Deep Learning-Based Identification of Illicit Activities in Dark Web Ecosystems
Abstract
This paper proposes a deep learning-based framework for detection of illicit activities on Dark Web based on semantic analysis of content as well as detection at ecosystem level using relational modelling. Transformer-based language representations are used to capture complex and obfuscated textual patterns, and graph-based reason based on relationships between vendors, listings, forums and extracted entities. To overcome the shortage of labelled data and class imbalance, weak supervision and active learning are included in the framework, allowing it to keep adapting to emerging illicit trends. An explainability layer will deliver to users’ case ready evidence through text highlights and relational subgraphs to support investigative transparency. Experimental evaluation shows that the proposed approach can obviously improve the detection accuracy and decrease the false positive rate as well as improve the operation efficiency compared with conventional methods. The results show the effectiveness of deep learning-driven, context-aware analysis as a scalable and ethical solution for Dark Web illicit activities detection.