Al-Noor Journal for Information Technology and Cybersecurity

Abstract Identifying abandoned accounts is crucial as they can be breeding grounds for fraud or used as fake accounts, and addressing platform disorder is an important part of sustainable development , Their accumulation causes chaos .This research aims to use the C4.5 algorithm to build a decision tree due to its speed and ability to manage categorical and numerical inputs in identifying features to classify Facebook accounts as active, abandoned, or deactivated by determining the number of days since the last interaction on the account. Data was obtained from a website and manually from other sources. The results of the decision tree demonstrated the discrimination capabilities derived from machine learning. The research presented a solution for recycling or deleting accounts to mitigate the damage caused by the accumulation of abandoned and deactivated accounts on the Facebook platform. The biggest obstacle for the researcher was the sample, as Facebook accounts are subject to privacy laws, and previous literature and studies have not addressed the treatment and recycling of these accounts. The results were satisfactory in the diagnostic process, providing a reliable method to ensure the safety and sustainability of interactions on this platform, which includes billions of users around the world.

Abstract This paper presents the “Delay Minimization with Random Mobility” (SMADM) framework as a theoretical controller and router algorithm for VANETs, aimed at minimizing end-to-end delay under random availability of links and rapid topology changes. The key contributions are the derivation of a mobility-aware “drift-plus-penalty” that converts long term delay minimization into decisions for each time slot, allowing for well defined routing options even when the network topology is changing rapidly over time. By integrating link continuity into the delay penalty while regulating queue growth, SMADM decouples the delay minimization objective from transient topological fluctuations, which typically destabilize greedy or pure geometric algorithms. The SMADM analysis aligns with well-established Lyapunov optimality bounds for access rates within the capacity region, and the resulting queue lengths are stable, with the achieved long-term delay penalty close to the upper bound. Furthermore, the framework has been validated through intensive simulations using SUMO (Simulation of Urban Mobility) in a representative urban environment. The results show SMADM significantly reducing end-to-end delay relative to GPSR and AODV, achieving randomized stability with a packet delivery ratio ranging from 45% to 75% under high traffic conditions.

Abstract Urban air quality assessment consists of both categorical interpretation of pollution severity and continuous estimation of pollutant concentrations. While the Air Quality Index (AQI) categories are effective at communicating with the public, particulate matter concentrations, such as PM2.5 and PM10, are also essential to the quantitative assessment of air quality. Although there is a high degree of correspondence between the two types of assessments, they are often modeled separately. This paper proposes to use a multi-task learning (MTL) approach that can be used to assess urban air quality using data from meteorological and air-quality related tabular datasets. Specifically, under the MTL approach, the proposed model will jointly conduct AQI category classification and PM2.5/PM10 regression using a shared neural network backbone with outputs specifically designed for each task. The performance of the MTL framework will be tested on the TRAQID dataset and compared with that of single-task models using the same data-preprocessing and data-partitioning techniques. Results indicate that the MTL approach creates an integrated modeling framework that provides stable performance across both tasks, achieving an AQI classification accuracy of 93.72%, and offering insight into the trade-off between task specialization and joint representation learning.

Abstract Plant leaf diseases present significant risks to agricultural productivity. The process for diagnosing plant leaf diseases involves the examination of leaves by experts, but this method is quite inefficient and inaccurate. This study presents a novel parallel scheme architecture of deep learning networks, which is based on parallel operation of individual binary classifiers, wherein each classifier is dedicated to identifying a particular disease. This architecture improves feature extraction while keeping low computational complexity and ensures reliability and portability of the model to be utilized on lightweight platforms such as mobile phones. Moreover, the proposed architecture is flexible and extensible to use new classifiers for particular diseases without adding any complexity. The model was trained with a part of the Plant Village data set that comprised images of corn, tomato, and apple crops, each with three types of diseases and a healthy one, with 513 images per category. Preprocessing techniques and data augmentation have been used to enhance generalization. This model has just 710,786 parameters and consumes 1.327 GFLOPs. Field images collected from nature were used to check the robustness of the model; however, environmental conditions and the scarcity of plants made it impossible to capture images of only 3–4 samples per crop through a mobile phone. This reason makes the evaluation preliminary. The accuracy by Parallel-CNN are: corn-Cercospora Leaf Spot 98.54%, Common Rust 100%, Northern Leaf Blight 100%; tomato-Bacterial Spot 100%, Early Blight 100%, Late Blight 98.54%; and apple-Apple Scab 95.63%, Black Rot 100%, and Cedar Apple Rust 100%.

Abstract Mental and cognitive fatigue have become more common with prolonged computer use or prolonged sitting in front of smart device screens in modern digital environments, significantly impacting concentration, individual performance, and overall health. This comprehensive review covers the scientific literature published from 2012 to 2025, examining AI-based approaches for detecting mental and cognitive fatigue using webcam data and nonverbal behavioral cues. Recent innovations in artificial intelligence and computer vision have led to the design of intelligent systems capable of detecting fatigue through behavioral and visual cues. This literature review examines some studies on AI-generated fatigue detection, focusing on approaches that use webcams to analyze nonverbal behavior. It also highlights the main trends in detection methods—physiological, visual, and dual—with an emphasis on the growing importance of deep learning models and multi-source data fusion in improving the efficiency and accuracy of detection. Research indicates a significant evolution from traditional feature extraction methods to intelligent network models that automatically learn features, which can recognize subtle indicators of fatigue such as eye blink rate, head movement and direction, and subtle facial expressions. This study represents a fundamental step towards building a sophisticated intelligent system based on nonverbal behavior analysis that can continuously analyze nonverbal behavior to detect fatigue during actual computer use.
This review aims to analyze and compare AI-based approaches for detecting mental and cognitive fatigue through nonverbal behavior and computer vision techniques, including recent methodological progress and important discoveries to inform future research directions.

Abstract State-sponsored cyberattacks are carried out to achieve pre-planned objectives, so their impact is significant. Defenders must respond, but the scale of the attack is large and there is a possibility that unknown vulnerabilities may be exploited, making response difficult. Furthermore, excessive response can reduce user availability and disrupt work. Therefore, a response policy that can effectively defend against attacks while ensuring user availability is necessary. This paper proposes a method to address this issue by collecting the number of process hydration sessions of Bob's assets in real time and utilizing this for learning. Using this method, we trained a reinforcement learning-based policy on a cyberattack simulator. As a result, the attack duration for two attacker models was reduced by 279 and 31 time-steps, respectively, based on 100 time-steps. Furthermore, the number of "resource actions" that impede user availability during the defense process was also reduced, resulting in a policy with better overall performance.

Abstract Deep learning has proven to be effective tool for improving the performance of wireless networks, for addressing many networks important problems including routing, security, spectrum sensing, resource allocation, and localization. Furthermore, integrating deep learning tools into networks can greatly increase the network's capacity to adjust to changing into smart environments and improve the overall stability and robustness of the system. As the demand for large network nodes and huge amounts of information, it can be used to determine and identify the network characteristics (such as best path, hotspots node, interference distribution, congestion points, traffic bottlenecks, spectrum availability, etc.) by analyzing a large number of network parameters (such as energy consumption, delay, lifetime, loss rate, packet overhead, etc.). This paper presents a review on recent published research that used different deep learning models to enhance performance of wireless network. In addition, the review focused primarily on the following topics in this field: user localization, routing, security, big data, mobility, network control, and other application. This article aims to help the readers awareness of the most recent proposed models and algorithms in deep learning-driven wireless network, as well as to identify relevant unresolved challenges that could be addressed in future study.

Abstract The rapid development of the Internet of Things (IoT) and other resource-constrained applications has increased the importance of research and development in lightweight cryptographic algorithms. This has led to a greater focus on security and performance in both hardware and software, aiming to optimize them to meet the demands of resource-constrained applications in terms of power and memory consumption, acceptable security levels, and the urgent need to fulfill these requirements. Numerous algorithms have been proposed, and international competitions have been held to select the best algorithm based on a simple structure and low computational complexity. This study reviews and analyzes several early and proposed lightweight cryptographic algorithms, focusing on their structure, mechanism, and parameters, as well as their characteristics, limitations, and the problems they solve. Block size, key size, memory space, execution time and etc. used as a matrix for comparison between the LWC algorithm. The selection of a good LWC algorithm depended on this matrix also with the algorithm's feature limitation and the environment nature that used the algorithm. The research provides a scientific reference basis for selecting the most suitable algorithm for a specific resource-constrained application.

Abstract There are several uses for the Generative Adver- sarial Network (GAN) technology. Anonymity, personal privacy, and protections for officials, managers, and powerful people are important legal concerns. These significant advancements, including facial recognition, are led by GAN-based technologies. This paper compares several Generative Adversarial Network types used in the de-identification field, depending on state of the art, such as privacy protection Generative Adversarial Net- work (PPGAN), conditional identity anonymization Generative Adversarial Network (CIAGAN), and semantic aware Genera- tive Adversarial Network (SAGAN), among others, to high-end products presented by researchers through multiple databases, including Celeb Face Attributes (CelebA), among others, to obtain the most accurate expressions and characteristics of real face images. Researchers used a range of techniques and strategies to present their findings and compare them to previous findings to obtain the best responses for de-identification. The strengths and weaknesses of developing new faces depend on the additions made to each proposed structure and the exploitation of raw resources into the basic system, which is reliant on the network’s structure. After discussing each technique and the relevant technique for assessing the output—such as Siamese for true/face verification and Learned Perceptual Image Patch Similarity (LPIPS)—they were grouped in a table with the other techniques to clarify the differences. This research has produced several insightful findings, such as increased interest in the subject of identity concealment and developments in GAN technology, which provoked scholarly discussion.