sources, including social media, IoT sensors, and user-generated content. To get over this restriction, more and more people are turning to NoSQL databases, which are ideal for big data because of their distributed architectures and varied data models. The goal of this research is to examine and improve query performance in four prominent NoSQL systems: document-based MongoDB, column-family Cassandra, key-value Redis, and graph-based Neo4j under different real-world workloads. Through the use of real-world (Twitter, Stack Overflow, MovieLens) and artificial-intelligence (graph) datasets, a Kubernetes-orchestrated testbed was set up to measure execution time, throughput, latency, scalability, & resource utilization. Using instruments like as YCSB and Apache JMeter, the technique included controlled trials with read, write, update, aggregation, and difficult traversal queries. The results show that Redis always has the lowest latency as well as maximum throughput since it stores data in memory. This makes it perfect for real-time analytics. Cassandra is well-suited for workloads with a high volume of writes, as it scales efficiently for workloads with a lot of writes because it scales well. MongoDB, on the other hand, is great for a wide range of query types since it has good indexing. Neo4j is better at graph traversal jobs, but it has greater latency when there is a lot of traffic at the same time. Mathematical models back up the trade-off between execution time and throughput even further by revealing a significant link between dataset size and query complexity. This research adds a complete benchmarking methodology for comparing NoSQL systems that helps developers choose and tweak databases depending on how they will be used. In the context of processing massive amounts of data, it emphasises how important it is to make sure that database structures meet the needs of certain applications.
Keywords:	NoSQL, Query Performance, Big Data, Database Optimization, MongoDB, Cassandra.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	CAPACITOR AND SUPER CAPACITOR RELIABILITY ANALYSIS UTILISING NEO-FUZZY NEURAL LEARNING
Author:	DESIDI NARSIMHA REDDY , RAHUL SURYODAI , DR B.V.S.ACHARYULU , MYLAVARAPU KALYAN RAM , PUPPALA RAMYA , B SWARNA
Abstract:	As energy storage components are employed more and more in mission-critical applications such as smart grids, electric vehicles and aerospace systems, supercapacitors' reliability has emerged as the main issue. Both traditional physics-based models and deep learning (DL) techniques frequently fail in two areas: adaptability and real-time usefulness. This is particularly true when there are constraints on data and computational power. The Neo-Fuzzy Neural Learning (NFNL) method is used in this study to analyse the reliability and forecast the degradation of supercapacitors by utilising the interpretability for fuzzy logic as well as the adaptive learning capabilities of neural networks. Using time-series data, this study classifies supercapacitor health states across lifecycle stages and accurately predicts Equivalent Series Resistance (ESR) degradation. To train and validate the proposed NFNL model, a real dataset containing ESR values collected for 1500 hours within the accelerated ageing setting was used. The research employed both traditional Weibull-based statistical models and state-of-the-art models like Long Short-Term Memory (LSTM) and Support Vector Regression (SVR). The NFNL model achieves a classification accuracy of 91.2% and a Mean Absolute Error (MAE) of 0.111, outperforming all baselines. Error distribution study confirms the accuracy and stability of NFNL's ESR degradation curve, which closely matches the actual data. The NFNL framework finds the ideal balance of precision, readability and computing economy, making it perfect for condition tracking systems that operate in real-time. A trustworthy and interpretable prediction approach for supercapacitor health estimation is presented in this research, which has far-reaching implications for predictive maintenance as well as lifespan optimisation on embedded and resource-constrained systems.
Keywords:	Energy Storage Components, Supercapacitors, Equivalent Series Resistance, Fuzzy Logic, Neural Networks.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	ENHANCING CHATBOT RESPONSES THROUGH CONTEXT-AWARE NATURAL LANGUAGE UNDERSTANDING
Author:	DR. GUNDA SWATHI , ASHISH GUPTA , JATIN ARORA , LAVANYA KONGALA , DR.SHIRISHA DESHPANDE, ELANGOVAN MUNIYANDY
Abstract:	The development of intelligent conversational agents has brought significant improvements in user-machine interaction; however, most existing chatbots still struggle with maintaining contextual coherence and delivering relevant responses in multi-turn dialogues. In this research, a realisation of a context-aware Natural Language Understanding (NLU) framework is proposed, which increases the competence of chatbots in processing user intents and entity recognition by increasing semantic awareness. The system maintains historical context and guarantees continuity across conversation turns by merging pre-trained transformer-based models like BERT and Sentence-BERT with a bespoke dialogue context encoder. The architecture uses a combination of human-centric assessments (appropriateness, contradiction rate) and automated metrics (BLEU-4, ROUGE-L, contextual coherence score) to evaluate performance thoroughly. Experimental results on benchmark datasets indicate that there is a great mark of improvement of the proposed model with respect to conventional single-turn NLU systems and the baseline. It improves in terms of accuracy, coherence, and lowered hallucination. In entity recognition, it obtained a 0.88 F1-score, a 0.85 SBERT-based semantic similarity score, and an intent detection accuracy of 92.1%. Additionally, there was a 25% boost in human-rated contextual appropriateness on top of the baseline. Further evidence of reduced conflicts, enhanced entity tracking, and increased alignment with user expectations comes from a thorough mistake analysis. The results show that the method works well for practical, ever-changing uses like AI assistants and customer service. To further human-centric conversational AI, this work shows how dialogue memory techniques and contextual embeddings might make conversations more meaningful and logical. Based on these findings, future work will focus on multilingual adaptability, low-resource optimisation, and deployment in edge environments.
Keywords:	Chatbot Response, Natural Language Understanding, Transformer Model, Semantic Score, Optimisation
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	ENHANCING COMMUNICATION PROTOCOL DESIGN FOR ENERGY CONSERVATION IN IOT NETWORKS
Author:	PANYA MANOJ SHARMA ,DESIDI NARSIMHA REDDY ,KALYANAPU SRINIVAS , MUKESH MADANAN , ELANGOVAN MUNIYANDY , A.SMITHA KRANTHI
Abstract:	The explosion in Internet of Things (IoT) installations has connected millions of low-power devices, which poses serious problems for maintaining reliable, energy-efficient connectivity in diverse settings. Traditional hybrid optimisation approaches like Particle Swarm Optimisation (PSO)-Low Energy Adaptive Clustering Hierarchy (LEACH) integrated with Random Forest (RF) improve clustering and traffic forecasting. Still, they struggle with the dynamic, heterogeneous traffic of smart cities, industrial IoT, and environmental monitoring. Emerging Deep Reinforcement Learning (DRL) algorithms, particularly Proximal Policy Optimisation (PPO), enable real-time routing and Medium Access Control (MAC) regulation but have not been extensively compared with hybrid algorithms in unified, realistic scenarios. This paper presents a thorough, application-oriented comparative analysis of a DRL protocol based on PPO and a PSO-LEACH + RF hybrid system, both of which are designed for energy-aware IoT communication. Both methods were evaluated using the same simulation scenario with consolidated traffic data reflecting the periodic, event-driven, and bursty nature of environmental monitoring, smart home automation, and industrial IoT applications. Key metrics include energy consumption, throughput, average packet delay, and packet delivery ratio (PDR). Results show that the hybrid model used 0.2–1.5 J less energy, while the PPO-DRL model achieved up to 23 kbps higher throughput, 8–10 ms lower latency, and up to 2% higher PDR.
Keywords:	Internet of Things (IoT), Deep Reinforcement Learning (DRL), PSO LEACH, Energy Efficiency, Quality of Service (QoS)
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	NUMERICAL METHODS FOR SOLVING NONLINEAR DIFFERENTIAL EQUATIONS IN INFORMATION NETWORK SECURITY PROBLEMS
Author:	MARYNA BELOVA, VOLODYMYR DENYSENKO, SVITLANA KARTASHOVA, VALERIJ KOTLYAR, STANISLAV MIKHAILENKO
Abstract:	The increasing complexity and dynamism of modern information networks makes the problem of their resistance to threats increasingly important. The use of differential equations, in particular, variants of epidemiological models, is one of the promising approaches to simulating the spread of harmful effects in such networks. This study proposes the use of numerical methods for solving nonlinear differential equations for modelling the dynamics of infection under different scenarios of cybersecurity threats. A scalable information network with a dynamic topology based on a stochastic block model is the basis of the experimental environment. The aim of the research is to determine the most effective numerical methods for modelling the spread of threats in information networks, taking into account accuracy, speed, and resistance to changes in parameters. Generalized models of the MeanField type were used to describe the spread of influence — both the basic one and its four nonlinear variations with exponential, logarithmic, quadratic, and power dependence, respectively. The models were solved using a wide range of numerical methods: classical adaptive methods (RK45, RK23, Radau, BDF, LSODA), as well as self-implemented schemes (Adams-Bashforth, Adams-Moulton). Large-scale experiments were conducted with varying network parameters (size, intensity of connections), initial conditions, model parameters, and integration step. The analysis was carried out using such metrics as accuracy (RMSE, Max Error), efficiency (execution time), and sensitivity to parameters. The obtained results gave grounds to determine the advantages of specific methods for different types of models and levels of system complexity. The prospects for further research include expanding models to multi-level networks, including stochastic components, and developing intelligent systems for choosing a numerical method in real time.
Keywords:	Nonlinear Differential Equations, Numerical Methods, Epidemiological Modelling, Information Security, Dynamic Network.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	ADAPTIVE CROWD FEEDBACK STRATEGIES FOR IMPROVED WIRELESS NETWORK EFFICIENCY
Author:	KROVVIDI S B AMBIKA, M.SELVI, PALLAVI SACHIN PATIL, V. RAMA KRISHNA, N.C. KOTAIAH, AMIT VERMA
Abstract:	Wireless networks today face increasing performance challenges due to dynamic conditions like user mobility, fluctuating density and diverse application demands. Conventional approaches, such as static resource allocation and offline machine learning (ML), lack the adaptability to respond effectively to real-time variations. To address these limitations, this research presents an Adaptive Crowd Feedback Strategy that combines the live, trust-filtered user feedback with a closed-loop optimisation system. The suggested framework includes 4 core modules: feedback collection, trust filtering, a Bayesian reinforcement learning (RL) engine and network control reconfiguration. Researchers use mathematical models to combine Quality of Experience (QoE) as well as Quality of Service (QoS) metrics, implement Bayesian inference to make policy changes and queueing theory to predict how the network will perform. Many real-world and fabricated datasets, like more than 18,000 mobile session logs, were used in the simulations. When contrasted with static as well as offline ML-based systems, the results show big performance improvements, with up to 35% more throughput, 30% less latency and over an additional 20% of energy efficiency. The adaptive system also achieves quicker convergence, making it highly responsive to changing network conditions. Comparative evaluation highlights the system’s ability to maintain a higher packet delivery ratio and minimal congestion by smarter, feedback-driven decisions. Practical issues such as computational trade-offs, feedback dependability and scalability are highlighted in the discussion of the results. The promising uses in forthcoming 5G/6G, smart city and edge computing infrastructures, the research finds that the suggested adaptive model improves real-time network performance while also laying the groundwork for smart, user-aware, as well as energy-efficient wireless communication.
Keywords:	Wireless Network, Machine Learning, Bayesian Inference, Packet Delivery Ratio, Congestion, Energy Efficiency.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	ADAPTIVE DEEP LEARNING FRAMEWORK USING BAYESIAN OPTIMIZATION FOR AUTISM SPECTRUM DISORDER PREDICTION FROM SCREENING DATA
Author:	B. DEEPA , Dr.K.S. JEEN MARSELINE
Abstract:	Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition marked by persistent challenges in communication, behaviour regulation, and social interaction. The heterogeneity of symptoms across individuals and age groups complicates early detection, as behavioural traits often overlap with other conditions or remain masked until later developmental stages. Traditional diagnostic methods are usually time-intensive, subjective, and rely on specialist interpretation, leading to delayed or inconsistent identification. Screening data offers a scalable, cost-effective, and non-invasive alternative for early ASD prediction, capturing observable traits through structured behavioural questionnaires. To overcome diagnostic inconsistencies and optimize model performance, this research proposes a Bayesian Optimization based Long Short-Term Memory (BO-LSTM) framework that adaptively learns temporal dependencies in screening responses while automatically tuning its parameters using a probabilistic surrogate model. The model was evaluated using the Autism Screening Dataset, comprising 6075 records and 20 structured attributes, sourced from a mobile-based application developed by Dr. Fadi Fayez. The dataset includes behavioural inputs from toddlers, children, adolescents, and adults, with questionnaires tailored to each age group. BO-LSTM achieved a classification accuracy of 74.375%, along with notable gains in sensitivity, specificity, and interpretability. These results demonstrate the framework's effectiveness in processing sequential screening data for timely and reliable ASD prediction across diverse age groups.
Keywords:	Autism Spectrum Disorder, Prediction, Screening Data, Deep Learning, Long Short-Term Memory, Bayesian Optimization
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	LIC-NET: LIGHTWEIGHT INTEGRATED CONVOLUTIONAL NETWORK FOR ACCURATE POLYP SEGMENTATION
Author:	SUCHITRA A. PATIL, CHANDRAKANT GAIKWAD
Abstract:	The segmentation of a small polyp present in intestinal regions, which tend to be malignant, is a basic and essential task for the detection of colon cancer. Segmenting a small polyp is challenging due to the higher similarity of tissues. Inaccurate segmentation results in higher false positives for colorectal cancer classification. This work suggests a two-stage deep learning network for polyp segmentation. In the first stage, the colonoscopy image is preprocessed to generate salient regions. In the second stage, the salient regions are processed by a novel U-Net structure network called LIC-Net, which integrates transfer learning and multiscale feature extraction to increase the segmentation accuracy. Testing with Kvasir-Seg and CVC-ClinicDB, both in direct and cross-learning mode, the proposed solution achieved more than 90% accuracy. The false positives are at least 2% lower compared to the most recent deep learning based segmentation works.
Keywords:	Convolutional Neural Network, Fuzzy-TAN, Colonoscopy, Dilated Convolution, Polyp Segmentation, U-Shaped Model
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	A REVIEW OF THE ECONOMIC SOCIOLOGY OF DIGITAL MARKETS: THE DYNAMICS OF SOCIAL INTERACTION IN E-COMMERCE PLATFORMS
Author:	M. RASYID RIDHA, ILHAM SAMUDRA SANUR, REZKY JUNIARSIH NUR, AHMADIN, M. YUNASRI RIDHOH
Abstract:	This research seeks to define, delineate, and examine the evolution of trends in the Economic Sociology of Digital Markets, emphasizing the dynamics of social interaction inside e-commerce platforms. Social connection is demonstrated to be a vital factor that affects customer purchasing intents and decisions, both directly and indirectly. Social aspects, including trust, community participation, feedback, online reviews, and emotional experiences, are increasingly intricately intertwined within social commerce and live-streaming functionalities. Elements such as reliable followers, electronic word-of-mouth (eWOM), and the streamer's social presence significantly influence users' impulse purchasing behavior and emotional involvement. Conversely, social interactions that are not appropriately tailored to consumers' psychological situations, such as loneliness, might diminish loyalty and purchasing inclinations. This study employs a systematic literature review methodology utilizing a bibliometric approach to examine the trajectory of research progress. On April 17, 2025, 29 items were retrieved from the Scopus database using the keywords “social interaction” and “e-commerce platforms.” All data were examined utilizing the Analyze Search Results function on Scopus.com and visualized with VOSviewer software version 1.6.20. Twelve pivotal studies from 2019 to 2025 were acquired through the selection process, directly reflecting the dynamics of social interaction within the realm of e-commerce. Bibliometric analysis shows a sharp increase in the number of publications, peaking in 2024, driven by global collaboration, increased funding, and better access to digital data. By viewing the digital market as an integrated social environment, this research contributes to information technology (IT) by explaining how the dynamics of social interaction on e-commerce platforms critically shape consumer behavior and support a sustainable digital ecosystem.
Keywords:	Economic Sociology, Social Interaction, E-Commerce Platforms, Systematic Literature Review
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	TEST CASE COVERAGE ANALYSIS USING CLUSTERING METHODS FOR TEST CASE MINIMIZATION IN SOFTWARE TESTING
Author:	SANJAY SHARMA, JITENDRA CHOUDHARY, GOVINDA PATIL
Abstract:	In test case minimization especially in software testing large-volume datasets are minimized the purpose is to separate unnecessary and duplicate datasets. The coverage of minimized datasets will be the same as original data sets. At present thousand techniques are available for dataset minimization and clustering is one of the main. The objective of the proposed paper is to justify different clustering methods, test case coverage and their role of in software testing, apart from the reason for choosing particular methods. As initial stats all clustering methods require data sets, data sets are generated by programmers or by automation tools. Duplicate data sets are the by-product of automation tools. This paper is based on an analysis of different test case coverage and test case minimization especially using clustering methods.
Keywords:	Test Case Minimization, Coverage Analysis, Data Mining, Clustering, Test Case Coverage
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	ENHANCING GESTATIONAL DIABETES MELLITUS PREDICTION USING WHITE TIGER SWARM OPTIMIZATION-ENHANCED MULTILAYER PERCEPTRON (WTSO-MLP)
Author:	D.SHOBANA , V VINODHINI
Abstract:	Gestational Diabetes Mellitus (GDM) is a critical pregnancy-related complication that affects both maternal and neonatal health. Although commonly used for diagnosis, the Oral Glucose Tolerance Test (OGTT) is invasive, time-consuming, and fails to provide early detection. Inconsistent screening guidelines further complicate the identification of high-risk pregnancies, emphasizing the need for more accurate and timely predictive tools. This research develops a robust, non-invasive prediction model for GDM risk using White Tiger Swarm Optimization-enhanced Multilayer Perceptron (WTSO-MLP). The goal is to enhance early detection by integrating bio-inspired optimisation techniques to improve model performance while reducing dependency on invasive tests, such as OGTT. The WTSO-MLP model combines White Tiger Swarm Optimisation (WTSO) with Multilayer Perceptron (MLP) to optimise weight configurations, trained on a dataset of 3,525 instances that contain clinical and demographic data. Class imbalance has been addressed through adaptive techniques. Model performance has been evaluated using the Matthews Correlation Coefficient (MCC), Error Rate, Youden’s Index, and Critical Success Index (CSI). This study contributes new knowledge by demonstrating how a bio-inspired optimization strategy can simultaneously refine neural network parameters and feature subsets using prospectively collected data, achieving superior accuracy, early detection capability, and adaptability across diverse clinical settings. The WTSO-MLP model outperformed traditional methods, achieving high performance in GDM prediction, especially for early-stage detection. The model demonstrated improved generalization, reduced misclassifications, and higher MCC scores, making it a reliable tool for clinicians. The WTSO-MLP model provides an innovative, efficient solution for early GDM risk prediction, improving diagnostic accuracy, generalization, and interpretability. It can seamlessly integrate into clinical workflows to enable early, non-invasive GDM assessments, ultimately enhancing maternal and fetal health outcomes.
Keywords:	Healthcare, Diabetes, GDM, WTSO-MLP, Swarm Optimization
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	ENHANCING HEALTHCARE SECURITY WITH BLOCKCHAIN-POWERED SMART CONTRACTS
Author:	DR. NAIM SHAIKH, DR. MAMATHA G, KUKATI ARUNA KUMARI, DR. M.S.GIRIDHAR, DR. KRISHNA NAND MISHRA, DR. A.PANKAJAM, DR. GANESH KUMAR R, SWATI GUPTA
Abstract:	The rationale behind this research stems from the increasing frequency of data breaches in healthcare and the inadequacy of centralized systems to ensure privacy, interoperability, and regulatory compliance. The Present study emphasizes the importance of applying security in healthcare. This model was prepared by utilizing Smart Contracts. It has been noted that there are some emerging concerns about data security and privacy as well as interoperability within healthcare organizations. The focus of a research paper is on the deployment of Smart Contracts along with blockchain technologies. The fundamental vision is to improve healthcare infrastructure’s security. Blockchain is transforming healthcare systems for the better by eliminating inefficiencies caused by fraud and outdated technologies, allowing for the efficient, transparent, and secure issuance of Smart Contracts. The challenges of confidentiality, data security, and access to relevant patient information for medical professionals have been a problem in the healthcare sector. Most of the existing EHR systems do not have adequate mechanisms for enforcing security access controls, which hampers cooperation between healthcare institutions. These security concerns pose risks for patients’ privacy and cripple the adoption of modern information technology within the health sector. Simulation works shows that Transaction processing time in case of proposed model is below 1.5 second where as it is 2.5 in case of conventional model. Security breach probability of proposed model has been reduced to 0.05 that was 0.35 in case of conventional model. Data integrity verification time in case of proposed model is below 1.0 that is above 1.75 in case of conventional model. While with the existing Electronic Health Record (EHR) systems face limitations in security, privacy enforcement, and interoperability, this study addresses the lack of automated, decentralized access control mechanisms. It proposes a blockchain-powered Smart Contract model to fill these gaps and enhance healthcare data governance and trust.
Keywords:	Blockchain, Smart Contract, Digital transformation, Healthcare system, Data security.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	MACHINE LEARNING-DRIVEN IMPROVEMENTS IN SOFTWARE DELIVERY PIPELINES
Author:	RAYAVARAPU SRIDIVYA,BEHARA VENKATA NANDAKISHORE,DR VED SRINIVAS, DR. R. NIRUBAN, RAJESH TULASI, AMIT VERMA
Abstract:	Modern software delivery relies heavily on Continuous Integration as well as Continuous Deployment (CI/CD) pipelines, yet these processes are frequently hindered by inefficiencies, including longer build times, faulty testing, and inefficient use of resources. This paper introduces a novel framework combining Graph Neural Networks (GNNs) and Reinforcement Learning (RL) to intelligently optimise CI/CD processes in DevOps environments. Generalised neural networks (GNNs) accurately forecast failure sites and bottlenecks by modelling the structural relationships among activities in a pipeline. RL agents dynamically learn optimal scheduling policies that reduce build time and resource usage through adaptive feedback mechanisms. The main contributions include 1) an RL structure that dynamically assigns computational resources to minimise build latency and test flakiness, 2) a hybrid GNN+RL architecture that collaboratively improves both predictive accuracy as well as decision-making efficiency, and 3) a GNN-based method for identifying critical path nodes as well as failure-prone components in the pipeline graph. Through experimental assessment on real-world DevOps pipelines, in comparison to baseline and RL-only techniques, it can reduce CPU consumption by 25%, increase failure prediction rate by 49%, and decrease build time by 38%. To create intelligent, self-healing CI/CD infrastructures, our study connects the gaps between optimisation driven by AI and actual DevOps.
Keywords:	CI/CD, DevOps, Graph Neural Networks, Reinforcement Learning, Pipeline Optimisation, Failure Prediction, Resource Allocation
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	CONTINUAL LEARNING-DRIVEN LAPLACIAN PALE TRANSFORMATIVE CONVOLUTION NETWORK FRAMEWORK FOR DIABETIC RETINOPATHY DETECTION FROM RETINAL FUNDUS IMAGES
Author:	K.S.NALINI, Dr.ARUNACHALAM A.S
Abstract:	Diabetic retinopathy (DR) is a diabetes-related eye ailment caused by retinal blood vessel (BV) damage. This manuscript presents a novel continual DL framework for efficient DR disease detection. Initially, input retinal fundus images are taken from the IDRI Dataset for accurate DR disease detection, which undergoes the preprocessing stage by employing a Color Wiener Filter (CWF) that can enhance image clarity by adaptively removing noise while maintaining edge details for further processing. After preprocessing, a novel Laplacian Pale Transformative Convolution network (LPTCN) is introduced, which classifies with more accuracy the distinction between different DR abnormalities. Moreover, the proposed framework integrates Elastic Weight Consolidation (EWC) and Herding Selection Replay (HSR) to prevent catastrophic forgetting on new data samples. The proposed framework is simulated in the Python platform. In the simulation part, the average Accuracy of 97%, Matthew’s Correlation Coefficient (MCC) of 0.936, symmetric mean absolute percentage error (SMAPE) of 2.07, and Computation Time (CT) of 4.9s, Youden’s index (YI) of 0.89 are obtained by the suggested framework on DR identification.
Keywords:	Diabetic Retinopathy (DR), Retinal Fundus Image (RFI), Multi-class Classification, Image Preprocessing, Continual Learning, Laplacian Pale Transformative Convolution Network.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	SMART MEDPLANT: A HYBRID DEEP LEARNING AND RANDOM FOREST MODEL FOR GEOTAGGED MEDICINAL PLANT DETECTION
Author:	BEULAH JABASEELI .N, D. UMANANDHINI
Abstract:	Accurate and scalable medicinal plant identification is critical for biodiversity conservation, pharmaceutical development, and ethnobotanical research. Traditional methods, reliant on expert knowledge and manual morphological analysis, are often labor-intensive, error-prone, and limited in scope. To overcome this issue, our research work introduces a hybrid model, TLBR101RF(Transfer Learning Based ResNet101 with Random Forest), to find the medicinal plants from it leaf image with high accuracy. The model employs ResNets’s deep residual blocks for effective hierarchical feature extraction and Random forest is used for classifying the result based on robustness against over-fitting and promote generalization on imbalanced datasets. The process consist of data acquisition , data augmentation by GAN , segmentation by LeGrNet and further involved in Classification of leaf image. TLBR101RF achieves a high classification accuracy of 98.69%, when compared with CNN models like VGG16, MobileNet and Inception. The incorporation of geolocation metadata improves classification accuracy by accounting for regional morphological changes influenced by environmental factors. A GPS-enabled web application was developed for real-time species identification, relevant to conservation, agriculture and pharmaceutical research. In TLBR101RF, Deep learning and Artificial Intelligence are combine to form best solution for identifying medicinal plants accurately. The results indicate that TLBR101RF effectively categories medicinal plants with high accuracy of 98.61%, while comparing with CNN models like MobileNet (97.7%), ResNet50(97%), VGG16(96.5%) creating an effective application for identifying correct medicinal valued plants from it localization by achieving high accuracy of 98% by using ResNet101 and Random Forest algorithms.
Keywords:	Pharamaceuticals, Transfer Learning, ResNet101, Random Forest, localization
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	QUANTUM CRYPTOGRAPHY FOR SECURE CLOUD DATA STORAGE AND TRANSMISSION
Author:	RAHUL SURYODAI, DESIDI NARSIMHA REDDY, DR. HEMLATA MAKARAND JADHAV, ANIL KUMAR MUTHEVI, DR. V. V. R. MAHESWARA RAO, V SIVARAMARAJU VETUKURI
Abstract:	Traditional encryption algorithms like RSA and ECC are increasingly vulnerable due to advances in quantum computing, which enable attacks through techniques such as Shor’s and Grover’s algorithms. To address this challenge, the researchers proposed a hybrid encryption system that integrates Quantum Key Distribution (QKD) with the Advanced Encryption Standard (AES-256) to ensure secure data transmission and storage in cloud environments. The system employs the BB84 protocol over a virtual 50 km quantum channel for key generation and distribution. Additionally, introduce the Hybrid Secure Transmission Protocol (HSTP) that rotates session keys every 5 seconds, enhancing security through continuous key renewal. By adapting the Key Management Service (KMS) to utilise QKD-generated keys, this approach is compatible with popular cloud platforms such as AWS S3 and Google Cloud Storage. Experimental evaluations comparing AES+QKD with conventional AES+RSA demonstrate that AES+QKD achieves higher key generation rates (>1050 MB/s), superior data consistency (99.9%), and maintains low latency under heavy workloads, while effectively resisting both classical and quantum attacks. This work presents a scalable, quantum-safe cloud security architecture, showcasing the practical integration of QKD in large-scale cloud infrastructures through the novel HSTP protocol and validated performance models.
Keywords:	Quantum Key Distribution (QKD), AES-256 Encryption, Cloud Data Security, Post-Quantum Cryptography, Hybrid Secure Transmission Protocol (HSTP).
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	HYBRID DEEP ATTENTION-BASED EMOTION RECOGNITION USING TEMPORAL-SPATIAL OPTIMIZATION FOR MULTI-SUBJECT VIDEO ANALYSIS
Author:	DR. RAMARAJU NAGARJUNA KUMAR, DR. G. PRASANNA LAKSHMI, DR. RABINS PORWAL, DR. C. MADHUSUDHANA RAO, SATYANARAYANA MURTHY VALLABHAJOSYULA, M. LAKSHMANA KUMAR, DR. M. NAGABHUSHANA RAO
Abstract:	Group-level emotion recognition (GER) is essential in applications involving human-computer interaction, public surveillance, and affective computing. This paper proposes a novel hybrid framework that integrates Enhanced Particle Swarm Optimization (EPSO) for feature selection with Recurrent Neural Networks (RNN) for modeling temporal emotion dynamics in video sequences. The system begins with preprocessing and frame extraction, followed by deep and statistical feature extraction. EPSO is employed to select the most informative features, which are then input into an RNN for sequential emotion prediction. Evaluations conducted on the AFEW dataset demonstrate that the proposed EPSO-RNN model outperforms traditional classifiers such as CNN, VGG-16, and SVM in terms of accuracy, precision, recall, and F1 score. The EPSO-RNN model demonstrated smooth convergence with minimal overfitting, aided by early stopping. The training accuracy peaked at 92.3%, with a validation accuracy of 89.4%, outperforming CNN (78.2%), VGG-16 (82.5%), and SVM (69.7%). Corresponding loss curves showed a steady decline, reinforcing the model’s stability. The results affirm the robustness and scalability of the proposed approach in complex, real-world group emotion recognition scenarios.
Keywords:	Affective Computing, Emotion Recognition, Attention Mechanism, Temporal-Spatial Optimization, Deep Learning, Multi-Subject Video Analysis.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	HYBRID DRL-CTO: OPTIMIZATION DRIVEN DEEP LEARNING FRAMEWORK FOR CEREBELLAR ATAXIA BASED ON HUMAN GAIT ANALYSIS
Author:	EDARA SREENIVASA REDDY, SUNIL PREM KUMAR PRATHIPATI
Abstract:	This study presents a Hybrid Deep Repeated Learning with Competitive Tuning Optimization (Hybrid DRL-CTO) framework for detecting and classifying Cerebellar Ataxia (CA) using human gait analysis. The proposed approach integrates a repeated learning mechanism with the CTO algorithm to enhance parameter optimization, improve convergence, and reduce computational complexity. Deep S3P features extracted from regions of interest in gait frames enable the model to capture discriminative representations while minimizing feature dimensionality, thereby improving classification accuracy. Experimental evaluations demonstrate that Hybrid DRL-CTO achieves 96.8% accuracy, 95.1% precision, and 97.1% recall, consistently outperforming existing techniques in k-fold validation. The combination of CNN-LSTM with the CTO algorithm further prevents overfitting and ensures reliable performance, offering a robust solution for CA classification with high efficiency and reduced computational effort.
Keywords:	Cerebellar Ataxia, Repeated Learning, Deep Learning, Gait Analysis, Competitive Tuning Optimization.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	BLOCK CHAIN BASED SECURED EHR WITH UNIFIED SIGNATURE ENCRYPTION SCHEME
Author:	VENKATESWARAN S , VIJAYARAJ N
Abstract:	Electronic Health Record(EHR) manage medical information in digital form and allow the patient to view their information and distribute the same to the doctors to diagnoses. Identity-based cryptosystem, specifically Unified Signature-Encryption Scheme(US-ES) is designed to enable the efficient and secure exchange of healthcare information within a data-sharing network. Leveraging bilinear pairings, the proposed system integrates a distributed, tamper-proof database that replicates health records across peer-to-peer (P2P) network. EHRs are treated as individual events, timestamped, and assigned a cryptographic hash. To ensure transparency and data reliability, these entries are organized into transaction blocks and each node in the P2P network maintains a copy of the ledger. Additionally, the healthcare blockchain includes user permission lists, which dictate access control and serve as essential network instructions. This framework ensures secured exchange of EHR both within and between medical institutions, eliminating the need for a third-party provider. To guarantee the integrity of the EHRs, US-ES cryptographic technique is applied on the blockchain, ensuring the security of decentralized healthcare data exchanges. The US-ES mechanism combines the functionalities of both digital signatures and encryption, providing confidentiality, authenticity, and efficiency in data transmission. Ultimately, the proposed US-ES technique is combined with digital signature to present a robust and secure method for healthcare data exchange, significantly reducing vulnerabilities associated with traditional systems. By removing the necessity for a trusted third party, the proposed solution enhances both the security and integrity of EHR transactions, fostering safer data sharing in the healthcare.
Keywords:	Blockchain, Electronic Health Record(EHR), Unified Signature-Encryption (USE), Identity-Based Cryptosystem (IBC).
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	DEPTH-FIRST SEARCH (DFS) BASED CHARACTER RECOGNITION USING GRAPH MATCHING ALGORITHMS
Author:	M.SARAVANAKUMAR ,Dr.S.KANNAN
Abstract:	This study focuses on full Multi lingual Printed character English Tamil, Malayalam, Hindi character recognition in UTF-8 encoding, utilizing both Times New Roman and Courier New font variations and hand written variation of styles Recognition. The proposed method employs Shi-Tomasi corner detection combined with a Depth-First Search (DFS) approach for adjacency matrix transformation. The adjacency matrix, represented as binary values (0 or 1), undergoes conversion and swapping to facilitate comparison across all scale variants, determining similarity using hierarchical matching techniques. The recognition process follows a structured approach: characters are matched based on the highest degree first, followed by the next highest degree, and subsequently, partial exact character recognition is conducted. The recognized character matrix is then utilized to print Unicode values and generate recognized character text. The recognized text is systematically converted into a Word file and saved, while all recognition input characters are preserved in a designated folder. Additionally, recognized text is stored in a structured format for further analysis. The system exclusively visualizes the program's input plot, corner detection plot, and shell recognition text. Furthermore, character recognition is enhanced using a comparative approach based on Graph Edit Distance, ensuring accurate shape differentiation between the two font styles. Character recognition has been extensively studied using feature-based and machine learning approaches, yet graph-theoretic methods remain underexplored despite their potential for providing structural interpretability. A particular knowledge gap exists in the literature regarding the application of Depth-First Search (DFS) traversal on adjacency matrices derived from character images. While prior studies focus on generic graph isomorphism or statistical descriptors, they do not systematically evaluate the role of DFS traversal in capturing distinctive structural patterns of characters for recognition purposes. This study addresses that gap by proposing a DFS based character recognition framework combined with graph matching algorithms. Characters are represented as graphs constructed from corner-detected nodes and their adjacency relationships, and recognition is achieved by comparing DFS traversal sequences with predefined templates. The framework is tested on English (uppercase and lowercase), Tamil, Hindi, and Malayalam characters, including both printed fonts (Times New Roman, Courier New) and handwritten style variations. The results demonstrate that DFS traversal effectively captures structural uniqueness, achieving an average accuracy of 98% for both printed and handwritten datasets. The approach also provides insights into computational efficiency and misclassification patterns, particularly in noisy or structurally similar characters. The novelty of this work lies in its systematic integration of DFS traversal with graph matching for character recognition, offering a transparent and interpretable alternative to black-box machine learning models. The new knowledge created by this research is the demonstration that DFS based graph traversal is not only computationally feasible but also robust across multiple languages and character styles, thus contributing a structural recognition paradigm that complements existing CR techniques.
Keywords:	UTF-8 Character Recognition, Shi-Tomasi Corner Detection, Adjacency Matrix, Depth-First Search (DFS) Using Graph-Based Character Recognition, Times New Roman & Courier New Font, Graph Edit Distance, Graph Based Matching (GBM).
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	IMPROVED BINARY DIFFERENTIAL EVOLUTION BASED CLASSIFIER SELECTION IN STACKED ENSEMBLE FRAMEWORK FOR EFFECTIVE DECEPTION SYSTEM
Author:	VIJAYASREE BODDU , PRAKASH KODALI
Abstract:	Deception is a state that refers to acting in a way that causes another person to believe something that is not true. Deception is a national security concern when investigating crimes. Accurate deception system is a critical challenge in criminal analysis necessitating the development of efficient predictive models. Despite advancements in ensemble-based machine learning models, selecting diverse classifiers to enhance model performance remains a significant hurdle. Past efforts to boost classification accuracy using ensemble learning encountered constraints. To enhance deception performance in this study a two level stacking framework is proposed. Selecting best classifier combination for this stacking framework is a hard problem. This problem is formulated as combinatorial optimization problem and attempted to solve using Binary Differential Evolution (BDE) algorithm. For effective solution and better results, in the proposed approach, base learners are encoded using binary encoding, while meta learners are encoded using one-hot encoding. Further, the proposed BDE uses dynamic mutation scaling factor and cross over rate. Finally, the continuous solution space is converted using sigmoid transfer function followed by thresholding. In this study nine diverse base learners and four meta learners are used to construct the stacking ensemble model. The proposed framework optimizes the combination of these classifiers using BDE, aiming to improve predictive performance. Our optimization process relies on a fitness function derived from the ensemble accuracy score and number of classifiers. A Concealed Information Test (CIT) is performed to collect the deception dataset. This dataset is used to evaluate the proposed model. The proposed model outperformed in terms of accuracy, sensitivity, specificity and F1-score when compared with State-Of-The-Art (SOTA) models in the literature. Next, when the proposed model was compared with state-of-the-art (SOTA) ensemble models, it not only achieved the best performance in terms of accuracy and sensitivity, but also reduced the ensemble model complexity drastically. Our findings demonstrate that improvements in performance across many metrics, highlighting the effectiveness of the BDE-based ensemble approach in designing a more accurate deception system.
Keywords:	Binary Differential Evolution, Deception Detection, Ensemble Learning, Optimization, Stacking
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	ADAPTIVE CARIBOU DEFENSE PROTOCOL (ACDP): A BIO-INSPIRED INTRUSION DETECTION FRAMEWORK FOR SECURING IOMT NETWORKS
Author:	BASIL BABY K, Dr. A. NITHYA RANI
Abstract:	Network security has become a critical concern due to the increasing complexity of cyber threats targeting interconnected systems. The “Internet of Medical Things” (IoMT) has transformed healthcare by enabling real-time monitoring, remote diagnostics, and automated medical interventions. Integrating IoMT devices into healthcare infrastructures exposes networks to security vulnerabilities, requiring robust intrusion detection mechanisms. “Host-based intrusion Detection Systems” (HIDS) provide a localized security approach, monitoring system logs, processes, and behaviors to detect unauthorized activities. Traditional detection techniques often struggle with evolving threats and resource limitations in IoMT environments. Bio-inspired optimization techniques offer adaptive security enhancements, refining detection mechanisms while minimizing computational overhead. The Adaptive Caribou Defense Protocol (ACDP) leverages nature-inspired intelligence to optimize intrusion detection, ensuring enhanced security resilience. By integrating bio-inspired approaches with HIDS, intrusion detection frameworks can achieve improved adaptability, real-time threat identification, and efficient security enforcement across IoMT networks, mitigating emerging cyber risks effectively.
Keywords:	Host Intrusion Detection Systems - Internet of Medical Things - Intrusion Detection – Cybersecurity in Healthcare - Caribou Optimization
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	MACHINE LEARNING IN IMAGE PROCESSING DRIVING INNOVATION IN VISUAL DATA ANALYSIS
Author:	DEVAKI KUTHADI, ARCHANA NAGELLI, PARUCHURI JAYASRI, SUSHUMA NARKEDAMILLI, KARRI L GANAPATHI REDDY, ANDE SASI HIMABINDU, C. GNANAVEL
Abstract:	This work introduces a new model based on Hybrid CNN and GAN for image processing and use in medical and natural image analysis. We aim to see if using CNNs for extraction and GANs for enhancement makes both image improvement and the classification process more accurate. In this work, we used medical CT scans, MRIs, and natural images and measured performance by calculating accuracy, precision, recall, F1-score, PSNR, and SSIM. Compared with existing architectures such as VGGNet and ResNet, the proposed Hybrid CNN–GAN achieved superior medical image classification accuracy of 95.2% and delivered enhanced image quality with PSNR of 37.2 dB and SSIM of 0.97. The model also demonstrated strong performance on natural image datasets, confirming its ability to generalize across domains. These results indicate that integrating CNN-based feature extraction with GAN-based enhancement improves both classification accuracy and perceptual image quality in medical and natural image processing tasks. This work provides more precise methods of collecting and studying images, which can significantly improve many fields that depend on analyzing precise visual data, such as the healthcare and autonomous vehicles sectors.
Keywords:	Hybrid CNN, Generative Adversarial Network, Image Processing, Medical Imaging, Image Enhancement, Deep Learning
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	SEFEX: A NOVEL APPROACH TO ACCURATE FACIAL EMOTION DETECTION USING EFFICIENTNET AND ATTENTION MECHANISMS
Author:	S SAHAYA SUGIRTHA CINDRELLA , R JAYASHREE
Abstract:	Facial emotion detection has become a pivotal technology in applications ranging from human–computer interaction to significant challenges in accurately recognizing emotions, especially when expressions are subtle, overlapping, or distorted by variations in lighting, pose, or noise. These limitations reduce reliability in real-world applications and create the need for more robust solutions. To address these challenges, this paper introduces the SEFEX (Single Emotion Facial Expression Detection) model, which combines advanced preprocessing techniques, EfficientNet-based feature extraction, and an attention mechanism to improve recognition accuracy. The model is evaluated on a comprehensive dataset encompassing eight emotion categories: happiness, sadness, anger, surprise, fear, disgust and neutral mental health monitoring for security and customer experience management. However, existing approaches face contempt. Preprocessing steps such as bilateral filtering for noise reduction, facial alignment using Haar Cascades, and landmark detection ensure consistent and high-quality inputs for training. EfficientNet is then employed to extract robust features, while the attention mechanism highlights key facial regions, enabling the model to capture subtle expression differences. SEFEX achieves an accuracy of 95.34%, outperforming state-of-the-art models including VGG16, ResNet50, and DenseNet. These results demonstrate SEFEX’s capability to enhance emotion recognition reliability, making it suitable for real-time applications in healthcare, customer service, and human-computer interaction.
Keywords:	Facial Emotion Detection, Deep Learning, Single Emotion Facial Expression Detection (SEFEX), EfficientNet, Feature Extraction, Attention Mechanism, Human-Computer Interaction, Noise Reduction, Image Preprocessing, Emotion Recognition.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	IMPROVED GREY WOLF OPTIMIZATION BASED BAND SELECTION FOR HYPERSPECTRAL REMOTE SENSING IMAGE CLASSIFICATION
Author:	MAANVITHA M D, PHANEENDRA KUMAR B L N
Abstract:	The selection of bands in hyperspectral imagery represents a significant challenge within the remote sensing discipline, attributable to the dataset's high dimensionality and the presence of redundant information across the various bands. This manuscript introduces a novel framework aimed at the classification of hyperspectral images. To identify the optimal set of bands, an enhanced grey wolf optimization technique employing a non-linear function is implemented, which iteratively refines the band configurations based on the optimal, suboptimal, and tertiary solutions, mirroring the roles of the "alpha," "beta," and "delta" wolves, respectively. Subsequently, the intrinsic information from the band set is extracted through a spatial filter that isolates the albedo component. A non-linear Support Vector Machine is trained with the extracted features. Empirical results derived from the Indian Pines and University of Pavia datasets confirm the effectiveness of the proposed framework in attaining superior classification accuracy over the state-of-the-art methods by accurately identifying appropriate bands and categorizing all varieties. The empirical quantitative and qualitative results validate the efficacy of the proposed framework. The proposed methodology holds significant relevance for land use analysis, environmental monitoring, and remote sensing applications, as it markedly reduces the number of bands utilized while maintaining classification performance.
Keywords:	Hyperspectral Images, Support Vector Machine, Grey Wolf Optimization, Band selection
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	TRANSFORMING THE MONETARY GIFT EXPERIENCE: QUANTITATIVE EVALUATION OF A MOBILE APP WITH PLAYFUL INTERACTIONS USING THE IVPRD INDEX
Author:	ALAN M. INFANTE VIDALON, ROSARIO D. OSORIO CONTRERAS, WILSON A. LAZO TAPIA
Abstract:	The development of this study confirms the perceived value of digital gifts through the (IVVPRD), which is a multidimensional construct for the evaluation of mobile applications, which digitizes cultural practices for money. The objective of the study is to prevent the gap between economic digitalization and traditional cultural practices. Using a quasi-experimental design with 100 participants from Peru, a mobile application with fun elements was evaluated relatively against traditional methods (physical envelopes). IVPRD consists of four dimensions: monetary value (VM), interaction efforts (EI), social/cultural context (SC), and experience experience (EC), which is expressed with the equation IVPRD = 0.31 (VM) 0.28 (EI) 0.32 (SC) - 0.12 (EC). The results showed the durability of the model (r² = 0.847), revealing that the digital application has reached significantly higher values (IVPRD = 7.49 ± 0.75) against the traditional method (IVPRD = 5.10 ± 0.68). Interaction efforts (EI) was found to be the most important difference between the methods (5.23 points), while the social/cultural context (SC) appeared as a factor with the highest predictive weight (β = 0.32). Significant differences were observed according to demographic and contextual variables: preference for new users (18-24 years: 4.10 points) and the use of socioeconomic status (18-24 and 3.70 respectively), as well as the inverse correlation between formality and apartment choice. The experience analysis identified "game setup" as the most important point for improving the success rate (89.5%) and moderate disappointment (2.68/10). This study shows that the inclusion of recreational elements and a social and cultural context significantly transform perceptions of value in applications, providing a methodological framework that relates to culturally sensitive FinTech solutions
Keywords:	User Experience, Gamification, Fintech, Digital Monetary Gift, Mobile App.
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text

Title:	TRANSFER LEARNING BASED ROBERTA WITH SMOTE: A NOVEL APPROACH FOR TWITTER SENTIMENT ANALYSIS FOR HATE SPEECH
Author:	REKHA JANGRA , ABHISHEK KAJAL
Abstract:	In the current digital era, a big spike of social interactions among people on various social networking sites have been witnessed around the globe. X platform (erstwhile Twitter) has turned leading platform for users to express their opinion on different burning topics related to social, political, religious domains. Though such discussions are healthy for a politically and socially active dynamic generation, but unfortunately many tweets carry hate speech. In past few years, India too witnessed the exponential rise of posts containing hate speech. Hence, we introduced a deep learning technique for catching the hate speech generated posts using X platform as our source for the dataset to classify the posts sentiments as negative and positive. Current research has assessed traditional ML methods, including SVM, Naïve Bayes, and Random Forest, alongside DL techniques. In light of the limitations of current methods, the suggested work offers a solution by merging RoBERTa with transfer learning. During preprocessing, the crawled Twitter data has been filtered, case-folded, and stemmed. Subsequently, stop word removal and tokenization have been executed. Data labeling has been conducted via deep learning algorithms. A word cloud has been generated, and a frequency chart has been produced based on positive and negative sentiments. Ultimately, the accuracy and error rates of LSTM, BERT, optimized RoBERTa, and Hybrid transfer learning-based RoBERTa with smote have been simulated. Simulation results indicate that LSTM attains 94.99%, BERT 92.87%, Optimized RoBERTa 97.45%, Hybrid RoBERTa with transfer learning 98.72% and Proposed SMOTE-based Hybrid RoBERTa with Transfer Learning 99.12%. The proposed model demonstrates superior accuracy, precision, recall, and F1-score relative to traditional methods. Consequently, the suggested approach has addressed the accuracy issues present in traditional hate speech Sentiment analysis systems.
Keywords:	Sentiment Analysis, Hate Speech, LSTM, BERT, ROBERTA, Transfer learning, SMOTE
Source:	Journal of Theoretical and Applied Information Technology 15^th October 2025 -- Vol. 103. No. 19-- 2025
Full Text