|
|
|
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
June 2026 | Vol. 104
No.11 |
|
Title: |
FORTIFYING FIREWALLS AGAINST EVOLVING DDOS ATTACKS WITH CONTRASTIVE AI AND LLMS |
|
Author: |
SRINIVASARAO DHARMIREDDI , DESIDI NARSIMHA REDDY , MRS. MODUGULA SIVAJYOTHI ,
MARISETTI KALYAN RAMUDU , ELANGOVAN MUNIYANDY , UDAY KIRAN KASI |
|
Abstract: |
Distributed Denial of Service (DDoS) attacks continue to threaten network
availability and challenge firewall decision logic. Existing firewall-centric
defences often fail to detect low-rate, multi-vector, and carpet-bombing DDoS
patterns while meeting inline latency constraints. This paper proposes
Contrastive-LLM Firewall (C-LLM-FW), a hybrid defence that integrates
self-supervised contrastive representation learning with a distilled Large
Language Model (LLM) encoder to produce context-aware flow representations for
real-time classification and mitigation. The proposed method pretrained a
contrastive encoder on benign flows to form a stable latent manifold and then
used an LLM encoder as a contextualizer over short flow-sequence tokens; a
compact cross-attentive fusion and a lightweight classifier issued firewall
decisions while an enforcement agent applied actions in the Data Plane.
Experiments were performed on the publicly available
BCCC-cPacket-Cloud-DDoS-2024 corpus and on a simulated urban IoT emulation
dataset for large-scale botnet scenarios. The method was compared to XGBoost as
a classical baseline and to DoLLM as an advanced LLM baseline. C-LLM-FW improved
F1 to 93.5% on the primary dataset, an absolute gain of 2.8 points versus DoLLM
and 9.4 points versus XGBoost; inference latency reduced by ~28 ms on average
and throughput resilience doubled at 1 Gbps attack injection. The results
demonstrate that by leveraging the capabilities of contrastive AI alongside LLM
encoders, practical gains in firewall detection accuracy and robustness can be
achieved while maintaining production-grade inference latency that is ideal for
deployment on the edge. |
|
Keywords: |
DDoS Detection, Contrastive Learning, Large Language Models (LLM), Firewall,
Hybrid Model. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
LAB-CLAHE ENHANCEMENT AND SYMPTOM-CENTERED PREPROCESSING FOR ROBUST LEAF DISEASE
DETECTION |
|
Author: |
G NAGI REDDY, V VIJAYA KUMAR |
|
Abstract: |
Plant diseases represent a great challenge to agricultural production, with
significant economic losses and substandard food. The early symptom of plant
disease is often small, low contrast and contaminated with background noise, and
the symptom is hard to differentiate from healthy leaf area under different
illumination and background conditions. The key to efficient management lies in
the ability to detect and apprehend in time which consequently can protect
yields. The research indicates a hybrid framework of plant-disease diagnosis
(integrating preprocessing based on symptoms, handcrafted feature extraction and
deep learning classifier) that is presented in the study. Bilateral filtering
and LAB-CLAHE are also used in preprocessing phase to reduce noise as well as
increase the visibility of disease signs. Disease-specific characteristics are
represented in a set of manually designed descriptors like the gradient, edge,
texture, shape, and color features. The neural networks in question are combined
with such features as custom CNN, YOLO, and NASNetMobile, creating a complete
pipeline of detection. Better activation, which is called Enhanced Tans Network
A LL (tanhReLU), is suggested to enhance nonlinear feature learning. Experiments
were conducted using a dataset of 10, 000 tomato leaf images with ten disease
classes, in more than 25 epochs. The end result CNN had a training, validation,
and test accuracy of 93.02,78.33,78.12. Visualizations by use of feature-map and
grad-cam established that the network was paying attention to infected leaf
areas. Experimental results indicate that the proposed approach has better
robustness and performance in detecting tomato leaf disease images, and the
proposed visualization approach confirms that the proposed model provides
attention to the infected regions. All in all, the hybrid solution proves to be
more robust and has a greater diagnostic function than the conventional methods.
In conclusion, the proposed hybrid method is found to be an effective solution
in terms of efficient detection of plant diseases in complex real-world
applications. |
|
Keywords: |
Hybrid Feature Extraction, Convolutional Neural Network, YOLO, Nasnetmobile,
Bilateral Filtering, CLAHE, LBP, HOG, Custom Activation Function, Deep Learning
In Agriculture. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
CAN HYBRID DEEP LEARNING WITH DUAL EXPLAINABLE AI ENABLE CLINICALLY TRUSTWORTHY
AUTOMATED DIABETIC RETINOPATHY GRADING? |
|
Author: |
DR K.E. PURUSHOTHAMAN, K. JYOSTNA, ROOMANA HASAN, DR MAHAVIR A. DEVMANE,
GUNASUNDARI B, AMIT VERMA, DR R. SENTHAMIL SELVAN, DR N. DHASARATHAN |
|
Abstract: |
Manual screening on a broad scale is still inefficient, expensive, and
susceptible to inter-grader variability, even though diabetic retinopathy (DR)
is the top avoidable cause of blindness globally. A major obstacle to real-world
deployment is the absence of regulatory permission and the inherent limitations
on physician trust caused by deep learning models, even if these models have
attained clinical-grade accuracy in automated DR grading. Our solution to this
transparency issue is ResViT FusionNet, a CNN-Transformer hybrid that uses
ResNet50's local lesion sensitivity in conjunction with a lightweight Vision
Transformer's global contextual modelling and a dual Explainable AI (XAI)
pipeline that incorporates Grad-CAM and SHAP. On the APTOS-2019 benchmark
(5-class grading, n = 6,000 images), ResViT FusionNet achieves Accuracy =
0.9301, macro-F1 = 0.9275, and Cohen's Kappa = 0.8935, significantly
outperforming standalone ResNet50 (ΔF1 = 0.0145, p < 0.01) and ViT baselines.
When compared to expert annotations, Grad-CAM heatmaps pinpoint clinically
significant lesions with a median IoU > 0.58, and SHAP attributions pinpoint the
exact patches responsible for each grade choice. The combined XAI outputs
enhanced referral confidence and mistake detection, according to an informal
physician review. These findings support the idea of reliable AI in
ophthalmology by showing that automated DR screening with interpretable accuracy
is possible. |
|
Keywords: |
Diabetic Retinopathy, Explainable Artificial Intelligence (XAI), SHAP, Grad-CAM,
Deep Learning, Medical Imaging, Interpretability. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
HIERARCHICAL GRAPH ATTENTION NETWORK WITH BIDIRECTIONAL LSTM FOR REAL-TIME
MULTI-CLASS NETWORK INTRUSION DETECTION |
|
Author: |
DR.CH.V.SIVARAMPRASAD, S VINOD KUMAR, KORLA SWAROOPA, DR. PAVADA SANTOSH, DR K
NAGARAJU, DR A PHANI SRIDHAR, DR. HARI JYOTHULA, DR SIVA KUMAR SUBRAMANIAN |
|
Abstract: |
Network intrusion detection remains a critical unsolved challenge because modern
cyberattacks continuously evolve in sophistication, volume, and diversity,
rendering traditional signature-based and shallow machine learning systems
increasingly inadequate. Despite advances in deep learning, existing models fail
to simultaneously capture the relational topology and temporal dynamics of
network traffic, leaving significant detection gaps for rare and novel attack
variants. This paper addresses that gap by proposing HGAT-LSTM, a unified
architecture that achieves 99.14% classification accuracy with a false positive
rate of just 0.91%, demonstrating that joint graph-structural and temporal
modeling is both technically feasible and operationally viable for real-time
intrusion detection.
To meet the challenge of having to detect contemporary
threats posed to modern network infrastructure by increasingly advanced
cyberattacks, intelligent intrusion detection systems (IDS) need to perform
real-time accurate multi-class classification across heterogeneous traffic
patterns. Traditional features are limited in expressiveness, topological
relations among network entities cannot be modeled, and generalization to
zero-day attack variants is poor with such approaches [7]. The deep learning
techniques are much more capable, but they ignore the rich structural
dependencies of networked environments by treating network flows as independent
observations. To this end, in this paper, we develop and propose a new
Hierarchical Graph Attention Network integrated Bidirectional Long Short-Term
Memory encoder architecture called HGAT-LSTM, which is specialized for
high-performance network intrusion detection. The proposed architecture builds a
dynamic attributed graph based on observed network sessions, then uses
multi-scale graph attention to process each of the three
hierarchically-organized layers (i.e. with differentiable pooling) and finally
combines structural embeddings of the graphs (i.e. learned via GTN) and temporal
sequences from the bidirectional LSTM using cross-attention exploration module.
To combat the extreme class imbalance commonly found in intrusion detection
datasets, a focal cross-entropy loss function is employed for model training.
On four benchmark datasets, including NSL-KDD, UNSW-NB15, CICIDS-2017 and
CICIDS-2018 over extensive experiments show that the proposed method HGAT-LSTM
outperforms six competitive baselines (CNN-LSTM, Transformer IDS, GCN-LSTM
architectures) with state-of-the-art performance accuracy of 99.14%,
macro-averaged F1-score of 98.91%, and AUC-ROC of4962 0.9987. Ablation studies
further validate the essential role of each architectural part, and
cross-dataset evaluation demonstrates improved generalization ability. The
inference latency of 6.8 ms per sample makes HGAT-LSTM suitable for deployment
in real-time intrusion detection pipelines. |
|
Keywords: |
Intrusion Detection System, Graph Attention Network, Bidirectional LSTM, Network
Security, Deep Learning, Hierarchical Pooling, Multi-Class Classification, Focal
Loss |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
A MOTION GRAPHICS–ENHANCED LEARNING ENVIRONMENT FOR IMPROVING UNDERGRADUATE
STUDENTS’ UNDERSTANDING OF DEPRESSION IN SOUTHERN THAILAND |
|
Author: |
VEERAYUT BOONPIT, OPHAT KAOSAIYAPORN, SAKKARIN CHONPRACHA, JIRAWAT TANSAKUL |
|
Abstract: |
Depression among undergraduate students has become an increasing concern in
higher education, particularly because limited understanding of symptoms,
severity, treatment options, and help-seeking channels may prevent students from
recognizing and responding appropriately to mental health problems. Although
motion graphics have been used as digital learning media, limited empirical
evidence exists on how they can be integrated into a structured learning
environment to enhance students’ understanding of depression. This study aimed
to: (1) develop and evaluate the quality of a motion graphics–enhanced learning
environment, (2) compare undergraduate students’ understanding of depression
before and after learning through the developed environment, and (3) examine
students’ satisfaction with the learning environment. The sample consisted of
160 undergraduate students selected using stratified random sampling by year of
study. The research instruments included a motion graphics–enhanced learning
environment, a depression understanding test, and a satisfaction questionnaire.
Data were collected through a pretest, learning activities within the developed
environment, a posttest, and a satisfaction assessment. The data were analyzed
using mean, standard deviation, and paired-samples t-test. The findings revealed
that the developed learning environment was rated at a very high level of
quality. Students’ posttest scores (M = 13.03, S.D. = 1.37) were significantly
higher than their pretest scores (M = 9.63, S.D. = 1.67) at the .01 level. In
addition, students’ overall satisfaction with the learning environment was at
the highest level (M = 4.93, S.D. = 0.30). The study contributes empirical
evidence that integrating motion graphics into a structured learning environment
can support mental health literacy by improving students’ cognitive
understanding of depression and providing an engaging digital learning approach
for higher education contexts. |
|
Keywords: |
Learning Environment; Motion Graphics; Depression; Undergraduate Students;
Mental Health Literacy; Learning Outcomes |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
CLOUD COMPUTING ADOPTION IN JORDANIAN E-GOVERNMENT: AN INTEGRATED TOE–TAM MODEL |
|
Author: |
RAGHED ALKHASAWNEH, ZAIHISMA CHE COB, ALIZA BINTI ABDUL LATIF, NOUR QATAWNEH |
|
Abstract: |
Cloud computing is a key enabler of digital transformation in e-government, yet
its adoption in public-sector organizations is shaped by technological,
organizational, and environmental factors. This study examines the determinants
of cloud computing adoption intention in the Jordanian public sector by
integrating the Technology–Organization–Environment (TOE) framework with the
Technology Acceptance Model (TAM) and trust theory. It investigates how relative
advantage, security, top management support, and government regulations
influence adoption intention through the mediating roles of perceived usefulness
and trust in cloud technology. A quantitative survey was conducted among
employees of the Civil Status and Passport Department in Jordan, producing 373
valid responses. Data were analyzed using Partial Least Squares Structural
Equation Modeling (PLS-SEM). The results show that top management support is the
strongest determinant of adoption, significantly influencing both perceived
usefulness and trust. Relative advantage positively affects both mediators,
while security mainly strengthens trust. Government regulations enhance
perceived usefulness but do not directly influence trust. Both perceived
usefulness and trust significantly predict cloud adoption intention. The study
contributes theoretically by integrating TOE, TAM, and trust theory to explain
cloud adoption in e-government. Practically, it provides guidance for
policymakers and public-sector managers to strengthen leadership support,
regulatory frameworks, and trust in cloud technologies to accelerate digital
transformation. |
|
Keywords: |
Cloud Computing, E-Government, Jordan, TAM-TOE, Trust, PLS-SEM. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
AN INTELLIGENT MULTI-MODEL APPROACH FOR STUDENT PERFORMANCE PREDICTION USING
TEMPORAL–SPATIAL LEARNING WITH CONVOLUTIONAL NEURAL NETWORKS AND GRADIENT
BOOSTING |
|
Author: |
BRIJESH KUMAR VERMA, DR. NIDHI SRIVASTAVA, DR. AJAY KUMAR BHARTI |
|
Abstract: |
The present research investigation is aimed at effectualising a more
corroborative and accurate educational predictive analytics so as to facilitate
and enhance data-driven academic decision-making for meeting the learning
objectives in Higher Educational Institutions. Given the rapid availability of
data drawn from the academic records, behavioral logs, and learning management
systems (LMS) it is now more efficacious to apply sophisticated machine learning
algorithms for generating the desired results to optimise on the benefits for
both the institutions and the learners. In this league, the study suggests a
hybrid predictive framework that combines Gradient Boosting and Temporal-Spatial
Convolutional Neural Networks (CNN) for effecting improved student performance
prediction. With the intent of finding the most pertinent academic, behavioral,
and socio-demographic characteristics, the study first employs data preparation,
normalization, and Recursive Feature Elimination (RFE). Thereafter, the
progression patterns and intricate feature interactions are accessed by using a
Temporal-Spatial CNN. To increase the accuracy of the predictions and accomplish
generalization, the study uses Gradient Boosting classifier. To assess the
efficacy of the suggested model, a train-test split of 80-20 is done on a
multi-dimensional student dataset. As authenticated by the experimental data, in
terms of accuracy and efficacy, the proposed multi-model architecture surpasses
the results drawn from the conventional machine learning and standalone deep
learning techniques. With an accuracy of 96.8% and an increased F1-score and AUC
(0.98), the devised model seeks to efficiently categorize the students into risk
groups. Thus, enabling targeted and prompt interventions to provide academic
assistance and niche the intended learning outcomes in the process. Although the
reliance on dataset could differ amongst institutions, the approach proffers a
confirmative and credible solution for educational predictive analytics. |
|
Keywords: |
Student Performance Prediction, Educational Data Mining, Hybrid Model,
Convolutional Neural Network, Gradient Boosting, Learning Analytics, Early
Warning Systems. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
TFF_ATNET: A TRIPLET FOLDED FEATURES LEARNING FRAMEWORK WITH AN ATTENTION-BASED
TRANSFORMER NETWORK-BASED INTRUSION DETECTION SYSTEM |
|
Author: |
INDIRA, DR. A. K. SAMPATH |
|
Abstract: |
Intrusion Detection Systems (IDS) play a vital role in protecting networks from
malicious threats, yet they often face challenges such as imbalanced categories
of traffic and distributional discrepancies between training and testing data.
To overcome these limitations, this study introduces an integrated methodology
that enhances IDS performance and robustness. The framework begins with data
preprocessing and systematic data splitting through Apache Spark to handle
imbalance and ensure efficient processing. For feature learning, a Triplet
Feature Fusion module is employed, comprising Shallow Early-Stacked Feature
Pooling (SESF), Deep Learning-based Statistical Features (DLSF), and Dual
Statistical Features (DSF) to extract diverse and discriminative
representations. These fused features are then processed by the Triplet Feature
Fusion Attention Transformer Network (TFF_ATNet), where the Trans_Attention
classifier captures long-range dependencies for precise classification. To
further fine-tune the classification outcomes, the Self-Adaptive Walrus
Optimization Algorithm (SA-WaOA) is applied, ensuring optimized detection
performance. This model is implemented in Python and tested using three major
datasets, namely CIC-IDS2017, UNSW-NB15, and TON_IoT. The evaluation of
performance achieves an accuracy of 0.9993 on CIC-IDS2017 and low computation
time (98 ms), superior sensitivity (0.9909), and specificity (0.9769), compared
to existing methods. The proposed approach ensures balanced detection, improved
classification efficiency, and contributes significantly to the advancement of
IDS research in cybersecurity. |
|
Keywords: |
Intrusion Detection Systems, Feature Imbalance, Triplet Folded Feature Learning,
Attention-based Transformer Network, Cybersecurity Measures. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
HETEROGENEOUS ACCELERATION OF REAL-TIME OPERATING SYSTEM DIGITAL TWINS VIA
CPU-GPU CO-SIMULATION FRAMEWORK |
|
Author: |
DUC-THANG NGUYEN , DONG-LUONG DINH |
|
Abstract: |
Real-time operating system (RTOS) digital twins are essential validation
platforms for safety-critical cyber-physical systems, including autonomous
aerial vehicles and 6G radio units. However, as sensor counts scale toward tens
of thousands of channels, legacy single-node simulators suffer from significant
temporal drift, where host CPU overhead for virtual peripherals exceeds kernel
execution time. This paper introduces a Heterogeneous Co-simulation (HCS)
framework that addresses this by separating the discrete RTOS logic from the
continuous sensor manifold. The implementation of HCS executes the RTOS model on
the CPU and evaluates the sensor manifold through a common array-backend
abstraction that resolves to a single-instruction-multiple-thread capable
accelerator backend when one is available and otherwise to a vectorized CPU
backend. Both paths communicate through a shared-state module, avoiding explicit
device-to-host copy logic in the coordinator. Results from thousands of sensor
unmanned-aerial-vehicle case study show that HCS achieves an order-of-magnitude
speedup over a serial CPU-only baseline on the evaluated host. A statistical
evaluation on five controlled scenarios including sensor-count scalability,
tick-rate sensitivity, noise sensitivity, interrupt-threshold sweeps, and
long-run latency characterization shows that the HCS plane is strictly faster
than the CPU-only baseline in every configuration. The accompanying
implementation and testable component boundaries ensure reproducibility,
establishing HCS as a scalable alternative to legacy single-node simulators. In
summary, this paper contributes a two-plane execution model that places the RTOS
kernel in a closed simulation loop with an accelerator-resident sensor manifold
through a bit-packed virtual interrupt controller interface. In practice, the
framework enables engineers to validate interrupt-dense, safety-critical
firmware at sensor scales, thereby shortening certification campaigns and
lowering energy costs. |
|
Keywords: |
Real-time Operating Systems, Digital Twins, Heterogeneous Co-simulation, CPU-GPU
Partitioning, Parallel Discrete-event Simulation |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
A PRIVACY-PRESERVING FEDERATED CNN FOR DIABETIC RETINOPATHY DIAGNOSIS |
|
Author: |
R. SINDHUJHA , Dr. K. PADMA PRIYA |
|
Abstract: |
Diabetic Retinopathy is a common diabetic condition that, if left undressed, can
affect in blindness and visual impairment. In this work, we present a new system
for DR using a Federated Convolutional Neural Network (Fed CNN) that has been
trained using the Adam supervised classification optimization algorithm.
Decentralized model training on distributed datasets is made possible via
federated learning, which protects patient privacy and enhances the model's
conception. Multiple convolutional layers are used in Fed CNN armature to
facilitate hierarchical representation learning and feature extraction. The Adam
optimizer is used to adaptively alter the learning rates for specific
parameters, perfecting both the speed of confluence and the delicacy of the
model numerous healthcare installations work together to train the model, each
furnishing a different collection of retinal images for the model to use. The
proposed system is enforced using MATLAB software. By easing ongoing model
enhancement without taking raw data sharing, the federated learning architecture
allays privacy enterprises related to centralized styles. The proposed Fed CNN
with Adam optimization system shows better delicacy with 98% which 2.25% is
advanced when compared with FR- CNN, ANN and SVM. The issues of the trials show
how well the suggested approach may classify the inflexibility degrees of DR.
The federated CNN's improved individual capabilities and sequestration-
conserving features are stressed by comparisons with other styles. The suggested
frame advances telemedicine and customized healthcare by offering a scalable and
private result for the opinion of DR, a serious consequence of diabetes. The
primary research contribution of this study is the creation of a
privacy-preserving and highly accurate diabetic retinopathy diagnosis system
based on a Federated Convolutional Neural Network (Fed CNN) and the Adam
optimisation algorithm. The work describes a way by which several healthcare
institutions can train an AI model without sharing sensitive patient data,
enhancing both data privacy and model performance. |
|
Keywords: |
Adam Optimization, Convolutional Neural Network, Diabetic Retinopathy, Federated
Learning, Supervised Classification. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
SCOPE: STOCHASTIC CONSENSUS PURIFICATION WITH EXPLANATION CONSISTENCY FOR
ADVERSARIALLY ROBUST IMAGE CLASSIFICATION |
|
Author: |
HONEY DIANA P, Dr N. SUPRIYA |
|
Abstract: |
Adversarial perturbations are very effective at fooling deep neural image
classifiers: adding subtle noise to the image can lead to unreliable
predictions. Current defenses typically have three drawbacks: low clean
accuracy, low defense strength against high-level adaptive attacks, and a lack
of evidence of semantically meaningful visual regions recovered from recovered
predictions. In response to these challenges, we present a novel adversarial
defense framework called SCOPE: Stochastic Consensus Purification with
Explanation Consistency for robust and explainable image classification. To
overcome these challenges, we propose a novel adversarial defense framework,
called Stochastic Consensus Purification with Explanation Consistency (SCOPE),
for robust and explainable image classification. SCOPE combines stochastic image
purification, randomized consensus classification and explanation alignment via
Grad-CAM into a single optimization framework. The proposed approach is based on
a stochastic robustness objective which combines clean risk, adversarial risk,
purification error, prediction variance, margin loss, and explanation
inconsistency. Moreover, multiple stochastic forward passes are aggregated by a
consensus inference mechanism, to get more stable predictions, estimate
uncertainty and enable abstention for unstable inputs. The framework is tested
on publicly available datasets such as a balanced subset of 10,000 images from
the ImageNet dataset, 4,000 of which are used for purifier training, 1,000 for
validation and 5,000 for independent testing. Comprehensive evaluation is made
on FGSM, BIM, PGD-20, PGD-100, DeepFool, C&W, AutoAttack, Transfer Attack and
Adaptive BPDA+EOT. Experimental results demonstrate that SCOPE gains the clean
accuracy of 94.62%, the FGSM accuracy of 89.42%, the PGD-100 accuracy of 79.64%,
the AutoAttack accuracy of 75.48% and the BPDA+EOT accuracy of 71.36%,
outperforming PGD-AT, TRADES, SmoothAdv, DiffPure, RDC, and ADBM. Moreover,
SCOPE's interpretability is enhanced by its explanation stability score of
0.892, pointing-game accuracy of 82.75%, and its AUC for insertion and deletion
of 0.742 and 0.218 respectively. The findings in this work confirm the ability
of SCOPE to deliver high-quality trustworthy image classification with all the
above-mentioned properties. |
|
Keywords: |
Adversarial Robustness; Stochastic Purification; Explanation Consistency;
Grad-CAM; Imagenet; Autoattack; BPDA+EOT; Consensus Inference. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
RATF-NET: A RELIABILITY-AWARE TRI-BRANCH FUSION FRAMEWORK FOR
BOUNDARY-CALIBRATED BRAIN TUMOR SEGMENTATION IN PUBLIC MULTIMODAL MRI DATASETS |
|
Author: |
NEERUKATTU SIVA KUMAR , JOGESWARA RAO BAIPILLI |
|
Abstract: |
Correctly segmenting a brain tumour in multimodal magnetic resonance imaging
(MRI) is crucial for computer-assisted diagnosis, radiotherapy planning,
assessment of treatment response, follow-up evaluation, and radiogenomic
studies. Despite their success, current deep learning-based segmentation models
still suffer from several drawbacks, such as low robustness to incomplete MRI
scans, poor accuracy in detecting irregular tumor boundaries, poor uncertainty
calibration, and poor generalization to heterogeneous clinical imaging
conditions. Therefore, the main problem addressed in this study is not only
achieving high segmentation accuracy, but also producing reliable,
boundary-sensitive, and uncertainty-aware predictions under clinically realistic
conditions where MRI sequences may be incomplete or heterogeneous. This study
introduces a reliability-aware tri-branch fusion framework for multimodal
brain-tumor segmentation (RATF-Net) to solve these problems. The proposed
framework consists of a local 3D convolutional encoder for the spatial texture
representation, a global multilayer perceptron-based context encoder for the
long-range semantic modeling and a boundary-guided U-Net decoder for the
accurate tumor contour reconstruction. The branches are gated with a
modality-aware gating mechanism which assigns variable weight to the available
MRI sequences and enhances robustness in the absence of modalities. In addition,
a composite objective function is used to optimize RATF-Net, including the
Focal-Tversky overlap loss, the boundary distance regularization, the topology
consistency control, the modality dropout consistency, and the calibration-aware
uncertainty refinement.
The proposed model was assessed on publicly available
brain multimodal MR datasets with internal validation, external validation and
missing-modality stress test. As can be seen from the experimental results, the
performance of RATF-Net is promising, with the internal mean Dice score of 0.922
± 0.011, the external mean Dice score of 0.897 ± 0.015, and an expected
calibration error of 0.039 ± 0.006. The proposed framework outperformed the
representative baseline models in terms of segmentation accuracy, boundary
fidelity, uncertainty calibration and robustness in varying test scenarios. The
ablation results also showed that the three components, namely modality-aware
fusion, boundary-topology learning, and uncertainty-guided refinement, all made
significant contributions to the improvement of the overall performance.
Overall, the study concludes that integrating adaptive modality fusion,
boundary-topology learning, and uncertainty-guided refinement can improve both
the accuracy and reliability of multimodal brain-tumor segmentation. The results
suggest that RATF-Net achieves a comprehensive and solid deep learning tool for
multimodal brain-tumor segmentation from publicly available MRI datasets with
clinically relevant accuracy. |
|
Keywords: |
Segmentation Of Brain Tumors; Multimodal MRI Data; Calibration Of Uncertainty;
Boundary-Aware Learning; Handling Of Missing Modality; Deep Learning. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
CROWD EMOTION DETECTION FOR PUBLIC SECURITY |
|
Author: |
JAHNAVI SOMAVARAPU, N SANDEEP CHAITANYA,RAVIKANTH MOTUPALLI, K SAI SRIHITHA, K
RAVALIKA5, P GOUTHAM, P YASHASREE |
|
Abstract: |
With growing urban density and large-scale public gatherings, conventional video
surveillance which depends on human observers monitoring live feeds is
increasingly inadequate for detecting early emotional warning signs in crowds.
This paper introduces a crowd emotion detection system that integrates a dual
model for face detection using YOLOv8 and RetinaFace, a hybrid CNN-LSTM emotion
recognition network, and a Crowd Sentiment Aggregator (CSA) for generating
alerts. The experimental analysis conducted on the FER2013 dataset showed that
the CNN-LSTM architecture achieved a testing accuracy of 73.79%, which is higher
than the baseline ResNet-18 (49.16%) by 24.63%, and also better than other
existing models, such as DAN (55.97%) and EmotionNet (62.81%). |
|
Keywords: |
CNN-LSTM, Crowd Emotion Detection, Public Safety, RetinaFace, YOLOv8. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
EDGE COMPUTING FRAMEWORK FOR IOT PERFORMANCE OPTIMIZATION USING A
WAVENET–BILSTM–XGBOOST HYBRID ARCHITECTURE WITH AUTOMATED HYPERPARAMETER TUNING
WITH OPTUNA |
|
Author: |
Mr. UMESH KUMAR, Dr. PARUL VERMA, Dr. SYED QAMAR ABBAS |
|
Abstract: |
The present paper presents a state-of-the-art edge computing system that is
meant to support the solution of the major issues impacting IoT systems, such as
delaying data processing systems, network latency, and resource limitations. It
is a composite of three machine learning models: WaveNet, which does the
temporal pattern recognition, BiLSTM, which does sequential dependency modeling,
and the XGBoost, which does the refinement of the ensemble prediction. The
hybrid structure addresses the drawbacks of the traditional cloud-based
strategies because it handles data at the edge using automated hyperparameter
optimization that scales to the accessible computational resources. This allows
them to operate efficiently on a wide range of edge devices with different
capabilities using smart resource allocation and model compression. The
framework is tested on actual real-world datasets of smart cities, industrial
monitoring, and environmental networks and has shown significant improvements,
27 percent faster predictions, 34 percent higher forecasting accuracy, and 42
percent less resource usage than the current solutions. The system can process
high-frequency sensor data, irregular sampling and multi-dimensionality and
still have real-time processing. The automated tuning system minimizes manual
tuning by three-quarters and allows an automated response to dynamically
changing IoT conditions. The superiority of the framework over existing edge
computing solutions with respect to both performance and the practicality of
deployment is confirmed through a comparative analysis, which makes the
framework particularly useful to latency-sensitive IoT applications that need
immediate data processing on the network edge. |
|
Keywords: |
Edge Computing, IoT Optimization, WaveNet, BiLSTM, XGBoost, Hyperparameter
Tuning, Real-time Processing, Resource Allocation. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
DEEP VEHICLE RE-IDENTIFICATION USING VGG 16 AND SIAMESE NEURAL NETWORK: A ROBUST
APPROACH FOR VISUAL IDENTIFICATION OF VEHICLES |
|
Author: |
MARY SHAJI MATHEWS, P K NIZAR BANU |
|
Abstract: |
Vehicle re-identification is a challenging task in computer vision that aims to
identify the same vehicle across images captured by different cameras in a
surveillance system. Despite recent advancements, several challenges continue to
hinder the performance of vehicle re-identification systems, including limited
data availability, subtle intra-class and inter-class variations among vehicle
instances, the need to handle diverse input modalities, and re-identification
under extreme weather conditions. To address these issues, this paper proposes a
vehicle re-identification framework based on VGG-16 and a Siamese Neural
Network. The VGG-16 model is utilized to extract discriminative features from
vehicle images by focusing on key regions such as the vehicle body and license
plate. These extracted features are subsequently fed into a Siamese Neural
Network, which learns a similarity metric between pairs of vehicle images while
effectively accounting for variations in viewpoint, illumination, and occlusion.
In addition, a triplet loss function is incorporated to improve feature
discrimination and effectively handle difficult image pairs. The proposed
network is trained on a large-scale vehicle image dataset with similarity
labels, enabling accurate similarity measurement between vehicles captured under
diverse conditions. Experimental evaluation on the VeRi-776 dataset achieves a
mean average precision (mAP) of 0.8853, demonstrating the effectiveness and
robustness of the proposed approach for vehicle re-identification tasks. |
|
Keywords: |
Vehicle Re-Identification, Image Classification, VGG16 Neural Network, Siamese
Neural Network |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
HYBRIDATTENTIONNET: A GRAPH CONVOLUTIONAL-TRANSFORMER HYBRID ARCHITECTURE WITH
MULTI-HEAD ATTENTION FOR INTELLIGENT REAL-TIME NETWORK INTRUSION DETECTION |
|
Author: |
DR. RAM PRASAD REDDY SADI, DR.CH.V.SIVARAM PRASAD, DR. CH RAMESH BABU, DR.
CHANDRA SEKHAR KOPPIREDDY, DR. K NAGARAJU, DR. SAMUEL SUSAN VEERAVALLI,
M.KEERTHI PRIYA, Dr. HARI JYOTHULA |
|
Abstract: |
Network intrusion detection systems face a critical dual challenge: the
exponential growth of encrypted and polymorphic attack traffic renders
signature-based defenses obsolete, while existing deep learning models fail to
jointly capture the spatial topology of co-occurring network flows and their
long-range temporal evolution. This paper addresses this concern by proposing
HybridAttentionNet (HAN), a unified Graph Convolutional-Transformer architecture
that simultaneously models structural inter-flow relationships and sequential
temporal dependencies through a dual residual fusion mechanism. The key
contribution of this work is a principled, end-to-end trainable framework that,
for the first time, integrates dynamic attributed graph construction, two-layer
GCN encoding, 12-head Transformer self-attention, and SHAP-based
interpretability into a single intrusion detection pipeline. Evaluated on the
NSL-KDD benchmark, HAN achieves 98.4% classification accuracy and a macro
F1-score of 97.6%, outperforming all compared baselines by up to 7.2 percentage
points. The practical impact is a deployable, interpretable detection model with
22.3 ms inference latency suitable for real-time Security Operations Centre
environments, directly reducing analyst response time and false-positive burden
in production networks. |
|
Keywords: |
Network Intrusion Detection; Graph Convolutional Network; Transformer;
Multi-Head Attention; Hybrid Deep Learning; Cybersecurity; NSL-KDD; SHAP
Explainability |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
HYBRID QUANTUM-CLASSICAL OPTIMIZATION: COMBINING PSO AND GA WITH QUANTUM
ALGORITHMS FOR BETTER DECISION-MAKING |
|
Author: |
K. YASUDHA, MUKTEVI SRIVENKATESH |
|
Abstract: |
Quantum computing is a big change in how to do calculations, especially good for
solving complex problems that are hard for traditional computers. These problems
often get worse quickly as they grow bigger, making them very difficult to solve
efficiently. Although quantum computers have some big advantages, they still
have some issues like limited number of qubits, noise, and problems with staying
stable for long periods. To deal with these problems and make the best use of
both quantum and classical methods, this research proposes a new system that
mixes classical methods such as Particle Swarm Optimization (PSO) and Genetic
Algorithms (GA) with quantum methods like Quantum Approximate Optimization
Algorithm (QAOA) and different versions of Grover's search. [14] [15] [1] [5]
The classical methods are good at finding good solutions and working reliably,
while the quantum parts can help in certain tasks, like searching or improving
calculations. The new system is tested on real problems in areas like logistics
(like planning delivery routes), network routing (like finding the fastest
paths), and job scheduling (like assigning tasks with limited resources). A way
is developed to turn these problems into a form that quantum computers can
understand, called QUBO. [10]
The tests use real data and simulated quantum
setups (like Qiskit and D-Wave Ocean SDK). The results show that this new system
works better than using only classical or only quantum methods in terms of the
quality of the solutions, how fast they find the answers, and how long they
take. This study shows the potential of using a mix of quantum and classical
methods to solve problems more effectively and sets up a foundation for future
research in using quantum methods to help make better decisions. [12] [18] |
|
Keywords: |
Quantum Computing, Hybrid Optimization, QAOA, Grover's Algorithm, Particle Swarm
Optimization, Genetic Algorithm, Scheduling, Logistics, Network Routing. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
ENHANCING NETWORK SECURITY WITH CONVOLUTIONAL NEURAL NETWORKS: AN INTRUSION
DETECTION MODEL USING THE NSL-KDD DATASET |
|
Author: |
AMMAR D. JASIM, SAIF S. KAREEM |
|
Abstract: |
Intrusion Detection Systems (IDS) are very significant in the protection of
network infrastructure systems against unauthorized malicious attacks and new
threats. In the given work, the Deep Neural Network (DNN)-based IDS is suggested
to be based on Convolutional Neural Networks (CNNs), and this suggestion is
tested on the NSL- Knowledge Discovery in Databases (KDD) benchmark dataset. The
proposed model, in contrast to the traditional machine learning techniques using
manual feature engineering, provides automatic feature extraction of a
hierarchical structure which is based on the time-frequency representation of
the network traffic, facilitating detection. To solve the problem of class
imbalance, weighted loss functions are used, and categories are represented by
categorical encoding, and feature standardization is used to learns the models.
The model had an outstanding performance of 99.87% accuracy and AUC 0.9265 and
which was quite higher compared to the baseline 5-layer Auto encoder and other
traditional methods. The experimental findings indicate the validity of the
model because it can differentiate normal and attack traffic with very low false
positives and false negatives. As a regularization method, a regular way to do
early stopping is to avoid overfitting and to generalize. Although the proposed
system proves to be very effective on benchmark data, a few issues are still
related to high computational requirements and real-time usability. Future work
should focus on optimizing the model for deployment, improving zero-day attack
resilience and ensuring robust testing of the model performance under variable
and realistic network conditions. This collection of results adds to the
combinability of CNN-based models and the most advanced IDS systems for more
proactive and intelligent cyber security. |
|
Keywords: |
Intrusion Detection Systems, Convolutional Neural Network, Deep Neural Network. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
FUZZY ADAPTIVE DENSITY WEIGHTED K MEANS FRAMEWORK FOR RECRUITMENT CANDIDATE
CLASSIFICATION USING DENSITY BASED CLUSTERING AND WEIGHTED SIMILARITY ANALYSIS |
|
Author: |
PETRICIA LEEMA ROSELINE, ARUNA P |
|
Abstract: |
Modern recruitment environments generate large volumes of candidate data that
require structured analytical mechanisms for effective competency evaluation and
decision support. Conventional recruitment analytics approaches mainly emphasize
recommendation systems or transparency-oriented analysis, offering limited
capability for discovering competency structures within candidate datasets.
Effective recruitment intelligence requires analytical frameworks capable of
identifying meaningful candidate groupings while preserving variations in skill
composition and experience distribution. This study introduces a Fuzzy-Adaptive
Density-Weighted K-Means framework designed for competency-oriented clustering
within recruitment data. The framework integrates adaptive density estimation,
weighted similarity assessment, and fuzzy membership assignment to capture
heterogeneous relationships among candidate attributes. Density-aware weighting
strengthens recognition of candidate distribution patterns, while fuzzy
membership modelling enables flexible representation of overlapping competency
profiles. Iterative centroid refinement further supports stable cluster
formation and improved structural consistency. Experimental evaluation conducted
on a recruitment dataset containing ten thousand records demonstrates strong
classification stability and reliable clustering behaviour. Analytical outcomes
highlight the capability of the proposed framework to support structured
candidate profiling and intelligent recruitment analytics within data-driven
human resource environments. |
|
Keywords: |
Fuzzy Clustering, Recruitment Analytics, Density-Based Clustering, Candidate
Competency Analysis, Weighted Similarity Measurement, Intelligent Recruitment
Systems. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
VGG19-FTKDDD: A DEEP TRANSFER LEARNING FRAMEWORK FOR MULTI-CLASS DRIVER
DROWSINESS DETECTION AND CLASSIFICATION |
|
Author: |
V.S. KRISHNA PRASAD , Dr. S. SARAVANAN |
|
Abstract: |
Driver drowsiness is a major contributor to road accidents worldwide, accounting
for thousands of injuries and fatalities each year. Early detection of fatigue
is therefore essential to prevent collisions and enhance road safety. This
research presents an advanced deep learning-based framework for driver
drowsiness detection and classification using a VGG-19 architecture. The
proposed model operates through four key stages: Image annotation,
pre-processing, feature extraction, and classification. Annotated driver images
were organized into three behavioral categories – Drowsy, Awake, and Yawning.
Preprocessing techniques such as resizing, normalization, and histogram
equalization were applied to enhance feature visibility and ensure input
consistency. The pre-trained VGG-19 model, fine-tuned on the dataset, was
employed for deep feature extraction, followed by customized fully connected
layers and a softmax output layer for final prediction. Hyperparameter
optimization was performed using the Keras Tuner framework, enabling the
fine-tuning of learning rates, batch sizes, and optimizer parameters to achieve
superior accuracy and reduced overfitting. Experimental results demonstrate the
model’s effectiveness in accurately identifying driver states under varying
lighting and pose conditions with 95.00% of accuracy. This study contributes to
the advancement of intelligent driver monitoring systems, providing a reliable
and automated deep learning based solution to mitigate fatigue-related road
accidents. |
|
Keywords: |
Driver Drowsiness Detection, Deep Learning, VGG-19, Keras Tuner, Computer
Vision, Real-Time Fatigue Detection. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
OPTIMIZING CONTENT MARKETING STRATEGIES BASED ON SEARCH ENGINE ALGORITHMS AND
ARTIFICIAL INTELLIGENCE TO INCREASE SEMANTIC RELEVANCE AND SEO EFFECTIVENESS |
|
Author: |
OLHA KLIMOVYCH, ALINA DANILEVIČA, MARIANA ABRAHAMYAN, LUDMILA LARKA, MARYNA
KHAZHEEVA |
|
Abstract: |
The integration of generative AI into search engine algorithms presents
fundamental challenges for contemporary content marketing strategies. This study
aims to quantify the impact of artificial intelligence on SEO performance and to
develop a composite index system for optimizing content marketing in AI-driven
search environments. This study employs content analysis of secondary data from
leading industry platforms. In addition, the current research identifies
interdependencies among key performance metrics and applies multifactor
aggregation using min–max normalization. Predictive modeling is used to assess
AI adoption trajectories, accounting for the volatility associated with the
rollout of AI Overviews. The analysis includes an interregional comparative
assessment of the United States and the European Union, incorporating
institutional differences in antitrust regulation. The empirical dataset
comprises more than 300,000 keywords, over 10 million queries, and large-scale
clickstream data from Semrush and Ahrefs spanning 2022–2025. The AI Search
Transformation Index (ASTI = 0.547) indicates a moderate level of adaptation of
the search ecosystem to generative AI. The PEAIO visibility loss ratio (34.74%)
captures a substantial decline in organic click-through rates for top-10
positions attributable to AI Overviews. In the US, the Traffic Redistribution
Index (TRI) declined by 15.3% reflecting increased click migration toward
Google-owned properties and a rise in zero-click searches to 58.5%. An AI
citation anomaly was identified, whereby pages ranking in positions 21–50
exhibit a 64.7% higher likelihood of being cited in AI-generated responses than
those in the top five results. The composite Global Score of 56.5 suggests
moderate market readiness for AI-dominated search, alongside relatively high
elasticity in retaining organic traffic. The proposed methodology enables
systematic diagnostics of SEO vulnerabilities, cross-regional benchmarking of
content strategies as well as the development of data-driven optimization
frameworks. Strategic implications include adopting hybrid positioning
strategies that balance traditional CTR optimization with AI citation
visibility. Thereby it entails increasing semantic content density and
diversifying traffic acquisition through platforms with higher AI citation
propensity. |
|
Keywords: |
AI SEO Transformation, Semantic Relevance, AI Overviews, Zero-Click Searches,
Organic Visibility, Content Marketing, Composite Indexes. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
AT-HGNN: ADAPTIVE TEMPORAL HYPERGRAPH NEURAL NETWORK FOR REAL-TIME NETWORK
INTRUSION DETECTION IN HETEROGENEOUS IOT ENVIRONMENTS |
|
Author: |
SATISHKUMAR PATNALA , DR. CH RAMESH BABU , Y LAXMANA RAO , DR.JALAIAH SAIKAM ,
GANDHIKOTA UMAMAHESH , V VIJAYAKUMAR DASARI , DR.V.VAITHEESHWARAN , DR. HARI
JYOTHULA |
|
Abstract: |
The growing number of Internet of Things (IoT) devices greatly widened the
attack surface of modern network architectures, leading to traditional
signature-based and shallow machine learning intrusion detection systems (IDS)
that generally use lightweight tries being not robust at all against advanced,
polymorphic and zero-day cyber attacks. In this paper we introduce AT-HGNN, an
Adaptive Temporal Hypergraph Neural Network — a state-of-the-art deep learning
framework that progressively harnesses newly invented hypergraph-based
relational modeling, multi-head cross-feature attention schemes and temporal
graph attention network (T-GAT) to extract higher order nonpairwise
relationships on different time granularity level from the in-depth flow
association information. If compared to classical graph neural network that only
consider pairwise edges, the hypergraph formulation allows AT-HGNN to capture
complex multi-flow correlated attack patterns including distributed
denial-of-service (DDoS) campaigns and coordinated reconnaissance probes. To
make the model more robust to adversarial flow perturbation during training, a
new adaptive edge-weight updater is employed to establish inter-node
connectivity adapted by evolving traffic semantics.Comprehensive experiments on
four benchmark datasets — NSL-KDD, UNSW-NB15, CIC-IDS-2017, and the novel
IoT-NID-2024 dataset — show that AT-HGNN achieves state-of-the-art performance
with overall accuracy of 99.1%, F1-score of 98.7%, AUC−ROC (Area Under Curve −
Receiver Operating Characteristic) score of 0.9987, and false positive rate of
0.21% outperforming seven competing baselines by statistically significant
margins including CNN-LSTM, GraphSAGE and standard GNNs.) A large ablation study
verifies that all design elements are critical and, thus, indispensable. The
proposed model also shows remarkable scalability, processing one million flow
records using commodity GPU hardware in less than 18 minutes. These findings
solidify AT-HGNN as an efficient and deployable, state-of-the-art accurate
architecture for next-generation real-time network security monitoring in
diverse IoT scenarios. |
|
Keywords: |
Hypergraph Neural Network, Network Intrusion Detection, IoT Security, Temporal
Graph Attention, Deep Learning, Cybersecurity, Graph Neural Networks |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
HYBRID CNN TRANSFORMER MODELS FOR LUNG DISEASE DETECTION FROM CT SCANS |
|
Author: |
ARUNA VIPPARLA, DR. LAKSHMI NAGA JAYAPRADA GAVARRAJU, DR. B. LEELAVATHY, DR.
JOHNWESILY CHAPPIDI, CHALLAPALLI SUJANA, KUMAR DEVAPOGU, AKKALA YUGANDHARA
REDDY, K. PRAVEEN KUMAR |
|
Abstract: |
Timely and accurate diagnosis is significant for enhancing patient survival and
reducing the strain that health care facilities would otherwise face in
providing diagnostic services. Computed Tomography (CT) imaging is a popular
imaging method that allows visualization of lung structure in detail, revealing
abnormalities, including lung cancer, pneumonia, and Coronavirus (COVID-19)
infection. Nevertheless, manual interpretation of CT scans is labor-intensive
and susceptible to discrepancies among radiologists. The work presents an
effective automated system for lung disease detection using a hybrid
CNN-Transformer architecture that utilizes local spatial attributes and
extensive contextual comprehension. The convolutional neural networks (CNNs) in
this paper formulate a model that extracts multi-level features, after which a
Transformer encoder is used to model long-range dependencies across lung
regions. Moreover, a new multi-scale attention fusion framework is proposed to
combine the spatial and contextual representations to enhance the feature
learning. The experiment was conducted on a multi-class lung CT dataset that
includes normal cases, lung cancer, pneumonia, and COVID-19. The proposed model
achieved a classification accuracy of 96.3%, a sensitivity of 95.8%, a precision
of 95.9%, and an ROC-AUC of 0.97, surpassing other conventional CNNs, including
VGG16, ResNet50, and DenseNet121, as well as single Vision Transformer models.
The results show that the hybrid CNN-Transformer architecture outperforms both
CNN and Transformer architectures in feature encoding and lung disease
diagnosis. The proposed framework constitutes a reliable computer-aided
diagnostic system that can assist radiologists in the initial identification of
pulmonary diseases and has great potential for use in AI-assisted clinical
decision support systems. |
|
Keywords: |
Lung Disease Detection, Computed Tomography (CT), Hybrid CNN–Transformer,
Multi-scale Attention Fusion, Deep Learning, Medical Image Classification |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
MCMC-GUIDED STABLE DIFFUSION FRAMEWORK FOR IMPROVED TEXT-TO-IMAGE GENERATION
WITH ENHANCED SEMANTIC ALIGNMENT |
|
Author: |
DR RASWITHA BANDI , B RAVALI REDDY ,KAMBHAM PRATAP JOSHI , M VANIAH SUCHARITHA
SANTOSH ,SUCHITRA PATTABIRAMAN ,DR J VAMSINATH |
|
Abstract: |
Text-to-image generation has gained significant attention due to its ability to
synthesize realistic images from natural language descriptions. Despite recent
advancements in diffusion-based models, challenges such as inefficient sampling,
poor semantic alignment, and lack of consistency remain unresolved. This paper
proposes an enhanced Stable Diffusion framework integrated with Markov Chain
Monte Carlo (MCMC) sampling to improve image quality and text-image alignment.
The proposed approach leverages CLIP-based scoring to guide the sampling
process, ensuring that generated images better correspond to the given textual
prompts. Additionally, optimization strategies such as classifier-free guidance
and LoRA-based fine-tuning are incorporated to further enhance performance.
Experimental results on benchmark datasets demonstrate that the proposed method
achieves improved CLIP scores, higher image sharpness, and better semantic
consistency compared to baseline diffusion and GAN-based approaches. The
findings indicate that integrating probabilistic sampling techniques
significantly enhances the effectiveness of text-to-image generation models. |
|
Keywords: |
Stable Diffusion, MCMC Sampling, Text-to-Image Generation, CLIP, Image Synthesis |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
FEATURE EXTRACTION USING GLOVE AND FASTTEXT FOR SENTIMENT ANALYSIS ON AUGMENTED
TEXT DATA |
|
Author: |
MOHAMMAD HASBI ABIYADI , AMALIA ZAHRA |
|
Abstract: |
In this modern era, technological development and globalization have completely
changed human habits towards the digital world. leading to a significant
increase in user activity on social media platforms such as Instagram, Facebook,
and Twitter, which in turn generates large volumes of data known as big
data.This research aims to improve the accuracy of sentiment analysis on
Indonesian text data using the GloVe and FastText methods as effective feature
extraction techniques in text processing. The research framework systematically
describes the research process, from problem identification to research
objectives. The identified problem is the low accuracy of sentiment analysis,
especially for rare words. The literature review explores various sentiment
analysis methods, word embeddings, and previous related research. The research
hypotheses are formulated based on the literature review, stating that GloVe and
FastText methods produce different accuracy in sentiment analysis on Indonesian
text data and that FastText yields higher accuracy compared to GloVe. Data
collection involves gathering Indonesian text data with sentiment labels from
sources such as social media, online forums, and product reviews. The research
provides a clear and detailed guide on the effective use of GloVe and FastText
methods in sentiment analysis, assisting researchers and practitioners in
decision-making when choosing methods |
|
Keywords: |
Sentiment Analysis, Feature Extraction, GloVe, FastText, Data Augmentation, Word
Embedding, Indonesian Text, IndoBERT, Machine Learning. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
A FEDERATED DEEP REINFORCEMENT LEARNING FRAMEWORK FOR PRIVACY-PRESERVING
REAL-TIME ANOMALY DETECTION IN SMART GRID IoT SENSOR NETWORKS |
|
Author: |
Dr. SHOBANA GORINTLA , BODIGIRI SAI GOPINADH , CHOPPARAPU SRINIVASA RAO ,
ANTHARAJU K CHAKRAVARTHY, Dr. MAREPALLI RADHA, Dr. K CHINNAIAH6, Dr. S.
BANUMATHI, VALETI NAGARJUNA |
|
Abstract: |
The recent trend of Internet of Things (IoT) sensor networks to facilitate smart
grids has presented serious challenges in detecting anomalies, ensuring data
privacy and enabling real-time decision-making. This paper presents a novel
framework for Federated Deep Reinforcement Learning (FDRL) to develop an
effective, privacy-conscious anomaly detection system for distributed smart grid
systems. The given approach applies federated learning along with deep
reinforcement learning to provide decentralized model training and the evolution
of attack patterns without access to raw data. A hybrid reward functional is
developed to maximize recognition accuracy, energy consumption and latency. The
results of the experiments prove that the proposed model achieved an accuracy of
97.2%, which is higher than those of traditional machine learning, deep
learning, and federated learning methods. Moreover, communication overhead and
latency can also be minimized by 38% and 27%, respectively. The findings provide
an informative background on the system's strength, scalability, and real-time
nature for addressing non-IID IoT data. The findings of the current piece of
work may be summarized as the following: the suggested FDRL method is a viable
and scalable way of securing next-generation smart grid infrastructure with
significant implications on privacy-sensitive smart energy systems. |
|
Keywords: |
Federated Learning, Deep Reinforcement Learning, Smart Grid, Anomaly Detection,
IoT Security, Edge Computing |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
CROSS-REGIONAL EARTHQUAKE-INDUCED SOIL LIQUEFACTION ASSESSMENT USING A
MECHANISM-INFORMED EXPLAINABLE AI-DRIVEN DEEP LEARNING FRAMEWORK |
|
Author: |
U.V NARAYANA RAO, TAKKELLAPATI. SUJATHA, PRAVEENA BAI DESAVATHU, DANAM NOELLE, P
SHABANA5, DR RATHNA JYOTHI CHADUVULA, KARNATAPU LEELA KRISHNA, PONDURI SIDDARDHA |
|
Abstract: |
Accurate assessment of earthquake-induced soil liquefaction is challenging due
to nonlinear interactions among seismic loading, soil properties, and
groundwater conditions. This study proposes a physics-informed and explainable
hybrid deep learning framework for reliable liquefaction prediction across
multiple seismic regions. Classical geotechnical parameters, including cyclic
stress ratio (CSR), cyclic resistance ratio (CRR), and factor of safety (FS),
are embedded into a constrained learning formulation to ensure physical
consistency. A 1D-CNN–BiLSTM–attention architecture captures localized and
depth-wise dependencies, while SHAP and sensitivity analyses enhance
interpretability. Cross-regional validation demonstrates strong generalization
capability and improved predictive performance compared to conventional machine
learning approaches. |
|
Keywords: |
Cyclic Stress Ratio (CSR) ,Cyclic Resistance Ratio (CRR), Factor Of Safety (FS),
Soil Liquefaction, Classical Geotechnical Parameters, Cross-Regional Validation. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
ADVANCEMENTS IN VISUAL QUESTION ANSWERING METHODOLOGIES: INCORPORATING LSTM AND
PRE-TRAINED CNN FEATURES |
|
Author: |
Y HARIKA DEVI , DR N CHAITANYA KUMAR |
|
Abstract: |
Visual Question Answering (VQA) remains a challenging task due to the difficulty
of accurately modeling detailed image content, object relationships, and
contextual information while simultaneously understanding natural language
queries. Existing VQA approaches often struggle to capture fine-grained visual
features and maintain contextual relevance in generated answers. To address
these limitations, this study proposes an enhanced VQA framework that integrates
Long Short-Term Memory (LSTM) networks for natural language processing with
Convolutional Neural Networks (CNNs) pre-trained on ImageNet for robust image
feature extraction. The proposed framework combines textual and visual
representations through an effective feature fusion mechanism to improve
contextual understanding and answer accuracy. Experimental evaluation conducted
on a dataset containing 10,000 images and 50,000 question-image pairs
demonstrates stable model convergence, balanced performance across diverse
question types, and strong contextual understanding across multiple image
categories. Comparative analysis further shows that the proposed model
outperforms baseline VQA approaches that utilize only LSTM or CNN-based
representations. Additionally, the model exhibits effective generalization on
unseen test data, confirming its robustness and practical applicability. The
results indicate that the integration of LSTM-based language understanding with
pre-trained CNN visual representations significantly enhances VQA performance,
providing a reliable and context-aware solution for answering questions related
to visual content. |
|
Keywords: |
Visual Question Answering, LSTM, Pre-trained CNN Features, Natural
Language Processing, ImageNet, Contextual Understanding. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
HIERARCHICAL EXPLAINABLE DEEP NEURAL ARCHITECTURE FOR REAL-TIME INTRUSION
DETECTION IN ENGLISH EDUCATION ECOSYSTEMS |
|
Author: |
DR S. REMA DEVI, DR. KHURSHEEDA KHATOON, VAMSIDHAR TALASILA, VIMOCHANA.M, DR.
PUNIT PATHAK, DR. R. BALAKRISHNA, DR.M.SHYAMALA BHARATHY |
|
Abstract: |
Some of the online English education systems that have become highly vulnerable
to advanced cyber attacks threatening to compromise personal academic and
administrative information include learning management systems (LMS), virtual
classrooms, student portals, and institutional networks just to mention a few.
Due to class imbalance, poor feature representation, and lack of
interpretability, the traditional intrusion detection systems (IDS) that use
shallow machine learning or the standard deep learning models often cannot
detect the unusual types of attacks, which reduces the trust in automated
security decisions. The proposed solution to these issues in this work is a
Hierarchical Explainable Deep Neural Architecture (HEDNA) of real-time intrusion
detection in an English teaching environment. The proposed system integrates
explainable AI methods, explainable AI methods, oversampling and cost-sensitive
learning to reduce class imbalance and multi-level hierarchical feature
engineering to generate intelligible security decisions. Network traffic
characteristics such as flow duration, number of packets, rate of the bytes and
indicators of session-level behavior are used to identify low-level and
contextual trends by hierarchically aggregating the characteristics. The model
was tested using the TII-SSRC-23 dataset that contains realistic benign and
malicious network flows and was implemented in Python. It is a suitable
architecture to be deployed in real-time since the experimental results indicate
that the proposed architecture can achieve an accuracy of 96% in terms of
intrusion detection with low false-positive rates and latency. Also,
explainability modules provide useful information to the administrators by
highlighting significant contributing traffic characteristics. The results
confirm that explainable AI and hierarchical deep learning have a significant
impact on enhancing accuracy, transparency, and operation feasibility in the
protection of modern English education ecosystems. |
|
Keywords: |
Hierarchical Deep Neural Network, Explainable AI, Real-Time Intrusion Detection,
English Education Ecosystems, TII-SSRC-23 Dataset |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
LEGAL UNDERPINNINGS OF THE USING DIGITAL EVIDENCE IN CRIMINAL PROCEEDINGS IN THE
CONTEXT OF ADMINISTRATIVE-DIGITAL TRANSFORMATION |
|
Author: |
VLADYSLAV VEKLYCH, VITALII ANDRUKH, MYROSLAV POPOVYCH, OLEKSANDR KVASHUK, ANDRIY
TYMCHYSHYN |
|
Abstract: |
To ensure the reliability of digital evidence in criminal proceedings and,
consequently, its admissibility in court, it is imperative to apply both
technical and legal mechanisms. Such an integrated approach enables effective
control over the origin, authenticity and integrity of digital information and
reduces the risk of judicial challenges. The aim of this study is to conduct a
comprehensive assessment of the reliability of digital evidence used in criminal
cases within the context of administrative and digital transformation. Other
than that, the study proposes the development of an integrated framework known
as the Digital Evidence Reliability Framework (DERF). The research methodology
combines comparative legal, technical, analytical and system-based approaches,
supported by quantitative analysis of four interrelated components: technical
integrity (TI), procedural correctness (PC), access control (AC) and regulatory
compliance (RC). The study draws on criminal procedural law and legislation from
Ukraine, Spain and the United Arab Emirates. The findings reveal significant
differences in the balance between technical and legal mechanisms governing
digital evidence across the examined jurisdictions. Based on the average values
of the General Reliability Index (DERF = 0.66), Ukraine demonstrates a moderate
level of overall digital evidence reliability in criminal proceedings. This
outcome reflects a discrepancy between the technological robustness of digital
data and the procedural mechanism required to legally validate such reliability.
Thus, Spain (DERF = 0.78) represents a more balanced model, combining
established technical safeguards for digital information with well-defined
procedural requirements and complementary regulatory mechanisms, including data
protection standards. The highest overall reliability score was recorded in the
United Arab Emirates (DERF = 0.87), where technical, organizational and legal
mechanisms operate in a balanced and cohesive way. Analysis of structural
imbalances indicates that the most significant losses in digital evidence
reliability occur during the translation of technical actions into formalized
regulatory and procedural frameworks. The scientific novelty of this study lies
in the development and empirical validation of the integrated DERF model, which
shifts the evaluation of digital evidence from predominantly descriptive
approaches to a quantitative assessment of legal system readiness. From a
practical perspective, the proposed framework offers a diagnostic tool for
identifying critical risk areas and formulating targeted recommendations to
enhance legislation and improve the handling of digital evidence across
different legal systems. The revised manuscript makes the operational logic of
the DERF model explicit by linking each score to defined legal, procedural,
organizational and technical criteria, and by presenting DERF as a reproducible
conceptual and quantitative framework rather than as a descriptive comparison
only. |
|
Keywords: |
Criminal Justice, Innovation, Legal Administration, Digital Evidence, Criminal
Proceedings, Technical Integrity, Procedural Admissibility, Integrated Model,
Administrative-Digital Transformation. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
A HYBRID VARIATIONAL QUANTUM–DEEP LEARNING FRAMEWORK FOR ELECTRICAL LOAD
FORECASTING IN SMART GRIDS |
|
Author: |
RAVURI DANIEL, BODE PRASAD, PASAM PRUDHVI KIRAN, KANDRAKUNTA CHINNAIAH,
INDHUMATHI RAVICHANDRAN, RAMANJANEYULU DAYINABOYINA, MANIDHEER BABU GORIKAPUDI,
DORABABU SUDARSA |
|
Abstract: |
In the smart grid, electrical load forecasting plays an important role in its
operation, especially as renewable energy sources are increasingly integrated
into the grid, increasing demand variability. Recurrent neural networks (RNNs)
and Long Short-Term Memory (LSTM) networks can learn temporal patterns but are
often ineffective at handling long-range dependencies, complex feature
relationships, and fluctuations caused by renewable energy sources. In this
study, a hybrid Variational Quantum–Deep Learning model for short-term
electrical load forecasting is proposed. To enhance feature representation,
sequence learning, and time-step weighting, the model incorporates a Variational
Quantum Circuit (VQC), LSTM layers, and an attention mechanism. Temporal
forecasting is preceded by preprocessing and encoding of historical load data
and exogenous variables, such as weather, calendar, and renewable-energy
indicators. The proposed model is compared with the baseline models, such as
Multilayer Perceptron (MLP), RNN, LSTM, Convolutional Neural Network–LSTM
(CNN-LSTM), and the coefficient of determination (R²) using Root Mean Square
Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error
(MAPE). The forecasting errors of the proposed Variational Quantum
Algorithm–LSTM–Attention (VQA–LSTM–Attention) model are found to be the
smallest, with RMSE and MAPE 22.64% and 25.64% lower than those of the
standalone LSTM model, respectively. These results show that the accuracy of
short-term load forecasting in uncertain smart grids can be enhanced by
combining quantum-enhanced feature transformation with LSTM and attention-based
learning. The paper proposes an applicable hybrid quantum–classical forecasting
model for demand response, renewable integration, market bidding, and grid
stability. |
|
Keywords: |
Electrical Load Forecasting; Variational Quantum Circuit (VQC); Long Short-Term
Memory (LSTM); Quantum Machine Learning; Smart Grids. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
AN INTERPRETABLE TYPE-1 FUZZY SYSTEM FOR DATA-DRIVEN OCCUPATIONAL STRESS
PREDICTION |
|
Author: |
KHALID ALMOHAMMADI |
|
Abstract: |
Job stress affects employee well-being, organisational performance and
psychological and physical health, leading to significant organisational costs.
Nevertheless, evaluating it remains a challenge because it differs according to
employee characteristics and the complex and correlated nature of other
workplace factors. Traditional methods for assessing occupational stress are
typically based on the guidelines and opinions provided by specialists. Thus,
they fail to capture differences in stress perceptions between specialists and
employees and do not use accumulated data to foster learning and
interpretability. This study presents a novel type-1 fuzzy logic-based system
for predicting occupational stress by correlating the impact of five
inputs—workload, working hours, job satisfaction, managerial support and job
security—on employees’ actual perceived stress levels. The proposed system
incorporates self-learning capabilities from data, enabling it to predict
employee stress levels. The interpretability of the extracted rules enhanced
specialists’ understanding of the impact of workplace environmental
characteristics within a lifelong learning framework. Various preliminary
experiments were conducted on 31 employees at Tabuk University. Data on employee
characteristics and corresponding stress levels were collected and used as
primary inputs and outputs for developing transparent and interpretable type-1
fuzzy logic models. The results indicated that the proposed system could predict
employee stress levels with high accuracy, exhibiting a low average error and
standard deviation. Moreover, it could effectively address the uncertainty and
complexity inherent to employee characteristics and work environment stress. The
proposed system enabled specialists to understand and predict work-related
stress and gain insights into its contributing factors and the relationships
among them. |
|
Keywords: |
Type-1 Fuzzy Logic, Fuzzy Rule-Based System, Occupational Stress Prediction,
Workplace Stress Assessment, Interpretable Machine Learning |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
DYNAMIC FUSION OF CNN AND TRANSFORMER REPRESENTATIONS FOR ROBUST
PARKINSONIAN GAIT CLASSIFICATION |
|
Author: |
Dr G. SAVITHA, S. GAYATHRI, Dr.P.VIMAL KUMAR, DR.N.HANUMANTHARAO, M.RAMA
KRISHNAN, Dr.M.GOMATHY NAYAGAM, DR.S.S.ANANTHAN, R.VANIDHASRI, DR.K.SELVAM,
DR.T.VENGATESH |
|
Abstract: |
Parkinsons Disease (PD) diagnosis via gait analysis is a critical yet
challenging task. While deep learning models like Convolutional Neural Networks
(CNNs) and Recurrent Neural Networks (RNNs) have been applied, they often fail
to simultaneously capture the intricate spatial features and long-range temporal
dependencies inherent in gait data [3, 4, 5]. To overcome this, we propose a
novel Hybrid CNN-Transformer model with Dynamic Feature Fusion (CT-DFF). Our
architecture synergistically combines a CNN for local spatial feature extraction
and a Transformer for global temporal context modeling. The core innovation is a
Dynamic Feature Fusion (DFF) module that adaptively weights and fuses
multi-scale features from both components. Evaluated on a public PD gait
dataset, our model achieves a state-of-the-art accuracy of 96.8% and an F1-score
of 95.4%, significantly outperforming standalone CNN, LSTM, and Transformer
models. The results demonstrate the model's robust capability to capture complex
gait dynamics, offering a powerful tool for enhancing clinical PD diagnosis. |
|
Keywords: |
Parkinsons Disease, Gait Analysis, Deep Learning, Hybrid Model, CNN,
Transformer, Dynamic Feature Fusion, Diagnosis. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
LCSR–OPE: A MACHINE LEARNING–DRIVEN LEVEL-WISE COMPOSITE FRAMEWORK FOR
OPERATIONAL PROFILE–BASED SOFTWARE RELIABILITY PREDICTION |
|
Author: |
MANTENA JEEVANA SUJITHA, KODUKULA SUBRAHMANYAM |
|
Abstract: |
Accurate estimation of software reliability and ex plication of critical quality
attributes affecting failure behaviour are long-standing issues in the area of
software engineering. Traditional reliability models that are based on a profile
of operation approach the software as a homogenous entity, which overlook
in-vivo influence of intra-code metrics and the dynamic variation in its
operation. In order to address these limitations, the current research proposes
the Level wise Composite Software Reliability with Operational Profile
Evaluation (LCSR-OP) framework, which combines the frequency models of the
software’s operation and defect classification methods based on machine learning
to enable adaptive risk estimation of the software reliability. The framework
combines static code measures, like the number of lines of code, cyclomatic
complexity, number of branches, and the measure of operators, with probabilistic
operational profile distributions to create usage based module clusters. A
level-wise composite reliability function is then built; within this reliability
function, the probability density of each operational profile used to inform the
strategy of fault detection and prioritization for tests. XGBoost based fault
classification coupled with correlation driven feature importance analysis is
presented for improving the reliability estimations and also to rank the windows
attributes. Experimental validation on the basis of the data sets of the
Modelled Data Provider (MDP) of the National Aeronautics and Space
Administration, i.e., JM1, KC1, KC2 and PC1 shows that the proposed model
improves the fault detection efficiency and accuracy of reliability prediction
by 8 12%. |
|
Keywords: |
Software Reliability Engineering; Operational Profile Evaluation (OPE);
Level-Wise Composite Reliability (LCSR); Software Quality. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
DESIGN AND OPTIMIZATION OF A LOW-POWER HIGH-EFFICIENCY GAN-BASED LNA FOR X-BAND
MILITARY |
|
Author: |
SIVAPURAM DILEEP KUMAR , SHAIK MASTHAN BASHA , MADAM ARAVIND KUMAR |
|
Abstract: |
The Due to growing interest and demand for applications of military radar and
communication systems, the necessity appears in creating more advanced and
efficient, as well as less energy-consuming front-end devices for the
utilization in such systems. It should be mentioned that it becomes essential to
create more advanced low-noise amplifiers operating in the frequency range of
X-band (8-12 GHz). It will be discussed here about the development of innovative
LNA for military use with very low power consumption (under 3 W) and high
efficiency. Taking into consideration all advantages of GaN technology including
high electron mobility, breakdown voltage and high thermal stability, the
presented solution was created with the objective to achieve high RF parameters
in respect of low power usage. The stacking of CS topology provides a great
increase in gain with relatively small power usage. Moreover, the impedance
matching network is optimized to reduce reflection losses in order to achieve
efficient power usage. Uniform biasing of all LNA stages is achieved using
current-mirror biasing method. The designed LNA is analyzed and simulated
through EM and circuit simulations. Through simulations, a gain of around 18–22
dB along with a noise figure of about 1.5–2.2 dB and return loss less than −15
dB are achieved for the entire X band range while maintaining an input power
consumption of no more than 3 W. |
|
Keywords: |
Gallium Nitride (GaN), Low Noise Amplifier (LNA),X-band (8–12 GHz),Low-power
design, Impedance matching, Current mirror biasing. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
Title: |
A ROBUST FRAMEWORK FOR VIDEO FRAME EXTRACTION USING ENHANCED DEFORMABLE TEMPLATE
MODELS |
|
Author: |
USHA RANI J , RAVIRAJ P |
|
Abstract: |
The Enhanced Deformable Template Matching (EDTM) framework is designed to
achieve efficient and highly accurate frame extraction from video sequences.
EDTM integrates deformable model analysis with dynamic template adaptation to
effectively handle non-rigid object variations, occlusions, image noise, and
changes in illumination or viewpoint. The method employs oriented multi-scale
Gaussian derivative filter banks along with a robust similarity measure,
enabling precise matching of deformable objects using only gray-level
information—particularly advantageous for low-quality or color-limited videos.
To strengthen temporal coherence, EDTM incorporates flow-guided deformable
convolution for video frame interpolation. By combining optical flow estimation
with adaptive sampling, the framework improves intermediate frame generation,
resulting in smoother transitions and more reliable frame selection. The system
further utilizes two complementary template-matching modules: a multi-level
similarity computation unit and an edge-based control-point analysis unit, both
of which enhance geometric robustness and matching accuracy.
Experimental
evaluations demonstrate that EDTM consistently outperforms existing approaches
in both accuracy and computational efficiency across diverse video datasets.
Owing to these strengths, the framework is well-suited for applications
including video summarization, surveillance, facial recognition, and medical
imaging. |
|
Keywords: |
Deformable Template Matching, Object matching, Traditional Template matching,
Deep Learning, Active Shape Models. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
30th June 2026 -- Vol. 104. No. 12-- 2026 |
|
Full
Text |
|
|
|
