Detecting Fraudulent URLs on Arabic Social Platforms Using an Explainable Machine Learning Approach

Online advertisements are increasingly used to spread phishing URLs on Arabic social media platforms. This study aims to detect malicious URLs targeting Arabic social media platforms by building the Arabic Social Media Phishing URL Dataset, which contains 1,600 URLs, where 800 are malicious URLs and the rest are legitimate. Moreover, this research compares the performance of machine learning models, namely Logistic Regression, Random Forest, Gradient Boosting, XGBoost, and SVM, with a Deep Neural Network model. Furthermore, this research evaluates the decisions of the proposed model using the SHAP method. This study develops an explainable XGBoost model capable of detecting malicious URLs used in online advertisements on Arabic social platforms, achieving an accuracy of 96.57% and an F1-score of 96.55%. The results show that some features, such as uncommon top-level domains, increase the probability of classifying a URL as malicious. These findings provide valuable insights that can help users better identify malicious URLs, thereby enhancing their awareness and protection against phishing attacks.