Enhancing DarkWebActivities Classification Using Embedding Methods

The Dark Web is frequently associated with illegal operations, including drug trafficking, arms dealing, cybercrime, and hacking-related services. Mitigating these threats remains a major challenge for cybersecurity professionals and law enforcement agencies, given the complexity and evolving nature of cybercriminal activities. However, the recent progress in Artificial Intelligence and machine learning has introduced new approaches to analyse the Dark Web content and activities. This study proposes a novel classification approach capable of identifying ten types of illicit activities. The system is built upon the publicly accessible DUTA dataset, which offers a structured corpus for such analysis. To classify the Dark Web activities, ensemble learning algorithms were employed, including Random Forest, Light Gradient Boosting Machine, Extreme Gradient Boosting, Gradient Boosting, and CatBoost. To convert text-based data into numerical vector formats suitable for computational processing, five distinct word embedding techniques were utilized. The results of this study show that the combination of the RoBERTa embedding with the XGB classifier achieved promising results, with a precision of 99.98%. These results underscore the capability of the proposed methodology to detect and categorize illicit Dark Web activities.