Alkadhim Journal for Computer Science

Digital misinformation threatens news credibility through sophisticated image forgery techniques (Copy-Move, Splicing, Deepfake). This study introduces the first Hybrid CNN-ViT framework for 4-class news media forgery detection, achieving 94.2% accuracy and 0.935 F1-score on a novel 10K News dataset. The gated fusion architecture optimally balances ResNet50+EfficientNetB3 (α=52.2%, local features) with ViT-B/16(β=47.8%, global context), enabling superior discrimination across Authentic (96.5%), Copy-Move (95.8%), Splicing (93.7%), and Deepfake (90.5%) categories. Class weights mitigate multi-class imbalance, with training stability confirmed (12.8 min GPU, no overfitting).Comparative analysis against 12 state-of-the-art studies demonstrates 1.5-11.9% superiority, establishing news-specific SOTA absent in prior CASIAv2/CIFAKE binary approaches. Future work targets Deepfake enhancement (>95%) via frequency augmentation and multimodal text integration.

The increasing fake number accounts and bots on Twitter threaten cyber security, information integrity, and public discourse. modern methods to define and block bots\', primarily depend on user information metadata, content, and behavioral heuristics, have confirmed inactive against more developed bots that simulate actual user behavior. Graph Neural Networks (GNNs) can help. User interaction and Social media can be modeled as a graph, where users are represented as nodes and their interactions as edges, and GNNs can mesh individual and relational data into a single entity. This paper analyzes these accounts from two various methodological viewpoints looked at Conventional Neural Networks (CNNs) and Relational Graph Convolutional Networks (R-GCNs). While convolutional neural network (CNN) models have shown high efficiency in recognizing linguistic indicators and metadata, their reliance on surface features has limited their ability to handle computations designed to accurately simulate human texts and metadata. The R-GCN model architecture is the most outstanding, outperforming all models particularly in terms of F1 and ROC-AUC, by successfully capturing different relationship patterns—follow, rewet and mention. this paper highlights the quality and limitations of current detection schemes and proposes modern methods for automatic interpretation, multimedia integration, and cross-platform adaptation.

One of the main causes of morbidity and death is still cardiovascular diseases (CVDs) thereby stressing the need for developing effective early diagnostic tools. This paper proposes a hybrid approach using AI that combines deep ECG feature learning with clinical data to enhance early heart disease diagnosis. The ECG signals were preprocessed, divided into 2-second epochs, and then represented as 128-dimensional embedding using a 1D CNN-based feature extractor. At the same time, a set of 13 clinical risk factors was normalized and used to construct a lightweight clinical prediction branch. Both were then fused using a multimodal neural network that learns both ECG and clinical features. The findings of the experiment have demonstrated the effectiveness of the proposed hybrid framework, which has achieved 97.56% accuracy, 0.996 AUC, 99.0% sensitivity, and 96.0% specificity, significantly outperforming the individual models. Furthermore, the SHAP values-based explain ability method has assisted in revealing the most important clinical variables and ECG embedding features, thereby making the prediction results interpretable and meaningful from the clinical perspective. In summary, the proposed approach has shown how multimodal AI systems might help with early heart disease diagnosis and offers a promising solution for future real-time applications.

Machine learning (ML) and deep learning (DL) have advanced at a very rapid pace; thus, improving image classification in many fields. MNIST data is one of the most popular benchmark datasets to compare classification algorithms and it is represented by 70,000 grayscale images of handwritten digits (09). This paper will make a comparative analysis of a number of machine learning and deep learning algorithms and models, namely Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM) and Convolutional Neural Networks (CNN) using the MNIST dataset. All of the models were trained and tested using the same preprocessing methods, including normalization and feature scaling. Several evaluation metrics were applied to evaluate the performance, and they included accuracy, precision, recall, F1-score, and confusion matrix. The findings prove that conventional machine learning models, especially SVM and Random Forest, perform competitively. Nonetheless, CNN is more effective as compared to classical models because it is capable of learning hierarchical features of space automatically using image data. The results indicate the significance of choosing suitable models depending upon the nature of data and computing limitations. This paper presents a comparative study in detail and in a systematic manner that would improve the comprehension of the trade-offs between the model performance and computational efficiency and could act as a valuable source of information to researchers and practitioners in image classification exercises.

Cyber-attacks are advancing toward secretive behaviors that avoid classical signature detection systems and single-metric abnormality approaches. This paper is used to be proposed a framework-based Hybrid Entropy–Complexity Learning (HECL) for strong detection of cyber-attacks based on networking information models. The proposed approach cooperatively analyzes the entropy-based Shannon scheme, statistical complexity, and learning adaptability for characterizing normal and abnormal traffic activities. Different from conventional systems-based detection entropy, the proposed method utilizes the temporal coupling between both the entropy and complexity, allowing the correspondence of low-rate and whitewash attacks. Comprehensive simulations illustrate that the proposed model performs outstanding detection accuracy, speedy convergence, and a lower rate-based false-alarm when it is compared to the standalone entropy and approaches-based Machine Learning (ML). The key results emphasize that integrating an information-theoretic system with the learning-based indicators gives an influential and lightweight security technique appropriate for recent information security systems. The experimental results on the CICIDS2017 dataset show that it outperforms the state of the art deep learning techniques with 98.85% accuracy and 0.95% false positive rate.

Cyberattacks are becoming more advanced, challenging the traditional defenses based on signature-based and rule-based approaches. The current literature on generative artificial intelligence (GenAI) in the context of cybersecurity is disjointed based on model family or area of application, and it does not combine technical performance indicators with practical implementation limitations, making a comprehensive picture impossible. To fill this gap, the study conducts a comprehensive review of peer-reviewed articles and conference papers (2021-2026) indexed in Scopus and ScienceDirect, and Springer databases, involving the use of GenAI as a defensive cybersecurity tool. These papers are divided into (I) GenAI model family: generative adversarial networks (GAN), large language models (LLM), variational autoencoders (VAE), diffusion models, and hybrid GenAI; (II) application domain: intrusion detection, malware detection, anomaly detection, threat intelligence, privacy-preservation, vulnerability detection, and phishing and spamming detection; and (III) defense strategy: reactive, proactive, and adversarial. GenAI typically increases the accuracy of detection and data efficiency and provides active defense. Nevertheless, the practical validation is usually limited to offline tests that apply imprecise metrics. The paper provides the performance–efficacy trade-off model, which relates technical standards and realistic implementation limitations. It also identifies a research roadmap in the future focused on creating autonomous, privacy-protective, and trustful GenAI-powered cyber defenses and suggests a living review platform to keep track of advances in the fast-changing area.

The recent rise in the use of Internet of Things (IoT) devices has further increased the pressure on cryptographic solutions that will meet the challenge of high-security standards, yet nearly meet the extreme resource requirements imposed by embedded computers by nature. The paper introduces a new hybrid encryption architecture that is aimed at meeting these conflicting requirements of the IoT setting. The scheme suggested is a variant of AES-r cipher combined with the ASCON authenticated-encryption primitive giving the AESASCON algorithm, and extended with integrity verification by the SHA3-256 hash to create the AESASCONH construction. Stochastic key material is calculated using a four-dimensional NoseSprott-Jerk-Rossler (4D-NSJR) chaotic system, which is highly sensitive to initial conditions causing the effective key space to become large and making brute-force enumeration challenging. The architecture is segmented into three layers that collaborate with one another and they are chaotic key generation, symmetric data encryption and authenticated-integrity verification. General testing against the complete NIST SP 800-22 statistical test is done that indicates that all the 15 randomness tests have been passed. Experiments with Five data sizes are used in experiments to demonstrate speedups ranging from 42% for large data payloads (1.5 MB) to 175% for small data sizes compared to AES. This demonstrates that the suggested approach provides efficient and secure protection for IoT. and entropy values of data of up to about the theoretical maximum of 8.0 bits per symbol, and thus confirm speed improvements as well as cryptographic strength is available to support high-assurance IoT deployments.

Abstract It is necessary in various real-time applications such as teleconferencing, live streaming, and remote monitoring. However, applying standard encryption methods on high resolution video streams can cause computational delay. We propose a scalable video encryption framework, based on AES-256 and multi-threaded parallel processing to boost the performance of cryptography without compromising security in the paper. Video frames are sliced into independent chunks and assigned to available CPU cores so that encryption can occur in parallel. In order to recover frame ordering correctly, we use a synchronization mechanism. Extensive experimental evaluation on Full HD video streams shows that the proposed framework improves throughput and results in a four-times reduction in encryption time when compared to sequential execution. Using eight threads with small CPU utilization efficiency, the system achieved a speed up as high as 4.62×. These results show that parallel software encryption on CPU is a realistic method for secure real-time video applications

The rapid development of face generation technologies in recent years has enabled the creation of stunningly realistic human face images, effectively blurring the lines between authentic and synthetic content. While these technologies offer positive applications in education, entertainment, and digital art, their potential for misuse—particularly on social media—is profound. The proliferation of deep-fakes poses a significant threat to digital trust, facilitating electronic fraud and identity manipulation, which necessitates the development of effective detection solutions. This paper addresses this escalating threat by proposing an adaptive detection framework based on pre-trained Convolutional Neural Networks (CNNs). Unlike traditional methods that struggle with unseen data and environmental noise, our approach optimizes ResNet50, DenseNet201, and Alex Net architectures through customized data preprocessing and hyper parameter tuning. Evaluated on the rvf10k dataset—selected for its diverse illumination and ethnic representation—the proposed model achieves a peak accuracy of 99.89%. Furthermore, this work provides a comprehensive analysis of computational cost and inference speed, demonstrating robustness against over-fitting via cross-validation techniques. The results confirm that the proposed framework significantly enhances detection reliability and contributes to securing digital content integrity.

This paper explores how e-governance principles and Machine Learning (ML) technologies can be integrated to support participatory design in educational systems, and, in particular, learning the programming language curriculum with programming language curriculum reform. Based on the socio-technical systems theory of Mumford and the ETHICS methodology, we are able to propose a new framework, which will utilize three supervised ML classifiers, including Support Vector machine (SVM), Random Forest (RF) and Decision Tree (DT), to systematically analyze the data of teacher questionnaires collected via e-governance channels. The data set included the answers of 94 teachers in Baghdad, Iraq, and was used to categorize the needs of instructional improvement into six categories: Hardware Development (HD), Demonstration Tools Development (DDT), Teacher Experience Development (DTE), Redistribution of Teacher Sessions (RDTS), Syllabus Development (SD) and Student Ability Development (SAD). The results of the experiment show that SVM attains the maximum classification accuracy of 99.90% in training and 98.00% in testing is supported by the consistent performance across all six classes and the minimal train-test gap (≈1.9 percentage points), indicating effective generalization rather than data leakage or overfitting. The analysis of the results of questionnaires has shown that the most pressing areas of improvement are Syllabus Development (85%), and Student Ability Development (75%). This paper highlights that an active participatory, socio-technical approach to e-governance, one that actively engages educators in decision making, can go a long way in improving the quality and relevance of the educational policy outcomes.