|
|
|
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
September 2026 | Vol. 104
No.18 |
|
Title: |
DIGITALISATION OF ENVIRONMENTAL MANAGEMENT IN LOCAL SELF-GOVERNMENT BODIES |
|
Author: |
HENNADII MAZUR, YURII YEFIMOV, VOLODYMYR SERVETNYK, OLHA VASYLENKO, MYKOLA
ANDRUSHKO |
|
Abstract: |
In the context of growing challenges related to environmental pollution, climate
change and increased anthropogenic pressure on natural systems, the
digitalisation of environmental management plays a key role. The study aims to
identify key problems that limit the effectiveness of local self-government
bodies and to determine ways to improve their institutional capacity to ensure
an adequate level of environmental safety at the local level. The study applies
a systematic approach and uses methods of comparative analysis, statistical
evaluation of results, content analysis of regulatory and legal acts, and
questionnaires. Data from 10 territorial communities in different regions of
Ukraine, selected based on criteria such as digital maturity, population size,
and environmental impact, were analysed. The results of the study showed that
the key barriers are insufficient funding for environmental programmes and
limited human resources. Large cities scored highly at 4.5 out of 5 in
institutional readiness; medium-sized cities scored an average of 3.6; and small
communities scored lower (3.1), compensating for limited resources with high
public engagement. Institutional capacity (coefficient 0.12; p = 0.009) and the
effectiveness of law enforcement agencies (0.085; p = 0.01) had a positive
impact on the level of environmental security. A positive correlation was found
between GDP per capita (0.05; p = 0.013) and foreign direct investment (0.04; p
= 0.027). In contrast, the share of industry in the economic structure has a
negative impact (-0.03; p = 0.046), caused by increased environmental pressure.
The index of fiscal autonomy of communities was 0.06 (p = 0.007). A conceptual
model of a digital ecosystem for environmental management of communities has
been developed. The practical significance of the results lies in the potential
to use the developed recommendations to improve local environmental safety
strategies, optimise resource allocation, and establish an effective
environmental monitoring system. Prospects for further research include
analysing the possibilities for digitalising management processes. |
|
Keywords: |
Natural Resources, Agricultural Land, Digitalisation, Environmental Management,
Multilevel Governance, Environmental Policy, Sustainable Development. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
BEYOND FUNCTIONALITY: WHY BRAND ASSOCIATION OUTWEIGHS PERCEIVED QUALITY IN
FOSTERING LOYALTY TO AI CHATBOTS |
|
Author: |
LUKMAN KHAKIM, AFDAL MAKKURAGA PUTRA, HERI BUDIANTO |
|
Abstract: |
This study investigates the influence of Brand Association and Perceived Quality
on the loyalty of ChatGPT users among Master of Communication Studies students
at Mercu Buana University. With ChatGPT holding a 59.5% market share in the AI
chatbot sector, understanding the drivers of user loyalty is crucial. Despite
its widespread use, limited studies have simultaneously explored how brand
perception and quality affect loyalty, especially among student users. Using a
quantitative survey method, data were collected from 63 students via a
Likert-scale questionnaire and analysed using statistical tools such as
normality tests, correlation analysis, multiple linear regression, t-tests, and
F-tests. The findings reveal that Brand Association has a significant positive
effect on user loyalty, indicating that students who strongly identify with
ChatGPT’s brand are more likely to remain loyal users. In contrast, Perceived
Quality alone does not significantly influence loyalty, suggesting that
favourable service assessments are insufficient to guarantee continued use.
However, when combined, Brand Association and Perceived Quality jointly
influence user loyalty, highlighting the synergistic value of brand image and
service quality. Among the two variables, Brand Association emerged as the more
dominant factor. The study underscores the importance of branding strategies in
strengthening user loyalty to AI-based platforms. |
|
Keywords: |
Brand Association; Perceived Quality; Brand Loyalty, ChatGPT |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
EDGE IOT INTEGRATED WEARABLE SYSTEM FOR REAL-TIME CARDIOVASCULAR MONITORING AND
EXPLAINABLE AI |
|
Author: |
DR N SREEKANTH, KAVITHA G L, DR. MOHAMMAD SIRAJUDDIN, SADINENI NEELIMA, KOYYA
AVINASH, DR B. SIVA PRASAD |
|
Abstract: |
Cardiovascular diseases (CVDs) remain the foremost cause of global mortality,
necessitating continuous and intelligent monitoring systems capable of detecting
pathological events before they become life-threatening. This paper proposes a
cardiovascular monitoring wearable system based on Edge Internet-of-Things
(Edge-IoT) technology that incorporates Explainable Artificial Intelligence
(XAI) for real-time monitoring. The new architecture features flexible skin
conformal electrocardiogram (ECG) and photoplethysmography (PPG) sensors,
together with models that are deployed at the edge to classify arrhythmia in
real-time on the sensor device, as well as estimate blood pressure and extract
heart rate variability (HRV) analytics with sub-second latency. A multi-site
clinical ECG database is used to train a hybrid Convolutional Neural
Network–Bidirectional Long Short-Term Memory (CNN-BiLSTM) model to detect
ten-class arrhythmias, which is then evaluated in a leave-one-patient-out
protocol on 1,240 subjects with a macro-average AUC of 0.963. Additionally, a
combination of SHAP value analysis and Grad-CAM temporal heat maps is used to
produce clinically interpretable model predictions, which allow the cardiologist
to review the rationale for each prediction. A lightweight TinyML variant,
compressed using knowledge distillation, achieves a 78% reduction in the model
size with 96.4% of the teacher model performance compared to the full model,
while being deployed on an ARM Cortex-M7 microcontroller at 3.1 mW.
Communication between the system and the cloud dashboard is made possible
through a Bluetooth Low Energy (BLE) and 5G-NB-IoT dual radio link, and the
cloud dashboard, which combines patient data from the longitudinal patient
record with clinical decision support. The edge-XAI pipeline is integrated to
reduce the number of false positive alerts by 79.3% over threshold-based rule
systems when evaluated in real-world ambulatory applications, and to achieve
clinician trust scores that are much higher than those of current automated
clinician monitoring tools. These results pave the way towards a research
prototype to a clinically deployable, privacy-preserving and interpretable
wearable cardiovascular intelligence. |
|
Keywords: |
Wearable Sensors, Edge Computing, Internet of Things, Cardiovascular Monitoring,
Explainable Artificial Intelligence, ECG, Arrhythmia Detection, TinyML,
CNN-BiLSTM, SHAP, Grad-CAM |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
INTELLIGENT SMART SYSTEMS AND AUGMENTED REALITY IN INCREASING THE EFFICIENCY OF
THE EDUCATIONAL PROCESS |
|
Author: |
LIPING GONG, ALMAZBEK ARZYBAEV, YULIIA SUPRUNCHUK , ALINA DANILEVICA, ANDRII
BILIAVETS |
|
Abstract: |
This study addressed the rapid expansion of smart digital infrastructures and
the need for empirically validated, data-driven learning models. The study
evaluated how Intelligent Smart Systems and augmented reality affected learning
efficiency, retention, engagement, cognitive load, and HCI/UX performance. A
five-arm controlled quasi-experimental design was implemented over 6–8 weeks.
Distributional screening of 600 students identified 200 eligible participants,
followed by controlled random allocation into five groups of 40. The conditions
comprised traditional instruction, an AI-enabled Smart System, visual AR, audio
AR, and hybrid visual–audio AR. Baseline domain-specific performance was 72.4 ±
6.1, with normality (p > .12), variance homogeneity (p = .41), and inter-group
deviation below 3.5%. Outcomes included normalized learning gain, delayed
retention, immersion, cognitive load, adaptivity, interaction error rate,
usability, and interaction density. ANCOVA, repeated-measures modelling, and
mediation analysis were applied. Smart/AR conditions produced normalized
learning gains of 0.68–0.92. Post-test means increased from 75.4 ± 6.8 in the
control group to 92.4 ± 5.1 in the hybrid group. Delayed scores ranged from 70.2
± 7.1 to 90.1 ± 5.4. ANCOVA confirmed a large group effect, F(4,194) = 18.72, p
< .001, partial η² = .278. The time × condition effect was also large, F(4,194)
= 21.36, p < .001, partial η² = .306. Mediation was significant, ab = 0.214, SE
= 0.052, 95% CI [0.118, 0.324]. The resulting framework combined screening,
multimodal Smart/AR intervention, cognitive–behavioural–HCI assessment, and
operational deployment thresholds for evidence-based educational implementation. |
|
Keywords: |
Smart Learning, Augmented Reality (AR)-Based Education, Immersive Learning
Technologies, Cognitive Efficiency, Analytical Development of the Learning
Process |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
ADAPTIVE ROBOTIC ARM FRAMEWORK FOR SURGICAL TOOL PICK-AND-PLACE USING REAL-TIME
PERCEPTION AND TRAJECTORY OPTIMIZATION |
|
Author: |
M. SENTHILKUMAR, K. SHANMUGASUNDARAM, A.T. RAJAMANICKAM, D. ANTONY PRADEESH |
|
Abstract: |
The rise in popularity and use of robotic manipulators which act autonomously
can be noted in the process of surgery. This represents the significant step
towards achieving error-free and precise healthcare. In the paper, there is
presented the design and implementation of the system based on the robotic arm
performing pick-and-place tasks on surgical tools in a reliable and adaptive
manner. The system which has been proposed involves real-time computer vision,
planning of trajectory, and force feedback-controlled grasp to provide precise
detection, robust grasping, and contextual placement of the tools in the
operation theatre. The multi-stage pipeline has been implemented in the ROS
Noetic environment with the application of MoveIt for trajectory planning,
Gazebo and RViz for physical simulations and visualization, and deep
learning-based modules for segmentation and prediction of the phase of work.
Mathematical models for kinematic control, trajectory optimization, and grasp
stability have been formulated and experimentally verified using realistic
scenarios. It has been demonstrated that the system which has been proposed
allows achieving a tool-handling success rate greater than 95% for the five most
used surgical instruments and at the same time provides the time variability
which is constrained to less than plus or minus two seconds. Comparative
analysis has shown that the optimized trajectory decreases the rate of errors by
70% and smoothness index increases more than 20% comparing to the base-line
trajectory. Thus, the combination of robust perception and adaptive control can
be considered as evidence of the possibility to use the architecture as the main
module for the surgical assistance in semi-autonomous systems. |
|
Keywords: |
Robotic Arm, Multi-Stage Pipeline, ROS, MoveIt, Gazebo and RViz. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
CLASSIFICATION OF ANCIENT TAMIL PALM LEAF MANUSCRIPTS USING AN OPTIMIZED CNN
BASED ON AN ENHANCED WHALE OPTIMIZATION ALGORITHM WITH PARTIAL OPPOSITION-BASED
LEARNING |
|
Author: |
S. KEERTHIKA, P. TAMIL SELVAN |
|
Abstract: |
Tamil palm-leaf manuscripts are an inestimable source of historical and cultural
information. Still, their conservation and electronic reading are difficult due
to severe deterioration, noise, uneven lighting, and disordered structures in
handwritten characters. To achieve proper classification of characters extracted
from such documents, preprocessing, segmentation, and well-trained deep learning
models are necessary. The proposed study suggests a smart character recognition
system that combines the Horizontal Projection Profile (HPP)-based line
segmentation with a Convolutional Neural Network (CNN) optimized by a Whale
Optimization Algorithm (WOA) along with Partial Opposition-Based Learning
(POBL). The hybrid WOA-POBL (PWOA) method better balances exploration and
exploitation in hyperparameter tuning, increases the diversity of a population,
and speeds up convergence, which overcomes the shortcomings of manual or
grid-based CNN tuning. An end-to-end preprocessing pipeline, including denoising
and removing the background, binarization, and line segmentation, is used to
extract high-quality character samples from damaged manuscript images. The
experimental findings of various learning rates and batch-size setups illustrate
that the suggested PWOA-CNN is much more effective than the baseline CNN,
GA-CNN, PSO-CNN, and WOA-CNN. The convergence analysis also validates the
efficiency and stability of the offered method. Altogether, the framework is a
strong and scalable system of digital preservation and automatic identification
of ancient Tamil scripts and could be applied to other historical collections of
handwritten documents. |
|
Keywords: |
Convolutional Neural Network (CNN), Tamil palm-leaf manuscripts, Hyperparameter
Optimization, Partial Opposition-Based Learning (POBL), Horizontal Projection
Profile (HPP) |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
ENHANCED FUNCTIONAL AND NON-FUNCTIONAL REQUIREMENT CLASSIFICATION USING FEATURE
EXTRACTION, FEATURE SELECTION AND MACHINE LEARNING CLASSIFIERS |
|
Author: |
CHANDANI KUMARI , MAHENDRA KUMAR GOURISARIA , JUNALI JASMINE JENA |
|
Abstract: |
Classification of software requirement is very important in the software
development field. Categorization of non-functional and functional requirements
at early stage help to improve the design phase, quality of the software and
reduce the probability of risks. Manual classification of software requirement
is cumbersome due to huge volume of data and customer requirements that lead to
error & delay in the development of the software. To resolve this issue, feature
selection techniques such as Chi2 and PCA combined with feature extraction
techniques such as TF-IDF, BOW, Fast-Text, GLOVE and WORD-2-VEC have been
simulated with machine learning models such as NB, SVM, LR, D-Tree, Stacking,
K-NN, Voting, GB, RF and AdaBoost to automatically classify the software
requirements. In multi-class classification, Fast-Text and Chi2 simulated with
Stacking and Voting classifiers achieves an accuracy of approximately 90%. In
binary classification, TF-IDF simulated with SVM, Stacking and Voting
classifiers reached an accuracy of 98%. Hence, supremacy of ensembles models
like Stacking and Voting classifiers has been observed in automated software
requirements classification task. |
|
Keywords: |
Functional Requirements, Non-Functional Requirements, Feature Extraction,
Feature Selection, Software Requirements Classification |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
DETERMINANTS INFLUENCING CUSTOMERS’ INTENTION TO USE AI-DRIVEN ROBOTIC SERVICE
IN MALAYSIAN RESTAURANTS |
|
Author: |
SHIAU WEI CHAN , HIROHIKO MORI , WAN THENG LOH , MD FAUZI AHMAD |
|
Abstract: |
As artificial intelligence (AI) and service robotics increasingly permeate the
restaurant industry, new IT-enabled service models are emerging. In IT research,
customer acceptance is critical; a robotic system's value extends beyond mere
functionality, requiring users to perceive the technology as useful, easy to
use, trustworthy, enjoyable, and low-risk. But there are practical issues as
well. For example, customers at two restaurants reported that robots struggled
to accommodate special requests for additional items like sauces or tableware
and frequently caused order confusion by delivering food to the wrong customers.
This study aimed to explore the expected level of customers' intention to use
AI-driven robotic services and the determinants that influence that intention in
restaurant contexts in Batu Pahat, Johor, Malaysia. A sample of 307 restaurant
customers who had experienced robotic services at two restaurants was surveyed
using questionnaires, and the data were analyzed quantitatively by employing
SPSS software. The study found a high level of intention among customers to use
AI-driven robotic services. Results indicated positive correlations of intention
to use with perceived usefulness (0.743), perceived ease of use (0.624),
perceived need for interaction with the technology (0.516), perceived enjoyment
(0.645), and perceived trust (0.609); in addition, a weak negative correlation
was found between intention to use and perceived risk (0.197). For IT
contribution, the contribution of this study was based on the contextual
extension and empirical testing of a TAM-based acceptance framework integrating
functional, interactional, hedonic, trust, and risk factors within an
understudied restaurant setting in Malaysia. |
|
Keywords: |
Determinants, Intention to use, AI-driven Robotic Service, Restaurants |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
INTEGRATION OF ARTIFICIAL INTELLIGENCE TECHNOLOGIES IN THE FORMATION OF
STUDENTS’ COGNITIVE INDEPENDENCE |
|
Author: |
ALBINA VOLKOTRUBOVA, JIE SUN2, OLHA PROKOPENKO, OKSANA NIKOLAIENKO, KHRYSTYNA
KONDAUROVA |
|
Abstract: |
Generative artificial intelligence is increasingly embedded in higher education,
yet its contribution to students’ cognitive independence remains uncertain
because prior research has often examined self-regulation, critical thinking,
learning outcomes, and AI acceptance as separate effects. The problem addressed
in this study was the absence of an integrated, empirically validated framework
that simultaneously measures cognitive independence and structures AI-supported
learning without replacing students’ own cognitive effort. The study therefore
developed and experimentally validated an AI-driven adaptive model combining
learning analytics, generative support, adaptive scaffolding, reflective
revision, and teacher pedagogical control. A multi-site quasi-experimental
pre-test/post-test design involved 240 students from four higher education
institutions. The experimental groups used an integrated environment comprising
ChatGPT, Gemini, Claude, and Microsoft Copilot, whereas the control groups
followed comparable learning activities without systematic generative AI
support. Effectiveness was assessed using the Cognitive Independence Index (CII)
and the supplementary SRL-S, TPQ, and DSS indicators. In the experimental
groups, CII increased from 0.50 to 0.71, compared with 0.49 to 0.56 in the
control groups; SRL-S increased from 0.52 to 0.73 versus 0.51 to 0.58; TPQ from
0.55 to 0.75 versus 0.54 to 0.61; and DSS from 0.51 to 0.72 versus 0.50 to 0.57.
Analytical processing and self-correction showed the greatest sensitivity to the
intervention. The main contribution is a validated measurement-and-support
architecture linking cognitive outcomes to behavioral patterns of editing,
critical evaluation, and reflective revision. The findings indicate that
pedagogically structured human–AI interaction, rather than the frequency of AI
use alone, is associated with stronger cognitive independence. |
|
Keywords: |
Learning, Pedagogy, Knowledge, Artificial Intelligence, Generative Artificial
Intelligence, Cognitive Independence, Adaptive Learning Model, Learning
Analytics, Self-Regulated Learning. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
AN INTELLIGENT DEEP LEARNING APPROACH FOR GASTROINTESTINAL DISEASE DETECTION IN
ENDOSCOPIC IMAGES |
|
Author: |
K. SHARIFA , S. MALARVIZHI |
|
Abstract: |
The rising number of gastrointestinal (GI) diseases demands accurate and early
diagnosis methods that are essential for effective treatment. The research
developed a new deep learning system that uses endoscopic images to classify GI
diseases. The Kvasir dataset provides the model with diverse medical images that
contain clinically significant samples. The system uses Contrast Limited
Adaptive Histogram Equalization (CLAHE) to enhance image quality by making
crucial image details more visible. The Min-Max normalization method converts
pixel intensity values into a standard range that helps maintain stable
performance during model development. The researchers applied data augmentation
methods such as rotation and flipping to enhance model performance while
diminishing the risk of overfitting. EfficientNetB3 model with transfer learning
capabilities to extract advanced feature representations from GI images. The
Diffusion Graph Transformer processes these features by using graph-based
learning together with attention mechanisms to represent complex relationships
between feature elements. The proposed model combines methods for extracting
local features with techniques that enable a complete understanding of global
context. The Kvasir dataset testing results show that the model achieves a
classification accuracy of 98.41%, which demonstrates its effective performance.
The proposed method shows better performance than all existing state-of-the-art
models. The results demonstrate that the model maintains its strong performance
with reliable efficiency to classify GI diseases. The proposed framework
provides a scalable and accurate solution for automated GI disease diagnosis,
supporting improved clinical decision-making and early detection. |
|
Keywords: |
Gastrointestinal, Kvasir dataset, Diffusion Graph Transformer, EfficientNetB3,
Contrast Limited Adaptive Histogram Equalization. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
PATIENT-INVARIANT ECG ARRHYTHMIA CLASSIFICATION VIA PER-BEAT NORMALIZATION AND
STRICT INTER-PATIENT VALIDATION |
|
Author: |
DEEKSHIKA SIVA VINNY, Dr. JEEVAN JALA, Dr V SANGEETHA, JHANSI PANDIRI, Dr
KANDUKURI GEETHA, Dr. A MAHENDAR, SURESH DAMERUPPULA |
|
Abstract: |
The cardiovascular diseases have been the most important cause of death
worldwide, and the electrocardiogram (ECG) analysis is the diagnostic foundation
of arrhythmia. Although accuracies of 99.0 have been reported in research, there
is a deadly industrial/clinical disconnect: 86.7% of existing studies use
intra-patient evaluation paradigms that do leak patient-specific morphological
patterns, giving inflated measures that collapse disastrously on unseen
patients. This research paper helps fill this important safety gap because it
makes the hypothesis that clinical utility is more important to evaluation
methodology and feature normalization than to architectural complexity. Our
suggested lightweight three component dual lead convolutional neural network
will include per beat z-score normalization to eliminate patient specific
amplitude leakage, inverse square root class weighting with focal loss to
overcome extreme class imbalance, and physiological R-peak jitter augmentation
to be robust to annotation noise. When strict patient-disjoint validation is
applied to the MIT-BIH Arrhythmia Database (trained on 32 patients, tested on 5
completely unknown patients), we obtain a macro F1-score of 0.714. It is worth
noting that the model has an outstanding recall (99.7%) and F1 (91.5) on
life-threatening Ventricular ectopic beats. Ablation experiments ensure that
normalization per-beat is the primary force behind the phenomenon ( +0.210 macro
F1 gain), whereas t-SNE plots ensure the creation of patient-invariant feature
embeddings. These results criticize the modern trend of growing the complexity
of architecture and show that strict assessment and normalization are the major
factors of the reliability of cross-patient arrhythmia detection. |
|
Keywords: |
ECG Arrhythmia Classification, Inter-Patient Validation, Patient-Invariant
Features, Deep Learning, MIT-BIH Database, Class Imbalance, Focal Loss. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
INTELLIGENT EXTENDED HEURISTIC META-LEARNING FRAMEWORK FOR ADAPTIVE
CLASSIFICATION OF DIVERSE AND HETEROGENEOUS DATA ENVIRONMENTS THROUGH
CONTEXT-AWARE REPRESENTATION HARMONIZATION AND DYNAMIC MULTI-EXPERT LEARNING |
|
Author: |
Dr. B RAMA, B ANITHA |
|
Abstract: |
The large volumes of heterogenous data are being collected across different
domains. Therefore there is a huge demand for intelligent classification systems
which can be adapted to the changing nature of the data. This includes variable
structural and feature distributions, variable degrees of noise, as well as
variable modality characteristics. Most existing methods of classification,
including traditional machine learning, deep learning and AutoML frameworks have
traditionally used fixed pre-processing techniques, rigid model selection
processes and very little ability to adapt when distribution changes occur. As a
result, their generalization capabilities are significantly impacted in such
dynamic environments. In order to meet these needs, this paper presents an
extended heuristic meta-learning framework (i.e., IH-Meta-Framework) to provide
adaptive classification across various types of heterogeneous and dynamic
environments. This involves an integrated set of five interdependent analytical
components: CASE-Net: A Context-Aware Heterogeneity Signature Encoder to
generate dataset fingerprints; ARHT: An Adaptive Representation Harmonization
Transformer to transform features based on the context; HMPAS: A Heuristic
Meta-Policy Architecture Selector to select models and hyperparameters based on
heuristics; DMEHC: A Dynamic Multi-Expert Heuristic Classifier to reason using
ensembles at a sample level; TDMCL: A Test-Time Drift-Aware Meta-Correction
Layer to adapt during deployment under evolving data distributions. Overall, the
architecture presented provides a single continuous data-flow pipeline that
enables the simultaneous optimization of representation quality, classification
accuracy, search efficiency, robustness, calibration reliability, and drift
resilience. The integration of contextual meta-learning, heuristic optimization,
adaptive expert routing and drift aware corrections enable the framework to
achieve high degree of reliable classifications across complex heterogeneous
environments. The proposed research will advance future generations of
intelligent analytics through enhanced decision consistency, scalability,
interpretability and robustness with regards to real world deployments for Data
Intensive Applications. |
|
Keywords: |
Context-Aware Meta-Learning, Adaptive Classification, Heterogeneous Data
Analytics, Multi-Expert Learning, Heuristic Optimization, Scenarios |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
T-DASH: INTEGRATING TASK-BASED TRAINING AND AN INTERACTIVE DATA-DRIVEN LEARNING
ENVIRONMENT FOR DATA VISUALIZATION AND DASHBOARD COMPETENCY DEVELOPMENT |
|
Author: |
PARWAPUN KAMTAB , NOPPARAT SABAYKAN |
|
Abstract: |
This study developed and preliminarily evaluated T-DASH, a five-stage task-based
training model supported by an Interactive Data-Driven Learning Environment for
data visualization and dashboard competency development. The three-phase
research-and-development process comprised conceptual synthesis; development and
expert review of task-based training activities, assessment instruments, and the
interactive environment; and field implementation using a one-group
pretest-posttest design. The T-DASH architecture links a staged data-to-decision
workflow, five competency domains, and an evidence-to-feedback cycle supported
by eight environment functions: website access, interactive media, multimedia,
learning resources, analytics, assessment, communication, and online activities.
Five experts evaluated the model and supporting components. Twenty-five support
personnel at Chulabhorn Royal Academy completed a 30-hour program, and one
external assessor scored parallel pretest and posttest performance tasks and
completed dashboards using analytic rubrics. Expert ratings were high for the
model (M = 4.41, SD = 0.57), activity plans (M = 4.40, SD = 0.59), and
interactive environment (M = 4.28, SD = 0.60). Overall competency increased from
pretest (M = 2.79, SD = 0.39) to posttest (M = 4.13, SD = 0.29), t(24) = 12.96,
p < .001, Cohen's dz = 2.59. Completed-dashboard quality (M = 4.18, SD = 0.31)
and participant satisfaction (M = 4.33, SD = 0.35) were high. The study
contributes an operational training architecture that connects authentic
whole-task performance, explicit competency criteria, and actionable learning
evidence within one dashboard-development workflow. The findings support the
feasibility of T-DASH in the study context and document substantial within-group
change; however, the one-group design does not establish a causal effect. |
|
Keywords: |
Task-Based Training, Data Visualization, Dashboard Development, Data-Driven
Learning Environment, Professional Competency, Learning Analytics |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
HUMAN PERFORMANCE IN DETECTING DEEPFAKE MEDIA: AN EXPERIMENTAL EVALUATION IN
INDIA |
|
Author: |
ANINDITA MALIK , DEEPAK RAJ RAO G , KIRAN KUMARI |
|
Abstract: |
The rapid advancement of deepfake technologies has made it increasingly
difficult to distinguish between real digital content from Artificial
Intelligence (AI) generated media. This study examines human efficiency to
detect deepfakes through a controlled media-classification experiment. It was
hypothesized that the human ability to detect deepfakes would be low in general
and vary significantly across the demographic and behavioral characteristics
such as age, occupation, social media usage, and prior familiarity with deepfake
technology. Participants classified a series of images and videos as deepfake or
real and identified the perceptual cues influenced their decision. The responses
were analyzed using Signal Detection Theory (SDT), Dual-Process Theory (DPT),
and the Heuristic-Systematic Model (HSM), to interpret the decision accuracy and
cognitive processes underlying it. Results confirmed the hypothesis:
participants showed substantial difficulty distinguishing genuine content from
manipulated content; with misclassification rates exceeding 53% across
categories. Detection accuracy differed meaningfully by age, profession, use of
social media and familiarity with deepfake technology, with senior, more
experienced, and more digitally active participants performing better.
Participants relied predominantly on surface level cues like media quality and
facial expressions instead of systematic analysis. The findings reveal the
critical gap between the advancement of generative AI and society’s ability to
detect it, emphasizing the need for stronger digital media literacy initiatives
and hybrid human-AI detection frameworks to ensure information safety. |
|
Keywords: |
Deepfake, Artificial Intelligence, Synthetic Media, Generative Adversarial
Networks, Deepfake Detection Framework |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
CYBERSECURITY AS A COMPONENT OF NATIONAL AND ECONOMIC SECURITY |
|
Author: |
LEONID MAIDANEVYCH , VIKTOR GUDZ , VIKTORIIA REINSKA , SVITLANA HURKOVSKA ,
SERHII MOTORNIUK |
|
Abstract: |
The article highlights the conceptual principles and practical aspects of
ensuring cybersecurity as a key component of national and economic security in
the context of the digitalization of society and the growth of hybrid threats.
The study is aimed at generalizing modern approaches to the formation of a cyber
defense system, analyzing the impact of cyber threats on the functioning of
state institutions and economic processes, as well as determining the role of
information influence as a tool for destabilizing society. Particular attention
is paid to the importance of youth media literacy as an important factor in
increasing cyber resilience and forming a safe information environment. The
methodological basis of the study is the analysis of scientific publications,
regulatory legal acts in the field of cybersecurity, analytical materials of
international organizations, as well as the application of systemic and
comparative approaches to assessing the effectiveness of cyber protection
measures. Methods of generalization, systematization and structural-functional
analysis were used to study the relationship between technical, organizational
and social aspects of cybersecurity. It has been established that modern cyber
threats are complex in nature and combine technical attacks with information and
psychological influence, which significantly complicates the process of their
detection and neutralization. It has been proven that effective cybersecurity
requires an integrated approach, which includes the development of institutional
capacity, improvement of regulatory and legal regulation, implementation of
modern information protection technologies and activation of educational
initiatives. It has been shown that increasing the level of media literacy of
young people contributes to reducing vulnerability to disinformation, forming
critical thinking and strengthening the information security of society. It is
concluded that cybersecurity is a strategic prerequisite for ensuring the
stability of the state, the sustainability of the economy and effective
counteraction to modern information threats. |
|
Keywords: |
Cybersecurity, National Security, Economic Security, Information Influence,
Youth Media Literacy |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
QUANTUM DRIVEN NEUROMORPHIC EDGE COMPUTING FRAMEWORK FOR ULTRA LOW POWER
INTELLIGENT IOT DEVICES |
|
Author: |
B DEEVENA RAJU, SUBHASHINI PALLIKONDA , G LAKSHMI , T SRINIVASULU , P. KIRAN
KUMAR, ATHMAKURI SATISH KUMAR, V. VIJEYA KAVERI, N. DHARANI KUMAR |
|
Abstract: |
The Internet of Things (IoT) has seen rapid expansion, creating a great demand
for smart edge computing architectures capable of real-time data processing
under limited energy constraints. However, traditional edge intelligence
solutions are often inefficient because they consume high power, exhibit high
latency, and lack flexibility in resource-constrained environments. To address
these challenges, this paper proposes a novel Quantum-Driven Neuromorphic Edge
Computing Framework (QDNEC) for ultra-low-power intelligent IoT devices. The
proposed framework integrates Quantum Feature Encoding (QFE), Quantum-Guided
Spiking Neural Networks (QGSNNs), memristive neuromorphic processing, and an
Adaptive Energy-Aware Neuromorphic Scheduler (AEANS) into a unified
architecture. First, sensory data from various IoT devices are preprocessed and
converted into quantum state representations using variational quantum circuits.
The QGSNN architecture is then trained to learn complex spatiotemporal
relationships, while the adaptive scheduler dynamically optimizes computational
resources based on workload characteristics and energy levels. Experimental
results revealed that the proposed framework consumed only 78 mW of energy per
inference and achieved an average inference latency of 14 ms, with an accuracy
of 98.42%, precision of 98.16%, recall of 98.28%, and an F1-score of 98.22%.
Compared to existing edge intelligence models, it reduced energy consumption by
46.8% and latency by 38.5%. The experimental results validate that the proposed
QDNEC framework is an accurate, scalable, and energy-efficient approach for
developing next-generation intelligent IoT ecosystems, thereby enabling
sustainable ultra-low-power edge intelligence for real-time IoT applications. |
|
Keywords: |
Quantum Edge Computing; Neuromorphic Computing; Spiking Neural Networks;
Ultra-Low-Power IoT; Edge Intelligence; Adaptive Energy-Aware Scheduling. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
AN EXTENSIVE SURVEY OF IN-BAND NETWORK TELEMETRY: HISTORICAL INSIGHTS, CURRENT
DEVELOPMENTS, AND SIGNIFICANT CHALLENGES |
|
Author: |
SAMI ABBAS NAGAR ADAM, AISHA HASSAN ABDALLA HASHIM , OTHMAN OMRAN KHALIFA |
|
Abstract: |
While In-Band Network Telemetry (INT) is revolutionizing network monitoring, the
current literature lacks in-depth synthesis of the shortcomings and open
questions associated with INT architecture. This paper provides critical review
of INT technology based on systematic analysis of 80+ primary studies published
from 2015 to 2026. There is a critical gap in understanding INT technology
because of the absence of the general framework of evaluating the solutions with
regard to visibility-overhead-security-ASIC capacity trade-off. The study finds
out that 76% of INT solutions are concentrated on solving the problem of
bandwidth, however, do not take into account issues like the effect of INT
traffic on depletion of ASIC resources, TCAM/SDRAM bottleneck and security risks
associated with INT. In this paper, a novel four-dimensional
Visibility-Fidelity-Overhead-Security (VFOS) framework is developed and used to
analyze the state of research in INT and demonstrate gaps such as absence of a
standard approach in modeling heterogeneous telemetries and insufficient use of
artificial intelligence. Contributions of this paper include (i) a critical
taxonomy of distinction between measurement fidelity and operational
sustainability, (ii) the first empirical proof that the scaling capacity of INT
is limited by a non-linear scaling wall that depends on the network diameter,
and (iii) an exciting future research agenda based on adaptive telemetry and
hardware/software co-design approach. |
|
Keywords: |
In-Band Network Telemetry, Programmable Data Plane, Network Monitoring, Critical
Analysis, Software-Defined Networking, P4 Programming. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
HYBRID INCEPTIONV3-CNN-RNN FRAMEWORK FOR MULTI-DISEASE CLASSIFICATION USING
CHEST X-RAY AND BRAIN MRI IMAGES |
|
Author: |
KANIKA PAHWA , RANJIT SINGH , SARUCHI |
|
Abstract: |
The increasing application of automated medical image-based analysis has
underscored the need to come up with effective multi-disease diagnostic systems
capable of processing heterogeneous sources of data. The authors present a
Hybrid Inception V3-CNN-RNN deep learning network for the classification of a
chest X-ray and a brain MRI image in this study. The model integrates spatial
feature extraction with gated recurrent feature aggregation to enhance
diagnostic accuracy across imaging modalities. The provided methodology consists
of the complicated preprocessing pipeline, including the image normalization,
the artificial noise simulation and its removal, and the region-of-interest
(ROI) segmentation with the assistance of the K-means clustering algorithm. Deep
feature extraction is also done with a pre-trained InceptionV3 network to
extract rich hierarchical representations, and then Chi-square-based feature
ranking is evaluated to identify a compact subset of discriminative InceptionV3
features, with the number of retained features determined through
validation-based ablation analysis. The selected features are reshaped into a
fixed representation and processed by the hybrid CNN–RNN architecture. The
convolutional layers extract local patterns from the grouped feature
representation, whereas the recurrent layer performs gated nonlinear aggregation
across the feature groups. Experimental assessment of seven diagnostic classes
indicates that the presented model has better classification performance than
the case of standard CNN and fine-tuned Inception V3 models, with an accuracy of
95%, a recall of 94.37%, a F1-score of 94.68%, and a precision of 95.15%.
Combining CNN and RNN blocks enables local feature refinement and gated
integration of information across the selected feature groups, which supports
better generalization and resistance to noise. The findings support the idea
that the hybrid deep learning (DL) model has validated to be effective in
addressing challenges of single-modality diagnostic tools and real-life
multi-disease conditions and offers a steady diagnostic accuracy across the
domains of both chest and brain imaging. The article reports the importance of
integrating spatial deep feature extraction with gated recurrent feature
aggregation for medical image classification. The future research will be on the
concept of multimodal data addition, enhanced explainability, and clinical
metadata addition to facilitate useful implementation in healthcare
decision-support systems. |
|
Keywords: |
Multi-disease Classification, Medical Image Analysis, InceptionV3, Hybrid
CNN-RNN, Deep learning |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
AI AND IOT-ENABLED SMART AGRICULTURE: PREDICTIVE MODELLING FOR CROP YIELD AND
RESOURCE OPTIMIZATION |
|
Author: |
DR. SHIKHA SINGH , A ADITHYA KASHYAP2 , DR.S. VIJAYAKUMAR , DR.R. NAVEENA
BHARGAVI , DR. SYED.SHAMEEM , DR. KONARI RAJASEKHAR |
|
Abstract: |
Sustainable food production under escalating climate variability, water
scarcity, and resource constraints demands intelligent, data-driven
decision-support systems for precision agriculture. Current deep learning
methods have significant challenges with spatio-temporal representation,
multimodal sensor fusion, adaptive optimization, model explainability, and
feasibility of edge deployment. This paper introduces a novel end-to-end
intelligent farming framework called AgriFormer-RL that combines a distributed
IoT sensor network with high-resolution multimodal data capture, a physics-aware
Graph Transformer encoder, which models spatial relations in the field with
multi-head relational attention over dynamic spatio-temporal graphs, and a
temporal foundation model (Chronos/TimeMixer) for long horizon crop yield
forecasting with uncertainty quantification. A cross-attention multimodal fusion
mechanism synergistically aggregates the satellite, meteorological and in-situ
sensor representations and a Soft Actor-Critic reinforcement learning optimizer
allows for adaptive, constraint-aware irrigation and fertilizer scheduling.
Agro-explainability modules based on SHAP and attention mechanism offer decision
attribution, which is interpretable by agronomists. Comprehensive experiments
conducted over four Indian agricultural datasets - ICRISAT, IMD, MOSDAC, and
India-WRIS - in addition to simulated LoRaWAN IoT streams are statistically
superior to its baseline methods of Informer, PatchTST, and STGNN, with an
improvement of 17.5% in RMSE, water savings of 31.4%, and fertilizer savings of
24.9%. The validation results using Wilcoxon signed-rank and Friedman tests
prove that AgriFormer-RL is a scalable, interpretable, and resource-efficient
precision agriculture paradigm. |
|
Keywords: |
Precision Agriculture, Graph Transformer, Temporal Foundation Model,
Reinforcement Learning, Explainable AI, IoT, Edge Intelligence, Crop Yield
Prediction, Resource Optimization, Spatio-Temporal Learning |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
ADVANCED MANET SECURE ROUTING PROTOCOL USING SIGNAL IMAGE PROCESSING |
|
Author: |
S. HEMALATHA, JENCY A, A. GNANASEKAR DR.D.SUGUMARAN, S. K. SATYANARAYANA,
CHARANJEET SINGH |
|
Abstract: |
The mobile ad hoc networks are dynamic and have been limited in their central
management, making them vulnerable to several security attacks. This article
provides a security-based routing mechanism to overcome such a security prone
issue. It proposes an advanced framework that integrates the signal and image
processing approaches to ensure a lower energy consumption along with improving
the packet transmission security and accuracy. There are a number of advanced
techniques that attempt to circumvent the limitations of packet communication's
security, such as signal identification, coding, and compression. The proposed
work will simulate the core MANET protocols AODV and DSR in NS3, OMNET ++ and
MATLAB. Performance criteria are as follows: Packet delivery ratio (PDR),
Throughput, Energy consumption, End-to-End delay, Energy consumption, Security
metrics. The simulation results indicate that the proposed framework is suitable
in terms of reducing the power consumption by 20%, the latency by 30%, and the
PDR by 25%. Finally, the incorporation of signal, image and coding with MANET
supports for secure routing in communication networks could provide a solid
basis for the development of secure communication wireless network systems. |
|
Keywords: |
Signal Processing, Secure Routing, Coding, Image Processing, Wireless
Communication Network |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
QUANTUM ENHANCED INTRUSION DETECTION FRAMEWORK FOR SECURING INDUSTRIAL IOT
SYSTEMS |
|
Author: |
B NARASIMHA SWAMY, V N V L S SWATHI, S.SAGAR IMAMBI, CHALLAPALLI SUJANA, POTU
BHARATH, PRITHVIRAJ R, RAJESH GARAPATI, G. MERCY RANI |
|
Abstract: |
In terms of industrial automation, intelligent monitoring, and real-time
decision-making, the IIoT has revolutionized all the above aspects, even as it
has simultaneously increased the exposure of critical infrastructure to advanced
cybersecurity attacks. Traditional intrusion detection systems have poor
performance in detecting complex and evolving attacks because of a lack of
sufficient feature representation and learning of spatial and temporal network
behaviors. This study put forward a Quantum-Enhanced Intrusion Detection
Framework (QE-IDF) to protect Industrial IoT systems, which is constructed by
combining Quantum Kernel Learning (QKL), Graph Attention Networks (GAT),
Transformer-based Temporal Learning (TTL), and the Adaptive Quantum Threat
Scoring (AQTS) mechanism. The dataset created in the Edge-IIoTset benchmark was
preprocessed using data cleaning, feature normalization, and categorical
encoding. To improve the separability of the nonlinear features, quantum feature
embedding was subsequently used. The features of the communication relationships
between the IIoT devices were then captured by graph attention learning, while
the features of the temporal attack patterns were captured by transformer-based
learning. Lastly, the Adaptive Quantum Threat Scoring module predicted whether
the network traffic was normal or malicious. From the results of the experiment,
it can be noted that the framework shows better results in comparison with
traditional machine learning and modern state-of-the-art intrusion detection
models based on deep learning in terms of several criteria. This is demonstrated
by the accuracy of 99.21%, precision of 99.08%, recall of 99.31%, F1-score of
99.19%, area under the receiver operating characteristic (ROC) curve (AUC) of
0.995, and a decreased false alarm rate of 0.84%. From these results, it is
possible to say that the QE-IDF framework will serve as a strong, scalable,
reliable, and highly dependable cybersecurity solution for real-time Industrial
IoT environments. Furthermore, the framework can be used to make the systems
more cyber-resilient and enable the secure deployment of next-generation
Industry 4.0 infrastructures. |
|
Keywords: |
Industrial Internet of Things (IIoT), Quantum Kernel Learning, Intrusion
Detection System, Graph Attention Network, Transformer-based Learning,
Cybersecurity. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
SCALABLE AI/ML APPROACH TO AUTOMATED ANALOG CIRCUIT DESIGN: FROM SPICE
SIMULATION TO LAYOUT-AWARE OPTIMIZATION |
|
Author: |
K. PURUSHOTHAM , T KRISHNA MOHANA2, VAKITI SREELATHA REDDY , Dr. T.RAJASANTHOSH
KUMAR , DR.R. JAYASUDHA , GANDHIMATHI K |
|
Abstract: |
The escalating complexity of modern analog integrated circuit (IC) design,
spanning deep-submicron technology nodes from 5 nm to 65 nm, has rendered
conventional manual SPICE-based optimization both computationally prohibitive
and commercially unsustainable. SPICEFormer is a unified end-to-end artificial
intelligence (AI) framework that combines the Graph Transformers,
Physics-Informed Neural Networks (PINNs), Soft Actor-Critic (SAC) Reinforcement
Learning, Bayesian Optimization, and layout-aware parasitic extraction to
automate analog circuit design, from parsing SPICE netlist to final layout
verification. SPICEFormer reduces the number of SPICE simulation calls by 85.1%
± 3.2% on average, with a 24.9% ± 1.4% more accurate gain prediction, reduces
the layout area by 28.4% ± 2.1% on average, and improves the speed of
optimization convergence by 51.3% ± 2.8% on average across 10 independent
experimental runs using different analog circuit topologies (OTA, Folded
Cascode, Telescopic, Current Mirror, Bandgap Reference, Comparator), 3
technology nodes (TSMC 65 nm, GF 45 nm, SKY130), and with all improvements
statistically significant at p < 0.01 (paired Wilcoxon signed-rank test).
Post-layout performance degradation is reduced to 1.2–2.3% as opposed to
7.6–12.4% for previous methods. An integrated explainability module, Grad-CAM,
offers interpretable attribution of design decisions, enhancing AI-based EDA
transparency. The results show the potential of physics-informed layout-aware RL
for large-scale analog circuit generation. |
|
Keywords: |
Analog Circuit Design, Graph Transformer, Reinforcement Learning, SPICE
Simulation, Bayesian Optimization, Layout-Aware Optimization, Physics-Informed
Neural Networks |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
EFFECTIVENESS OF LLM AND CONVENTIONAL SENTIMENT ANALYSIS MODELS IN UNDERSTANDING
DIGITAL PUBLIC OPINION |
|
Author: |
ASSEGAF INSANI , TANTY OKTAVIA |
|
Abstract: |
Sentiment analysis has become an important approach for understanding digital
public opinion expressed through social media, particularly in response to
government policies. In Indonesia, discussions related to issues such as the
Indonesian National Armed Forces Bill (RUU TNI), the Criminal Code (KUHP), the
House of Representatives (DPR), and other public policy topics are widely
discussed on X (Twitter), including through hashtags such as #IndonesiaGelap
that reflect socio-political criticism and public dissatisfaction. Rather than
focusing on a single issue, this study analyzes sentiment patterns across
multiple government policy topics. This research compares the effectiveness of
Large Language Models (LLMs), namely ChatGPT and LLaMA, with conventional
sentiment analysis models, including Support Vector Machine (SVM) and IndoBERT,
in classifying Indonesian digital public opinion. The study adopts the CRISP-DM
framework covering business understanding, data understanding, data preparation,
modeling, evaluation, and deployment. Quantitative evaluation was conducted
using accuracy, precision, recall, and Macro F1-score metrics, while prompt
sensitivity and stability analysis were performed on LLaMA using zero-shot and
few-shot prompting approaches. The results show that SVM achieved the best
performance with an accuracy of 0.8658 and Macro F1-score of 0.7651, followed by
IndoBERT with an accuracy of 0.8446 and Macro F1-score of 0.7440. In contrast,
LLaMA showed lower and less stable performance across prompting configurations.
Qualitative analysis further indicates that ChatGPT demonstrates stronger
contextual reasoning and better interpretation of ambiguous and implicitly
critical texts. These findings suggest that supervised models remain more
reliable for large-scale quantitative sentiment classification, while ChatGPT
provides complementary value as an interpretative and validation tool for
complex digital public opinion analysis. |
|
Keywords: |
SVM, IndoBERT, LLM, ChatGPT, LLaMA |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
A COMPARATIVE PARAMETER OPTIMIZATION STUDY OF YOLOV8, YOLOV9, AND YOLOV10 FOR
AUTOMATED PAVEMENT DISTRESS CLASSIFICATION IN POST-DISASTER EMERGENCY
INFRASTRUCTURE ASSESSMENT |
|
Author: |
MAWARDI ROSLI , NOR HAPIZA ARIFFIN , HASLIZATUL MOHAMED HANUM , NURUL A. EMRAN ,
NURUL AKHMAL MOHD DZULKEFLI , ABDULRAZAK F SHAHATHA AL MASHHADANI , RUHAILA
MASKAT |
|
Abstract: |
Road infrastructure plays a vital role in ensuring economic growth, safe
transportation, and human mobility. However, surface anomalies such as bumps,
potholes, and structural cracks compromise road safety, raise vehicle
maintenance costs, and place an operational strain on municipal management.
Furthermore, extreme natural events—such as landslides, flash floods, soil
subsidence, and seismic ground movements—inflict sudden, catastrophic structural
trauma on transportation corridors, isolating communities and obstructing
emergency relief logistics. Traditional inspection techniques involving manual
visual surveys are labor-intensive, time-consuming, and highly prone to human
error, making them inadequate for rapid post-disaster triage. To overcome these
constraints, this study presents an automated single-stage deep learning
approach to detect and classify multi-class pavement distress for rapid
post-disaster emergency infrastructure assessment. A comparative empirical
evaluation of three distinct modern architectures within the You Only Look Once
family was executed: YOLOv8s, YOLOv9s, and YOLOv10s. The networks were subjected
to comprehensive hyperparameter sweeps, including dataset split configurations,
batch volumes, optimization functions, and epoch sizes on a refined dataset
derived from the RDD2020 repository. Under an optimized parameter layout
featuring an 80:20 train-validation split, 100 training epochs, a batch size of
32, and a Stochastic Gradient Descent (SGD) optimizer, the YOLOv8s network
achieved the highest global performance boundary, tracking a precision of
0.8246, a recall of 0.7267, an F1-score of 0.7720, and a mean Average Precision
(mAP@0.5) of 0.8089. The results indicated that YOLOv8s balanced internal
structural complexity with abstract feature generalization capacity more
effectively than its architectural successors when processing irregular and
unstandardized disaster-induced surface distortions. The deployment of the
optimized model within a lightweight web prototype operating at a total
operational latency of 6.0 ms per frame (exceeding 160 Frames Per Second)
confirmed real-time offline inference feasibility for mobile emergency vehicle
platforms navigating communication-blind disaster zones. |
|
Keywords: |
Road Damage Detection, Deep Learning, YOLOv8s, Parameter Sweep, Computer Vision,
Smart Transportation, Post-Disaster Management, Emergency Infrastructure
Assessment |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
GENERATIVE AI ENABLED SUSTAINABLE CONSUMER BEHAVIOR PREDICTION MODEL IN DIGITAL
COMMERCE PLATFORMS |
|
Author: |
MADHAVI KILARU, D. KAVITHA, ANJANI YALAMANCHILI, MYLAVARAPU KALYAN RAM, NAGABABU
GARIGIPATI, SUKANYA METTA, P. LAKSHMI RANI, SRIKANTH SAKHAMURI |
|
Abstract: |
The growth of digital commerce has brought about an increasing number of worries
regarding the sustainability of the process of consumption. While current
consumer behavior prediction systems successfully enhance personalization and
recommendation accuracy, many of them fail to account for consumers' preferences
for sustainability and environmentally conscious consumer behavior. To overcome
this, this study presents a novel Generative AI-Enabled Sustainable Consumer
Behavior Prediction Model (GenAI-SCBP) for accurate sustainable consumer
behavior prediction and its promotion in digital commerce platforms. The
proposed framework combined GANs for sustainable behavior synthesis, a new
Sustainability Propensity Index (SPI) to quantify consumers' sustainability
orientation, Transformer-based behavioral sequence learning to capture long-term
purchasing patterns, and Graph Neural Networks (GNNs) for modeling the
sustainability relationships between a consumer and a product, and finally,
Explainable Artificial Intelligence (XAI) based on SHAP values for generating
transparent recommendations. Experiments were carried out with the benchmark
Amazon Sustainability, Alibaba Consumer Behavior, and Sustainable Product
Attribute datasets, each with over 2.4 million consumer interactions. The
proposed GenAI-SCBP framework outperformed the existing machine learning and
deep learning models with an accuracy of 97.42%, a precision of 96.85%, a recall
of 96.31%, an F1-score of 96.58%, and an AUC of 97.16% through experimental
results. Additionally, the framework raised the adoption of sustainable
purchases from 51.3% to 90.0%, increased the level of engagement in
sustainability issues from 46.5% to 88.8%, and raised consumer trust in the
recommendations from 61.4% to 93.8% due to the fact that the recommendations are
understandable to consumers. The results showed that combining generative AI,
behavior-based modeling with sustainability considerations, and explanations of
the decisions significantly improved the prediction of sustainable consumer
behavior as well as encouraged sustainable consumer buying. The proposed
framework provides an effective and scalable solution for developing
transparent, trustworthy, and sustainable digital commerce ecosystems. |
|
Keywords: |
Generative Artificial Intelligence, Sustainable Consumer Behavior, Digital
Commerce Platforms, Graph Neural Networks, Explainable Artificial Intelligence,
Sustainability Propensity Index. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
META-CONTEXTUAL HUMAN RESOURCE COMPETENCY INFERENCE THROUGH POLYSEMANTIC
NEURO-GRAPH FUSION AND AUTONOMOUS MULTI-AGENT POLICY DYNAMICS |
|
Author: |
DR AKASH DAULATRAO GEDAM, DR. TAVITI NAIDU GONGADA, SANDYA SNEHA SRI SEERA, DR.
P N V SYAMALA RAO M, DR GURU BASAVA ARADHYA S, DR. VEERA ANKALU. VUYYURU, PROF(
DR) MAITRI |
|
Abstract: |
The growing complexity of contemporary organizational ecosystems has unveiled
the drawbacks of conventional HR analytics, which are based on obviously fixed
rules, linear designs and data silos that do not encompass the nuances of
context, relations and behavior in workforce competency patterns. Overcoming
these limitations, this paper presents PolyNeuroGraphMARL a meta-contextual
inferential model of HR competencies that combines polysemantic neural encoding
with graph-transformer structural reasoning and autonomous multi-agent
reinforcement learning. The system integrates heterogeneous HR cues, employee
characteristics, behavioral histories, written records and organizational
graphs, into a single representation that is able to model latent semantics,
dynamical interactions, and role appropriateness changes. Under the framework of
multi-agent policy dynamics, the role alignment, training suggestions, and
performance forecasting are continuously optimized by the framework as the
organizational environment is in non-stationary state. Empirically tested
results indicate that it has greatly improved predictive accuracy and structural
faithful, with 93.5% accuracy, 94.0% F1-score, 0.110 RMSE, 0.91 edge precision
and Ravg/Rmax ratio of 0.87 that is more accurate than the state-of-the-art
predictive analytics baselines. The system provides strong decision support in
the areas of talent allocation, internal mobility, team composition, and
responsive workforce planning in the areas of sectors that must constantly
evaluate competency and organizational knowledge. The study paves the way to a
new frontier in HR computational modeling by providing an extensively scalable,
context sensitive, and interpretable neuro-graph reasoning structure that can be
used to provide anticipatory and data-driven strategic workforce management. |
|
Keywords: |
Five Keywords are Required Separated By Commas (Capitalize Each Work Italic) |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
QUANTUM MACHINE LEARNING BASED REAL TIME EPILEPTIC SEIZURE PREDICTION USING
WEARABLE EEG SENSORS |
|
Author: |
GATTU SHRAVANI , VENKAT LAKSHMI K , PAPPULA MADHAVI , RANJITH KUMAR CHINNAM ,
ARAVA NAGASRI , SRINIVAS MULKALAPALLI , DASARI YUGANDHAR , CH.LAVANYA SUSANNA |
|
Abstract: |
Epileptic seizures can be described as unpredictable neurological phenomena
which represent great dangers for the patients' safety and need constant
healthcare monitoring. The use of wearable electroencephalogram (EEG) devices
can allow accurate real-time seizure forecasting, enabling early warning and
improved clinical intervention. However, traditional machine learning and deep
learning algorithms do not have the ability to effectively exploit the complex
nonlinear properties and long-term temporal dependencies presented in EEG
signals. In this paper, we introduce the Quantum Machine Learning-Based Seizure
Prediction Network (QML-SPNet), which consists of the hybrid architecture based
on Quantum Feature Encoding (QFE), Variational Quantum Neural Networks (VQNN),
Transformer-based Temporal Learning (TTL), and Adaptive Seizure Risk Assessment
(ASRA). First of all, the preprocessing of the input data is performed by
cleaning artifacts, normalization, and segmentation of the raw wearable EEG
signals. The processed signals are encoded into a high-dimensional quantum
feature space and analyzed through quantum feature learning and temporal
analysis via a Transformer network for detecting pre-seizure patterns. Finally,
the Adaptive Seizure Risk Assessment module calculates the probability of
seizure occurrence and provides early-warning alarms. The proposed method allows
representing the input data in the form of quantum feature vectors in the
high-dimensional quantum feature space. Then, quantum feature learning is
performed along with the temporal analysis of the signals with the help of a
Transformer network to identify pre-seizure patterns. The Adaptive Seizure Risk
Assessment module calculates the probability of seizures and generates
early-warning alerts. The proposed QML-SPNet framework was tested on the CHB-MIT
Scalp EEG dataset and achieved 98.94% accuracy, 98.76% precision, 99.08% recall,
98.92% F1-score, an AUC of 0.994, and prediction latency of 31 ms, which exceeds
the performance of traditional machine learning and deep learning frameworks.
Therefore, the proposed framework is an effective, low-latency, and highly
reliable approach for wearable real-time epileptic seizure prediction, with
significant potential for improving patient safety and supporting
next-generation intelligent healthcare monitoring systems. |
|
Keywords: |
Epileptic Seizure Prediction, Quantum Machine Learning, Wearable EEG Sensors,
Variational Quantum Neural Network, Transformer-based Learning, Real-Time
Healthcare Monitoring. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
HIERARCHICAL CROSS-MODAL SPATIO-TEMPORAL TRANSFORMER FOR EARLY ALZHEIMER’S
DISEASE PREDICTION |
|
Author: |
SURESH BABU SINGAMSETTI , MURALI GUDIPATI |
|
Abstract: |
Early detection of Alzheimer’s disease (AD) and prediction of its progression
from cognitively normal (CN) to mild cognitive impairment (MCI) and ultimately
AD remain challenging because disease-related alterations are distributed across
multiple modalities and evolve over time. Recent CNN-, Vision Transformer-, and
multimodal learning approaches have improved AD classification by learning
discriminative anatomical and functional representations; however, existing
methods predominantly emphasize single-timepoint diagnosis, limited modality
combinations, or independent temporal modeling. Recent longitudinal and
transformer-based approaches have addressed some of these limitations, but
multimodal interactions, hierarchical disease-stage transitions, anatomical
priors, and prediction uncertainty are generally not modeled jointly within a
single framework. To address these limitations, this study proposes a
Hierarchical Cross-Modal Spatio-Temporal Transformer (HCST-Former) that extends
existing transformer-based AD modeling from predominantly cross-sectional
representation learning toward unified multimodal and longitudinal disease
modeling. The principal contribution is a unified architecture that integrates
Cross-Modal Attention Fusion (CMAF) for explicit interactions among structural
MRI, PET, cognitive, genetic, and clinical information; Longitudinal Temporal
Encoding (LTE) for learning disease trajectories across multiple timepoints;
Hierarchical Disease-Stage Modeling (HDSM) for explicitly representing the
CN→MCI→AD progression hierarchy; Graph-Guided Anatomical Attention (GGAA) for
incorporating prior knowledge of AD-relevant brain regions; and
Uncertainty-Aware Prediction (UAP) for estimating predictive confidence. In
contrast to approaches that address these aspects separately, HCST-Former
jointly learns multimodal, temporal, hierarchical, and anatomically informed
representations while considering computational scalability. The research
questions addressed by the framework therefore concern whether explicit
cross-modal interaction, longitudinal trajectory modeling, hierarchical stage
representation, anatomical guidance, and uncertainty estimation can collectively
improve AD detection and progression prediction. Experiments on the Alzheimer’s
Disease Neuroimaging Initiative (ADNI) dataset show that HCST-Former achieves
93.7% accuracy, 96.2% AUC, 92.8% sensitivity, and 94.5% specificity for
disease-stage classification. For longitudinal progression prediction, the model
achieves a C-index of 0.91 and time-to-conversion MAE of 8.7 months. Ablation
experiments demonstrate progressive improvements following the incorporation of
CMAF, LTE, HDSM, and GGAA, while the complete model achieves the reported
performance with 48.3M parameters, 92.6 GFLOPs, and 8.1 GB GPU memory. These
findings indicate that integrating complementary multimodal, longitudinal,
hierarchical, anatomical, and uncertainty-aware mechanisms within a unified
transformer provides a systematic extension of existing AD modeling approaches. |
|
Keywords: |
Alzheimer's Disease, Multi-Modal Transformer, Longitudinal Modeling, Cross-Modal
Attention, Medical Imaging, Interpretability |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
QSAF-IDS: QUANTUM-ENHANCED SECURE ADAPTIVE FRAMEWORK FOR SYBIL ATTACK DETECTION
AND POST-QUANTUM ENCRYPTED DATA TRANSMISSION IN URBAN VEHICULAR AD HOC NETWORKS |
|
Author: |
DR. NITHA C VELAYUDHAN , DR.MONG-FONG HORNG , DR. CHUN-CHIH LO , DR. SIVA
SHANKAR S |
|
Abstract: |
Urban Vehicular Ad Hoc Networks (Urban VANETs) constitute the communications
core of future smart city intelligent transportation systems, providing
real-time data exchange between vehicles (V2V) and vehicle-to-infrastructure
(V2I). Urban VANETs present a uniquely hostile security environment: rapidly
changing topologies from dense traffic and signalized junctions; severe signal
fading from multi-story buildings, tunnels, and urban canyons causing multipath
NLOS conditions; high RSU deployment density creating focal points for
coordinated attacks; and hard sub-100ms safety-critical latency mandates. The
Sybil attack — in which adversarial nodes impersonate multiple fake identities
to subvert routing and trust — is uniquely exacerbated in urban environments.
Simultaneously, prevailing cryptographic foundations (RSA, ElGamal, standard
ECC) remain vulnerable to Shor’s algorithm on quantum computers, creating a dual
security gap this work resolves.
This paper presents QSAF-IDS, a five-phase
quantum-secure adaptive framework: (1) QMKHM — Quantum-Modified K-Harmonic Means
behavior-aware clustering; (2) AFWA — Adaptive Floyd-Warshall multi-objective
cluster head election; (3) CMEHA-BiLSTM — feature-driven Sybil detection with
metaheuristic hyperparameter optimization; (4) PQMD5-ECC — post-quantum
encryption combining Kyber-768 lattice-based KEM with three-key ECC; and (5)
RL-ATM — Reinforcement Learning Adaptive Trust Management with Reward-Penalty
Trust Scoring (RPTS).
NS-3/SUMO simulations across 50–300 vehicle urban
scenarios yield: accuracy = 99.12%, F-measure = 98.87%, specificity = 99.23%,
FNR = 0.99%, FPR = 1.18%, malicious node isolation rate = 98.71%, encryption
security level = 99.1%, and minimum encryption latency of 6,187ms at 500 nodes.
QSAF-IDS outperforms all evaluated baselines (DNN, ANN, SVM, KNN, LSTM, CNN-IDS)
across all metrics. |
|
Keywords: |
Vehicular Ad Hoc Networks (VANET), Urban VANET, Sybil Attack Detection,
Quantum-Modified K-Harmonic Means, Bidirectional LSTM (BiLSTM), Chaotic Map
Elephant Herding Algorithm (CMEHA), Post-Quantum Cryptography, Kyber-768
Lattice-Based Encryption, Intrusion Detection System, Reinforcement Learning
Trust Management |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
DESIGN OF AN OPTIMAL PI SPEED CONTROLLER FOR BLDC MOTOR DRIVES USING MODIFIED
WHALE OPTIMIZATION ALGORITHM |
|
Author: |
CH. TRINAYANI , RAVI SRINIVAS LANKA |
|
Abstract: |
This paper proposes a novel, nature-inspired Modified Whale Optimization
Algorithm (MWOA) for the optimal tuning of a Proportional-Integral (PI) speed
controller in a Brushless DC (BLDC) motor drive. The IWOA metaphorically models
optimization processes based on the transmission dynamics of a virus and the
population's immune response, balancing global exploration and local
exploitation to minimize the Integral Square Error (ISE) objective function. A
detailed dynamic model of the BLDC motor is developed in the MATLAB/Simulink
environment to evaluate system performance. The proposed MWOA-tuned PI
controller is rigorously tested under various operating conditions, including
sudden load changes and step variations in reference speed. Its performance is
benchmarked against well-established optimization algorithms, namely Particle
Swarm Optimization (PSO) and Differential Evolution (DE), Whale Optimization
Algorithm (WOA) for a fair comparison. Simulation results show that the
MWOA-based controller has a better dynamic performance. It has faster settling
time, less overshoot and is more resistant to disturbances. The results show
that the proposed MWOA works to improve drive performance in all of the
operational scenarios that were tested. |
|
Keywords: |
BLDC motor, PI Controller, Speed Deviation, Modified Whale Optimization
Algorithm |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
ADVANCED BREAST CANCER PREDICTION BY COMBINING LARF FEATURE SELECTION, DYNAMIC
HYPERPARAMETER TUNING AND ENSEMBLE LEARNING |
|
Author: |
B. NARASIMHA SWAMY, L. PUNITHA, NIRMALA DEVI K, K. VENU GOPAL, NAGAMANI
CHIPPADA, S. LALITHA, S. SINDHURA |
|
Abstract: |
The classification of breast cancer by ML depends on the features selected, the
model configuration and the generalization ability of the classifiers. In this
study, an integrated framework of LARF-based feature selection, dynamic
hyperparameter tuning and ensemble learning is proposed for breast cancer
classification. For experimental evaluation, Wisconsin Diagnostic Breast Cancer
(WDBC) was used which contains 569 observations and 30 numeric features. LARF
was used to determine the most relevant predictors and then Logistic Regression,
Support Vector Machine, K-Nearest Neighbors, Random Forest and Gradient Boosting
classifiers were created. Dynamic hyperparameter tuning using stratified
five-fold cross validation was used to tune the configuration of the models and
the best classifiers were combined using a soft voting ensemble. The accuracy,
precision, sensitivity, specificity, F1 score and ROC-AUC were used to evaluate
the performance. Results from the experimental setup showed that the integration
of feature reduction, model optimization and ensemble integration can improve
the classification performance compared to baseline and individual models. The
proposed method offers an efficient and repeatable framework to classify breast
cancer using machine learning techniques. |
|
Keywords: |
Breast Cancer Classification, LARF Feature Selection, Dynamic Hyperparameter
Tuning, Ensemble Learning, Machine Learning. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
ADAPTIVE MODEL COMPRESSION OF YOLO11-N FOR ENERGY-EFFICIENT REAL-TIME DRONE
VISION |
|
Author: |
Dr.K.V.S.PRASAD, DR.K.E. PURUSHOTHAMAN, Dr.S.UMA, RAMESH KUMAR, M.R. EZILARASAN,
Dr.D. SUDHAGAR, Dr.R. SENTHAMIL SELVAN |
|
Abstract: |
Unmanned aerial vehicles (UAVs) equipped with onboard vision systems are
increasingly deployed for surveillance, traffic monitoring, and disaster
response, yet their limited battery capacity and embedded compute budget
severely constrain the use of modern convolutional object detectors. This paper
presents an Energy-Adaptive CNN Compression (EACC) framework, which combines the
structured channel pruning, knowledge distillation, and mixed-precision
quantization with an energy-aware controller that dynamically adjusts the
trade-off between the detection accuracy and on-board energy consumption.
Conventional compression pipelines optimize accuracy and model size
independently, while the proposed controller optimizes the compression ratio,
directly affecting the pruning and quantizing decisions at each layer, as a
constrained optimization problem in which the energy budget of the target
embedded platform is directly related to compression ratio. The framework is
developed based on the YOLO11-N baseline and tested on the VisDrone2021
benchmark, which is a large-scale benchmark with small, dense, and partially
occluded objects taken in various urban and suburban environments. Experimental
results show that the proposed model can achieve about 76% reduction in the
number of parameters and 66% reduction in energy consumption per frame compared
to the uncompressed baseline, while maintaining more than 92% of the mAP@0.5 of
the baseline and keeping the model running in real-time on a Jetson Orin Nano
platform. The results show that it is possible to implement energy-adaptive,
accuracy-aware compression as a viable strategy for deep object-detection
deployment on a resource-constrained aerial platform. |
|
Keywords: |
UAV Object Detection, Model Compression, Structured Pruning, Knowledge
Distillation, Quantization, Energy-Aware Deep Learning, VisDrone2021, Edge AI. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
AN APPROACH TO CLOUD SECURITY USING MACHINE LEARNING-BASED ATTACK DETECTION AND
MULTI-LAYER SECURITY AND GAP ANALYSIS |
|
Author: |
DR.N. SRIHARI RAO, DR.P. VINAY BHUSHAN , DR. ALGUBELLY YASHWANTH REDDY, DR.
NARASIMHA CHARY CH , DR. S .RAVI KUMAR , DR. KRISHNA ANNABOINA, DR. S SIVA
SANKARA RAO, DR. SHANKER C |
|
Abstract: |
Cloud computing environments are increasingly exposed to sophisticated and
adaptive cyberattacks that often bypass traditional static, single-layer, and
signature-based security mechanisms. To address these limitations, this study
proposes a Machine Learning–based Multi-Layer Cloud Security Architecture
(ML-MLCSA) that integrates adaptive attack detection with continuous security
gap analysis across multiple cloud layers. The proposed framework employs a
weighted ensemble learning strategy combining Support Vector Machine (SVM),
Gradient Boosting, and a lightweight Deep Neural Network (DNN) to enable
real-time threat classification at the network, virtualization, and
application/API layers. By dynamically adjusting detection thresholds and
updating learning rules, the architecture enhances resilience and mitigates
static vulnerabilities in cloud access control policies. The effectiveness of
the proposed approach is validated using the CADE dataset, comprising 151,200
cloud network flows. Experimental results demonstrate that the stacked ensemble
model (SVM + AdaBoost → MLP) achieves an accuracy of 98.6%, an F1-score of
98.0%, and a macro-averaged AUC of 0.987, outperforming existing machine
learning and case-based reasoning approaches, including Logistic Regression,
SVM, and Random Forest. The results confirm that integrating machine
learning-driven attack detection with multi-layer cloud security mechanisms
significantly enhances detection accuracy, adaptability, and overall cloud
security posture. |
|
Keywords: |
Cloud Security, Machine Learning, Multi-Layer Defense, Attack Detection,
Ensemble Learning, Gap Analysis |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
A CONTEXTUAL EMBEDDING FRAMEWORK WITH GRAPH RECURRENT MODELING FOR EMOTION
DETECTION AND CLASSIFICATION ON SOCIAL MEDIA DATA |
|
Author: |
K.HEMAKIRTHIGA, J.ARUNADEVI |
|
Abstract: |
In the modern era, a majority of users are accessing the internet globally for
communication through text, images, audio, and video. People from various
backgrounds share data on discussing current events and projecting their
personal views on social media. There is a necessity for identifying and
understanding the behaviours of massive textual data on people by examining
their emotions. People expressing their feelings through tweeter real reactions,
Facebook posts, and more. Nevertheless, handling large-scale data generated
across diverse social media platforms is challenging. In Natural Language
Processing (NLP), emotion recognition is a frequently studied task that can
identify these kinds of emotions. Deep learning (DL) approaches are widely used
for single-modal text-based emotion analysis. In this paper, we design a
Contextual Embedding Framework with Graph Recurrent Modeling for Emotion
Detection in Social Media (CEGRM-EDSM) approach. The main purpose of the
CEGRM-EDSM system is to develop an effectual and intelligent framework for
social media data classification and emotion detection by leveraging advanced
deep learning approaches. At primary stage, a comprehensive text pre-processing
pipeline is implemented to enhance data quality and semantic consistency,
comprising elimination of URLs, noise, and hashtags, hashtag segmentation
through word partitioning, substitution of slang and abbreviations, spelling
correction, emoji interpretation, and removal of unwanted numerals. Following
this, the high-quality contextual representations are generated using the
DeBERTa V3-base word embedding model to capture rich semantic and syntactic
dependencies within social media text. For classification and emotion
recognition, a hybrid graph attention network with bidirectional long short-term
memory architecture has been deployed to model both relational word dependencies
and sequential contextual patterns effectively. A wide range of simulation
analyses were conducted on two benchmark datasets. The comparative outcomes
exhibit the superiority of proposed CEGRM-EDSM methodology over recent
approaches in terms of various evaluation metrics. |
|
Keywords: |
Social Media, Emotion Detection, Data Classification, Graph Attention Network,
Contextual Embedding |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
Title: |
ENHANCING LOGO VERIFICATION AND RECOGNITION IN E-COMMERCE USING TRIPLET-LOSS
DRIVEN DEEP EMBEDDING NETWORKS |
|
Author: |
PREETI C M , Dr. T SANTHI SRI |
|
Abstract: |
The online business has contributed to the rapid proliferation of counterfeit
products with forged logos associated with serious risks to the reputation of
the brand and consumer trust. In these cases, when subtle intra-class changes
are involved, conventional classification-based convolutional neural networks
(CNNs) are not particularly effective. In order to overcome this problem, we
introduce a triplet-loss-based deep embedding framework that is based on the
EfficientNet-B0 backbone. The model is trained on a discriminative embedding
space with the authentic logos being tightly grouped and counterfeit logos far
apart.An adaptive training pipeline is developed in such a way that
anchor-positive-negative triplets are dynamically generated over a curated set
of logos with data augmentation, image preprocessing, and L2-normalized
embeddings to increase generalization. To test this, nearest-neighbor
classification is used on the embedding space with high accuracy and robustness.
The proposed architecture has some scalability, and can verify a logo of unseen
classes by simple distance comparisons in the learned space. The results of the
experiments demonstrate the high accuracy, recall, and F1-score in comparison to
the conventional CNN-based classifiers, and the effectiveness of the model in
real e-commerce verification systems. However, e-commerce platforms struggle
with accurately verifying and recognizing logos because of their different
orientations, sizes, illumination and image quality. Traditional approach has a
problem with differentiating between a real logos from similar-looking
counterfeit or altered logos. Hence, a Triplet-Loss Driven Deep Embedding
Network (Triplet-LDDN) is proposed to boost the accuracy of logo verification
and robustness of the recognition. |
|
Keywords: |
Triplet Loss , Logo Verification · Efficientnet ,· Fake Logo Detection ,· Image
Similarity · Convolutional Neural Networks ,· E-Commerce Security ,· Feature
Learning ·, Visual Similarity Modelling
|
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th September 2026 -- Vol. 104. No. 18-- 2026 |
|
Full
Text |
|
|
|
