Iraqi Journal for Computers and Informatics

This paper presents a newly collected and highly relevant dataset on students' abnormal behavior in online exams. This dataset focuses on assisting research in building machine-learning models that allow for maintaining academic integrity during the era of online exams. Properly, more than 8,500 annotated images of normal and abnormal behaviors of students during remote examination are held in the dataset hosted at the Harvard Dataverse repository. The dataset has two versions: the original and the augmented. We utilize semantic segmentation and deep learning techniques in the applied data augmentation; this dataset provides a crucial foundation for developing and benchmarking intelligent proctoring systems. We evaluate the dataset using YOLO5 and our improved SPL-YOLO5 model, and the resulting mean average precision (mAP) is close to 1.0.

Because of new multimedia types and advances in technology, protecting data has become extremely valuable. Because of the changes happening around us, cryptography continues to protect information that is crucial for security. Security methods and systems are continually being improved, but we should continue to review data protection methods while data travel through different services. One of the reasons data systems are secure today is because of encryption which makes plaintext into ciphertext. It examines the strengths and weaknesses of various cryptographic algorithms using what has been written by other authors. This paper reviews the use of cryptography in several literatures to make data security better nowadays. Data is protected using encryption during online and other transfers, yet the information may be obtained from an unauthorized attacker who is persistent enough. Two or more approaches to security should be combined, according to what is highlighted in the article. AES encryption and decryption can be improved by adding other algorithms such as the replacement algorithm. AES is used in this study to safeguard a message before it is presented. In 1998, Rijmen and Daemen designed the AES algorithm and named it Rijndael. Many people have started using it because its security is strong and it is not easily broken using brute force. The technique is applied on the plaintext of the AES algorithm. Thanks to this approach, breaking the encryption is extremely difficult, as you need the right key, though AES itself remains a simple system. The performances of both Advanced Encryption Standard and its modified option were tested after they were put into practice. There was an avalanche of 54.69% with the new version of Advanced Encryption Standard compared to the 50.78% observed in the standard AES. The strong encryption

One of the most fatal and well-known diseases worldwide, heart disease claims the lives of many people every year. In order to preserve lives, early detection regarding such disease is essential. One of the quickest, practical, and affordable methods of disease detection is Data Mining DM, an artificial intelligence AI technology. Human life is saved by healthcare services through prompt and efficient decision-making. For forecasting, decision-making, and disease prediction, DM technologies are essential. This research predicts heart disease using DM algorithms. There are 14 attributes in Cleveland dataset, including blood fat, blood pressure, gender, and age. The probability regarding patients developing heart disease in the future can be forecasted by analyzing such parameters. For classifying if heart disease is present or absent, two classification algorithms are used: Logistic Regression LR and K-Nearest Neighbor KNN. The precision, accuracy, f-score, and recall of the suggested model are evaluated. The outcomes of suggested model were tested using the heart disease dataset. Without preprocessing the dataset's variation values, the LR and KNN algorithms achieved the highest accuracy (61% and 71%, respectively). The algorithms (LR and KNN) preprocessed the dataset's variation values to get the highest accuracy (90% and 93%). In order to improve data driven medical decision-making, the presented research demonstrates how well DM algorithms work to increase heart disease prediction accuracy.

Excellent performance in artificial intelligence image classification leads to extensive applications throughout areas such as healthcare facilities, robotic systems and multimedia platforms. The research field has evolved through new developments in both Vision Transformers ViTs alongside Graph Neural Networks GNNs. A new image classification method utilizes integrated Vision Transformers ViTs and Graph Attention Networks GATs to improve results for difficult dataset types. The hybrid architecture made possible by combining ViTs with GATs successfully captures complex relationships within visual data because ViTs deliver powerful global feature extraction while GATs establish strong patch-level dependencies. The implementation of GATs via their built-in attention mechanism allows dynamic region prioritization for both accurate recognition and better interpretability of images. The experiments using benchmark datasets CIFAR-10, CIFAR-100, ImageNet, Fashion-MNIST, and SVHN show that ViT & GAT outperforms Swin Transformer and ConvNeXt for state-of-the-art architectures. The proposed ViT & GAT model establishes results above competing state-of-the-art architectures reaching 92% accuracy on CIFAR-10 and 90% accuracy on CIFAR-100 while attaining 85% accuracy on ImageNet. This framework achieves 99% top-5 accuracy on CIFAR-10 and 94% on ImageNet besides reaching 92% accuracy on CIFAR-10 and 90% on CIFAR-100 and 85% on ImageNet. The proposed model shows excellent ability to generalize while minimizing over fitting and demonstrating resistance against noise and adversarial disturbances. The evaluation with F1-score, AUC-ROC, and confusion matrices proves the effectiveness of ViT & GAT by establishing it as a new benchmark within image classification field.

In the always-evolving and dynamic field of software development, good decision-making is absolutely critical. Developers have to regularly decide how best to apply features, optimize performance and debug issues. This process could be much improved by extracting actionable insights from software code. The presented work explores the tools and metrics available to enable developers to make data-driven decisions, therefore enhancing the development efficiency as well as code quality. Also, it introduces a new automated tool called CodeLens which analyzes software code, extract lines of code (LOC), documentation quality, complexity, and other key criteria. Through a consolidated view of such metrics, the tool helps developers evaluate code fit, spot possible bottlenecks, and prioritize optimization or refactoring efforts. Furthermore, the tool's support of Java and Python languages guarantees general applicability, hence fitting for many software projects.

Artificial intelligence chatbots (e.g., ChatGPT, Claude, and Llama, etc.), also known as large language models (LLMs), are continually evolving to be an essential part of the digital tools we use, but are plagued with the phenomenon of hallucination. This paper gives an overview of this phenomenon, discussing its different types, the multi-faceted reasons that lead to it, its impact, and the statement regarding the inherent nature of current LLMs that make hallucinations inevitable. After examining several techniques, each chosen for their different implementation, to detect and mitigate hallucinations, including enhanced training, tagged-context prompts, contrastive learning, and semantic entropy analysis, the work concludes that none are efficient to mitigate hallucinations when they occur. The phenomenon is here to stay, hence calling for robust user awareness and verification mechanisms, stepping short of absolute dependence on these models in healthcare, journalism, legal services, finance, and other critical applications that require accurate and reliable information to ensure informed decisions.

Electronic Health Records EHRs have revolutionized healthcare by storing patient data in a digital format, increasing accessibility and efficiency. However, the flaws of traditional authentication methods necessitate the development of advanced security solutions. This study presents a novel methodology integrating AES-256 encryption with DNA-based steganography to enhance biometric verification in Electronic Health Records EHRs. The approach involves extracting Short Tandem Repeats STRs and Single Nucleotide Polymorphisms SNPs from DNA profiles, encoding the genetic data into binary format, and securing it with AES-256 encryption to ensure high confidentiality. Encrypted DNA profiles are embedded in MRI images by using the Discrete Cosine Transform DCT, which ensures the concealed data remains both imperceptible and secure against unauthorized alterations, and during the authentication phase, and the encrypted genetic information is extracted, decrypted, and matched with reference samples for verification, the experimental results indicate that the system significantly improves both biometric security and medical data protection, with average processing time of approximately 320 milliseconds, and as it exhibits strong resistance to tampering, achieving has a 99.8% success rate in preventing unauthorized modifications. The embedding method maintains a high level of image quality, reflected by a Peak Signal-to-Noise Ratio PSNR of around 47 dB, confirming that the diagnostic utility of MRI images remains unaffected. This approach effectively combines biometric security with medical data safeguarding, providing a dependable and scalable solution for patient authentication in electronic health record EHR systems.

Federated Learning FL is a rapidly evolving machine learning paradigm that enables collaborative model training across decentralized data sources while preserving data privacy. Since its inception in 2016, FL has emerged as a transformative approach in domains such as healthcare, IoT, and edge computing, where data sensitivity and regulatory constraints limit centralized processing. This systematic review consolidates findings from 50 high-quality studies selected from over 250 papers to present a comprehensive synthesis of FL methodologies, core aggregation techniques, privacy-preserving mechanisms, security threats, domain-specific applications, and emerging trends. We categorize challenges into communication overhead, heterogeneity (device, data, model), fairness, trust, and evaluation inconsistencies. We highlight research gaps, notably in standardized evaluation, incentive mechanisms, and deployment scalability. By reviewing recent advances such as vertical federated learning, federated unlearning, and blockchain-based incentives, this paper offers a roadmap for future research and identifies open questions vital for widespread FL adoption.

The ability to navigate safely is a critical issue for people with visual impairments, especially in urban cities where traffic flow is a constant danger. While traditional tools such as white canes and guide dogs provide some assistance, they are often limited in detecting fast-moving hazards (such as a vehicle) and lack integration with modern technologies. The goal of this research paper is to create and test a prototype of a smart system based on the Internet of Things (IoT) that will help blind individuals become more aware of their surroundings and be able to get around on their own. There are two primary pieces to the system: a smartwatch that one can wear and a mobile web-based application. The smartwatch is built on the NodeMCU ESP8266 module and has a GPS module, a buzzer, and an LED indication. All of them are built into a comfortable design that can be worn on the wrist. The system keeps an eye on where the blind person is and gets alerts through a real-time Firebase database. The web program runs on the driver's phone, tracks the car's location, figures out how close it is to the blind user, and sends warnings when the predicted arrival time drops below two minutes. We did field experiments at different speeds (10 km/h, 20 km/h, and 50 km/h) across a distance of 2 kilometers. The results demonstrated a high response accuracy exceeding 99.6%, stable GPS accuracy ranging between 96% and 98%, and an average alert response time of less than 1.5 seconds. These results confirm the system's ability to provide immediate alerts and improve the safety of the blind. By combining real-time cloud services, geolocation, and alerts, this prototype provides an effective system that enhances the safety and mobility of blind people in urban environments. The system's performance indicators support its effectiveness in providing two-way interactive alerts and proactive responses to protect the user.

Blockchain has become a critical enabler of secure data sharing in domains such as healthcare, finance, and digital identity. However, its reliance on classical cryptographic schemes (e.g., RSA, ECDSA, SHA-256) makes current systems vulnerable to emerging quantum computing attacks, raising risks to data confidentiality, integrity, and long-term trust. This paper addresses this challenge by proposing a modular hybrid framework that integrates post-quantum cryptographic (PQC) techniques into blockchain-based personal data sharing. The framework combines lattice-based encryption for protecting off-chain data, hash-based signatures for smart contract authentication, and quantum-safe zero-knowledge proofs and trusted execution environments (TEEs) for privacy-preserving verification and secure key management. To ground this design, we conducted a systematic literature review of 35 studies published between 2018 and 2025, analyzing security, scalability, interoperability, regulatory alignment, and user autonomy. Findings reveal that only 5 out of 35studies (14%) explicitly addressed quantum threats, with over 80% focusing on theoretical resilience without testing implementation constraints. Furthermore, 90% of proposals neglected smart contract compatibility, and only 8% (3/35) incorporated TEEs, underscoring implementation barriers in contract execution, secure key management, and performance integration. Prototype evaluation demonstrated that the framework sustained 1,500 TPSon Hyperledger Fabric, achieved a 75% reduction in storage bloatusing IPFS, and supported GDPR-aligned workflows with 99.98% audit log completion and 95% successful erasure requests. Privacy was further strengthened through zk-STARK proofs, which reduced unauthorized access by 40%, while TEEs improved key management efficiency by ~28%. Although PQC introduced 5–12 seconds of latency, consent revocation was processed in under 2.1 seconds, highlighting both the feasibility and trade-offs of practical post-quantum deployment. This work demonstrates a clear pathway toward quantum-resilient blockchain infrastructures that safeguard personal data, comply with regulatory standards, and maintain user trust in the quantum era.

Automatic Text Summarization is a crucial feature for managing the ever-increasing volume of textual data. However, existing methods often struggle with feature identification for sentence importance, which leads to a lack of maintained narrative coherence and accuracy. In this proposed approach, the summarization process leverages the Chi-square Binary Cuckoo Search (Chi-BCS) method for feature selection, this optimizes text features enhance the summary content and utilizes insights from classification to ensure summaries are contextually relevant and concise. Feature selection aims to improve the performance of machine learning models by reducing the dimensionality of the input data and removing irrelevant or redundant features. Classification, on the other hand, contributes to better summarization by distilling lengthy or redundant content into key points, thereby enhancing both efficiency and accuracy. The proposed approach implements a model that leverages advanced Natural Language Processing and machine learning techniques for effective extractive summarization on both BBC and CNN/DailyMail datasets. Key features extracted from the text include Named Entity Recognition, Cue phrases, TF-IDF, Sentence position, sentiment analysis, etc. Various algorithms are employed to improve classification performance, such as Decision Trees, Support Vector Classifier, Gradient Boosting, Random Forest, K-Nearest Neighbors, and Logistic Regression. Among all the methods evaluated, the Random Forest and Ensemble Hard Voting approach achieved the highest F-score of 96.26 and 0.9322 respectively on the BBC and CNN/DailyMail dataset. In the text summary evaluation, the ensemble method also delivered exceptional results, with ROUGE-2 and ROUGE-L F1 scores reaching 0.799 and 0.818, respectively on BBC. While our ensemble model achieved to high score on ROUGE1 and ROUGE 2, reaching 0.275, 0.5017, respectively on CNN/DailyMail when compared with state of art, highlighting the model's strong performance. These findings demonstrate that the proposed model is highly effective for both the classification and summarization of large-scale textual data.

The trends of resource consumption and optimization mechanisms for blockchain-enabled security in edge-fog computing environments. While blockchain provides robust security for fog networks in a decentralized fashion, its demand for resources creates tremendous challenge in resource-constrained settings. Through in-depth examination of a Practical Byzantine Fault Tolerance PBFT-based blockchain deployment across 50 edge devices and 10 fog nodes. The study reveals the most critical resource bottlenecks and proposes an adaptive resource management framework that maximizes the tradeoff between security requirements and operational efficiency dynamically. The proposed work shows that data-type-based optimization and intelligent workload distribution can reduce CPU utilization by 27%, memory by 22%, and network bandwidth by 38% without sacrificing security assurance. The introduction of a novel dynamic resource allocation algorithm that adjusts consensus participation and cryptographic strength to current system conditions, demonstrating that security-performance trade-offs can be optimally resolved through context-sensitive optimization. These advancements are a move towards resource-constrained security architectures for edge-fog computing, enabling the broader applicability of blockchain security in resource-poor IoT environments.

Emerging technologies have accelerated the digitalization of smart grids, improving demand-side management, sustainability, and operational efficiency. The attack surface is widened by this interconnection, though, leaving vital smart grid data and systems vulnerable to online attacks. Single points of failure, privacy violations, and a lack of robustness against sophisticated attacks persist in centralized data processing. Traditional cryptographic techniques are further threatened by the development of quantum computing, which raises significant security risks for smart grids. With a focus on post-quantum cryptography (PQC) resilience, this study examines 206 peer-reviewed research articles on blockchain-based federated learning (BFL) in smart grids that were published between January 2023 and July 2025. It assesses the advantages, limitations, and compromises of the current BFL models in this field. The paper suggests a unique post-quantum secure BFL (PQS-BFL) framework that integrates federated learning (FL), lightweight PQC protocols, and a scalable blockchain architecture to solve the vulnerabilities that have been uncovered. This design enables decentralized, private, and impenetrable cooperation among grid nodes. The results demonstrate that the system mitigates quantum-resilient attacks and inference threats while improving data integrity, key management, and secure model aggregation. A path for creating safe, scalable PQS-BFL solutions for upcoming smart energy systems is provided in the paper's conclusion, along with an overview of the main research issues. This study shows that using PQC, blockchain, and FL to secure next-generation smart grids is both feasible and important.

As digital threats become increasingly sophisticated, user authentication has become vitally important in cybersecurity. Traditional authentication methods such as passwords are under increasing assault from a range of attacks. Behavioral biometrics, such as mouse dynamics, have the potential to address these attacks in a way that is largely passive and continuous. In this paper, we present a new solution that rests on mouse dynamics behavior together with a lightweight deep learning model inspired by EfficientNet, specifically designed for Behavioral Assessment of Numerical Data (BAND). The SapiMouse dataset, consisting of mouse tracking data from 120 actual users, is harnessed. By applying preprocessing techniques such as Quantile Transformation and Min-Max Encoding, along with encoding, the raw data were prepared for model training. The modified EfficientNet model retains its computational efficiency while also being tailored to work with numerical input. Its structure uses compact convolutions along with compound scaling to capture time-series mouse data discriminative features, lowering the processing burden while maintaining accuracy. Moreover, to stabilize training and enhance generalization, dropout and batch normalization layers were added, ensuring robustness to overfitting, even when using data generated by a model. CGAN’s capacity for class sample synthesis was harnessed towards improving recognition of unused user profiles, resulting in a total of 240 unique classes (120 real + 120 synthetic). The model reached an accuracy of 99.24% for classification and a macro-averaged F1-score of 0.991 on the testing set. An inference time of only 0.2331 seconds per sample, alongside a cumulative training duration of 158.25 seconds, suggests real-time applicability. These findings support the promise of repurposing advanced deep learning models for behavioral biometrics, providing affordable, scalable, and efficient user verification for sensitive security contexts.

The predictive validity of machine learning models depends on the training data. In some cases, training data contains historical, social, or demographic inequalities, which leads algorithms to reproduce unfair results. This paper proposes a fairness-constrained stacking meta-learning approach for reducing bias in classification by aggregating a set of classifiers through a constrained ensemble learning scheme. A set of base classifiers, including Decision Tree, Naive Bayes, Support Vector Machine (SVM), and LightGBM, are trained and evaluated on the Adult Census Income dataset using both predictive and fairness metrics. The final meta-model is constructed as an aggregation of only the fair-performing models, while models failing to meet the fairness threshold are excluded. Learned weights are then optimized to maximize the F1-score while maintaining fairness constraints. Experimental results demonstrate that the proposed method achieves predictive performance (Accuracy = 0.91, F1-score = 0.82) while substantially reducing disparity between demographic groups (EOD = 0.03 for sex and 0.04 for race). These findings indicate that fairness-aware stacking ensembles can provide a solution for mitigating algorithmic bias through an aggregation framework that balances accuracy and fairness.

Phishing attacks continue to grow rapidly, often using obfuscated and deceptive URL patterns to mimic legitimate websites and evade detection. While traditional Machine Learning (ML) models perform well on benchmark datasets, they often struggle in real-world scenarios where phishing URLs are carefully crafted to resemble authentic domains. This study presents a dataset-driven comparison between traditional ML models—Logistic Regression, K-Nearest Neighbors, Support Vector Machine, and Random Forest—and advanced approaches such as Extreme Gradient Boosting (XGBoost), a tuned XGBoost variant, and a soft voting ensemble. Two datasets were used: (i) a high-variance global dataset from Mendeley Data, and (ii) a custom-built local dataset with region-specific phishing URLs designed with minimal alterations (e.g., character substitutions and deceptive subdomains) to simulate low-variance attacks. Preprocessing included feature engineering, variance analysis, and balancing with the Synthetic Minority Oversampling Technique (SMOTE). Experimental results show that Random Forest outperforms other traditional models but still struggles with low-variance phishing URLs. In contrast, advanced models—particularly tuned XGBoost—achieved significantly higher recall (0.99) and strong precision (0.81), while the voting ensemble further improved robustness by combining multiple classifiers. These findings emphasize the importance of realistic datasets and demonstrate that advanced ML strategies are more effective for detecting phishing attempts based on subtle obfuscation. This work contributes by (i) validating advanced ML models under realistic low-variance conditions, and (ii) highlighting precision and recall as more appropriate evaluation metrics than accuracy in cybersecurity.

The increasing prevalence of Internet of Things (IoT) devices has raised significant concerns regarding their security. Malware attacks on these devices can lead to severe consequences, including data breaches, privacy violations, and system failures. This paper proposes a novel approach for detecting malware in cross-architecture IoT devices using deep learning with Convolutional Neural Networks (CNNs). The proposed methodology involves converting binary files into grayscale images and utilizing a CNN model for feature extraction and classification. The model's performance was evaluated on a dataset comprising malware samples from various IoT devices, achieving an accuracy of 97%. The results demonstrate the effectiveness of the proposed approach, outperforming existing methodologies in detecting malware in cross-architecture IoT environments. The model's robustness and adaptability across various malware samples highlight its potential as a valuable tool for enhancing IoT security. Future work will focus on expanding the dataset to incorporate more diverse and complex malware samples and exploring additional deep learning architectures.

The Internet of Things (IoT) has risen rapidly and created a need for computing architectures with the ability to aid real-time applications with the least latency. Conventional cloud-only setups often experience high latency due to physical distance and network congestion, thus, unsuitable for time-sensitive applications. Four of the architectural models currently proposed in the literature are assessed in the present paper: Cloud-only, Edge-only, Fog Multi-layer, and Smart Hybrid; each of them is evaluated in a large-scale simulation aimed at simulating a smart city environment with 1,000 various IoT devices, such as sensors, cameras, and wearables generating dynamic traffic. The outcomes indicate that Edge-only architecture recorded the least mean latency of 0.2417 seconds compared to Smart Hybrid with 0.3156 seconds and Fog Multi-layer with 0.3874 seconds mean latencies whereas the Cloud-only model presented the highest latency with 0.5231 seconds. Given that there existed significant differences between the architectures (p < 0.0001) based on the results of the statistical analysis carried using the one-way ANOVA. Such results explain the promise of adaptive and context-sensitive hybrid systems and allow reducing latency in any IoT deployment in the future and designing an IoT-oriented system that supports latency-sensitive applications in smart cities in a more scalable manner.

Object detection and counting is a crucial problem in various areas like autonomous machines, health sector and industrial automation. This paper presents a comparative study between two Machine Learning (ML) methods, depending on Single Shot Multi Box Detector (SSD) with You Only Look Once (YOLOv3) and MobileNetv3 to focus on the object detection and counting. The well kown dataset (COCO) used to evaluate the two models where 44 images chosen randomly for testing. In addition SSD and MobileNet v3 compared in a real time mode. The performance metrics used in this paper were confidence scores, average over image and processing time. The results indicate that there was an tradeoff between accuracy and speed. Both models showed a fast inference time, which made it suitable for real-time applications, although (YOLOv3) was more confident in some cases because of its complex architecture. The performance difference is caused by the design of the model: MobileNet's simple structure aims at lightweight, while (YOLOv3)'s complex network enhance robustness in detection. This work highlights The necessity for application-aware model selection balancing between the fast performance on edge devices (such as drones) versus accuracy for accuracy-precision applications (like medical imaging). The findings provide practical insights for deploying ML-driven object detectors across various domain.