Journal of Information Technology Management (مدیریت فناوری اطلاعات)

Image classification is a significant process in the field of computer science. It has applications in every field, such as spam detection in emails, medical diagnosis, image recognition, sentiment analysis, object detection, weather forecasting, pattern recognition, and security. Image classification deals with the grouping of images based on labels or characteristics. Feature extraction, feature selection, feature reduction, and classification are the main steps used to classify images. A medicinal and non-medicinal flowers data set is prepared by clicking images for the study. Methodology is used to achieve satisfactory classification results on the seeds, Wisconsin Diagnostic Breast Cancer, Heart Failure Clinical Records, and Wisconsin Prognostic Breast Cancer data sets, which are taken from the University of California, Irvine (UCI) repository. The proposed methodology suggests an efficient feature extraction and selection approach for data sets under consideration. An information gain-based genetic algorithm is used for feature reduction. It is performed on the extracted features to retrieve an optimized feature set. Fitness of the features is evaluated to choose the most relevant features. A neural network is used to classify the obtained feature subset. Better classification results are attained with the help of feature extraction and feature reduction.

This paper focuses on a new hybrid machine learning model for classifying eye states from EEG signals by integrating traditional techniques with deep learning methods. Our Hybrid LSTM-KNN architecture employs KNN for classification and uses LSTM networks to extract features temporally. In addition, we perform extensive feature engineering, including statistical Z-test and IQR filtering, dimensionality reduction using PCA, and multivariate analysis to further model the performance. Moreover, an SVM-based unsupervised clustering approach is proposed to partition the EEG feature space, followed by ensemble learning in each cluster to improve accuracy and robustness. Using the EEG Eye State Dataset for the first assessment, the Hybrid LSTM-KNN model recorded an accuracy of 87.2% without PCA. Further improvements through statistical filtering outperformed initial expectations, achieving a 6% rise in performance to 89.1% after outlier removal, 89.1% with Z-test (σ = 3), and 88.3% with IQR (1.5x). After applying PCA along with ensemble learning post clustering, the final model exceeded expectations with an accuracy and F1 score of 96.8%, surpassing Ensemble Cluster-KNN and traditional models based on Ensemble Cluster-KNN, Logistic Regression, SVM, and Random Forest. The outcome demonstrates the robustness and noise-resilience of the model’s performance in practical real-time brain-computer interface and cognitive monitoring systems.

This research work aims to present a robust deep learning framework by devising a deep learning-based ensemble method of YOLOv8 and EfficientNet. The suggested model is evaluated on the dataset collected from Kaggle, comprising 10,000 high-definition images of stems, leaves, and cut fruits of banana and papaya. These images are captured under different lighting conditions and thus expanded to 80,000 images. Authors have proposed an ensemble model comprising YoloV8 and EfficientNet as base deep learning models to enhance prediction and classification performance. Here, authors combine the merits of both models, i.e., speed of YoloV8 and the accuracy of EfficientNet, by putting a majority voting method in place. The final forecast is determined by majority voting, and EfficientNet is given higher significance in the situation of a tie owing to its enhanced accuracy. The proposed model presents a robust solution for agricultural disease management and demonstrates significant improvements in the detection of diseases in papaya and banana, opening avenues for its widespread employment in real life.

Federated Learning (FL) has emerged as a revolutionary technique for distributed machine learning for training a model on shared data without sharing the data itself. Nevertheless, privacy-related concerns and scalability difficulties remain a problem. This paper discusses the state-of-the-art works to improve the privacy and convergence at FL frameworks for targeted healthcare and financial applications, as well as smart devices. It focuses on methodologies that preserve user privacy, such as differential privacy, homomorphic encryption, secure multi-party computation, and methods that enhance the model’s efficiency, including model compression, communication optimization, and adaptive optimization algorithms. To overcome these challenges, this study helps in the future design of FL systems for vital domains with high scalability.

Vertical handover (VHO) in heterogeneous wireless networks is essential to keep users continuously connected and also to guarantee that the Quality of Service (QoS) of mobile communications is adequate. The focus of this paper is to apply various Multi-Criteria Decision Making (MCDM) tools to model a comprehensive VHO decision-making framework. This proposed methodology brings together more than one parameter of a given networking environment, such as signal intensity, QoS, and energy requirements, into one or more decision models. Using analytical methods such as the Analytical Hierarchy Process (AHP) for criteria weighting, we continuously optimize network conditions, thereby enhancing the efficiency and reliability of Vertical Handover (VHO). Several experiments were made to test the efficiency of the given MCDM-based VHO algorithm. The performance evaluation of the proposed method reveals superior handover performances in terms of success rates, less latency, and better QoS as compared to other VHO techniques. In addition, our research conclusion implies that integrating MCDM into the VHO decision-making will not only facilitate the network resource optimization but also improve the user satisfaction in the heterogeneous networking domain. This paper, in the wireless communications area, makes a significant contribution by presenting a powerful framework for DNS and VHO. The subsequent studies will be directed towards improving this algorithm for real-time applications and experimentation in the new generation networks, such as 5G Networks.

Temporal databases, quickly rising in size, are distinguished by their capacity to maintain the older version of data objects against actions on them, allowing logical deletions. Queries for historical data are particularly costly due to the linear scanning of temporal versions. Temporal data structures like time-split B-Tree or multiversion B-Tree are working underlying the state-of-the-art temporal databases. So far, most efficient temporal data structures are partially persistent or fully persistent, but none of them support retroactive queries. On the other hand, efficient temporal indexing is required to address bulk loading in a real-life application. To the best of our knowledge, there is no efficient solution for bulk loading and updating retroactive index structures. This article seeks to offer a new data structure, the Retroactive B-Tree (RBT), to facilitate retroactive operations in temporal databases as well as bulk loading. It presents theoretical and empirical research and analysis of the suggested data structure and its relevant operations. The experiments were conducted to demonstrate the performance of the proposed retroactive B-Tree in terms of execution time, I/O complexity, space complexity, and bulk loading. The obtained results show that indexing with a buffer is the most powerful model for existing temporal databases for implementing a retroactive B-Tree. The tree of lists architecture is observed as an I/O efficient data structure for all variants of temporal indexing for large databases.

With an increasing reliance on mobile devices, continuous and assured user authentication is essential to protect sensitive personal data and digital interactions from unwanted access. Based on this background, this research proposed the development of the HybridTouch framework for smartphone-based continuous and passive user authentication. The proposed HybridTouch combines Convolutional Neural Networks for spatial feature extraction and Gated Recurrent Units for temporal sequence analysis. It uses accelerometer, gyroscope, and touch data to take advantage of the unique behavioral patterns captured by it. Innovative preprocessing techniques have been incorporated into the proposed approach: Discrete Wavelet Transform is used for signal denoising, and Variable-Length Adaptive Temporal windowing is used for segmentation based on signal entropy to enhance feature representation. To eliminate the data scarcity limitation, Generative Adversarial Networks were used to synthesize realistic behavioral data that considerably augmented the dataset and enhanced model generalization capability. Extensive experiments conducted on the Hand Movement, Orientation, and Grasp (HMOG) dataset showed that the proposed HybridTouch achieved excellent results with authentication accuracy up to 98.8% with real data, growing up to 99% with GAN-augmented data. The hybrid model further has an equal error rate of 1.4% on real data and 1.25% on synthetic data, which is better than any other models currently present (Sağbas et al., 2024; Siddiqui et al., 2022; Abuhamad et al., 2020) and all implementations of standalone convolutional neural networks and gated recurrent units.

This study aims to present a novel generative AI-driven system for hyper-personalized health monitoring. Dynamic data processing, predictive modeling, and flexible learning improve real-time health evaluations. By combining weighted feature aggregation, iterative least squares estimation, and selective feature extraction, the suggested strategy makes predictions that are more accurate while using less computer power. Abnormality detection methods like adaptive thresholding and Kalman filtering provide accurate health monitoring. Attention, gradient-based optimization, and sequence learning improve health trend forecasts as the model improves. Generative AI-driven wearables outperform conventional and AI-based alternatives in many key performance tests. These evaluations include prediction accuracy (94%), real-time monitoring efficiency (93%), adaptability (92%), data integration quality (95%), and system reaction time (90 ms). These devices are safer (96%), have longer battery life (32 hours), and are simpler, more comfortable, and scalable. The results suggest that creative AI can transform personal healthcare into something more adaptable, safe, and affordable. Generative AI-powered smart gadgets are the most sophisticated means to monitor health in real time and deliver individualized, data-driven medical treatment. Future research will concentrate on improving prediction models and developing AI-driven modification approaches to make them more effective in additional healthcare scenarios.

The widespread adoption of Android phones has heightened concerns about user privacy. This research presents an Adaptive Privacy Management System (APMS) that integrates Machine Learning (ML) models with Differential Privacy techniques to enhance privacy protection. The APMS monitors application behavior and employs ML algorithms to detect anomalies and enable context-aware privacy enforcement. Differential Privacy ensures that sensitive data remains protected through the addition of noise and privacy-preserving computations. Experimental results demonstrate that the APMS achieves a 92.5% accuracy rate in detecting the privacy leakage. The anomaly detection model, using Random Forest, shows high accuracy (92.5%), recall (89.5%), and precision (73.9%), effectively identifying both normal and anomalous behaviors. Additionally, the impact of noise on data utility, controlled by the privacy budget (ε), is manageable. The results show that APMS is a robust system for safeguarding user confidential information, contributing to a more secure and privacy-centric Android ecosystem.

The proliferation of fake news, characterized by the dissemination of inaccurate information to deceive audiences, has become a pressing concern in recent times. Traditional approaches to phony news detection, often focused on analyzing Twitter content, are susceptible to noise and variations in input sequences, leading to suboptimal performance. To address these challenges, this study proposes a novel method called Multi-Head Attention-Hierarchical Bidirectional Long Short-Term Memory (MHA-HBiLSTM) Networks. Our approach involves two phases: training and testing, wherein we employ tweet pre-processing techniques such as stemming, punctuation removal, stop-word elimination, URL handling, and Twitter control removal. Features are represented using the Glove word embedding technique for experimental evaluation and comparison. The MHA-HBiLSTM model integrates multi-head attention and hierarchical concepts, allowing meaningful information extraction from Twitter data. Notably, our model utilizes dual-level attention mechanisms and a hierarchical structure, reflecting the inherent hierarchy in documents and prioritizing key material during document representation. The effectiveness of the proposed MHA-HBiLSTM algorithm is evaluated using the Whale & Multi-Verse (W-MVO) Optimizer approach, with tests conducted on Kaggle and FakeNewsNet datasets. Comparative analysis with traditional machine learning approaches and deep learning models demonstrates the superior performance of the MHA-HBiLSTM approach in fake news detection.