A Dual Perspective on Digital Threats: Comparing Machine Learning and Large Language Models for Arabic Cyberbullying and DDoS Detection رسالة ماجستير

Abstract

In today's digital landscape, addressing harmful online behavior is essential for maintaining secure and trustworthy platforms. The study focuses on detecting anomalies in Arabic cyberbullying comments from the X platform, recognizing the unique challenges posed by its diverse dialects and informal usage social media. Also finding anomalies to detect Distributed Denial of Service (DDoS) attacks. The research aims to evaluate how well different models handle these complexities, with anomaly detection playing a vital role in identifying unusual patterns that could signify threats. It provides a comprehensive comparative analysis of models including Logistic Regression (used as the baseline), Support Vector Machines (SVM), BERT, BRAD, and XLNet. The methodology involved preprocessing datasets, managing missing values, and splitting them into training and testing sets. Each model was trained and evaluated on GPUs, with performance metrics including accuracy, precision, recall, and F1 score used for assessment. The findings reveal variations in model performance. Starting with Logistic Regression (LR) as the baseline model, it achieved an accuracy of 80.6%, demonstrating a reliable foundation for comparison. Support Vector Machines (SVM) performed slightly better, with an accuracy of 82%, effectively balancing precision and recall. BERT, an advanced transformer model, outperformed all others with an accuracy of 85% and a precision of 97%, highlighting its robustness in identifying cyberbullying comments. BRAD exhibited perfect precision but a lower recall and F1 score, indicating its strength in identifying true positives while lacking in comprehensive detection. XLNet, however, struggled significantly, with zero precision, recall, and F1 score, reflecting its limited effectiveness in this context. The research further integrates Explainable AI (XAI) techniques to improve model interpretability, offering insights into decision-making processes and enhancing trust in automated systems. For DDoS detection, all models achieved perfect performance metrics, indicating their effectiveness in classifying attacks. In conclusion, this thesis provides a comparative analysis of LLMs and traditional machine learning models in detecting anomalies in Arabic cyberbullying and DDoS attacks. It contributes valuable insights for developing more effective detection systems, and advancing online safety for Arabic-speaking communities.