University of Twente Student Theses


Attribute Extraction from Darknet Markets and the Applicability of Transfer Learning

Oonk, N.M.J. (2018) Attribute Extraction from Darknet Markets and the Applicability of Transfer Learning.

[img] PDF
Abstract:Cybercrime is an international threat that is continuously growing. One of the fundamental strategies in the combat against cybercrime is intelligence-led policing. Darknet markets (DNMs) are a rich source of information on cybercrime. However, gaining in depth insight into the products and services offered on DNMs is difficult. Whereas studies have been conducted to gain insight into DNMs, no study has yet focussed on extracting information from DNM posts to gain insight in the techniques that are exploited, the locations that are mentioned and other characteristics law enforcement agencies are interested in. In this study, attribute extraction was performed to automatically extract attributes from bulletproof hosting advertisements of a multi-lingual DNM. A variety of attribute extraction techniques was tested and compared. The extraction techniques included are dictionary-based, rule- and pattern-based and machine learning based attribute extraction. The machine learning algorithms that were compared are Maximum Entropy Model (MEM), Conditional Random Field (CRF) and bidirectional Long-Short Term Memory with Conditional Random Field (biLSTM-CRF). Furthermore, transfer learning by fine-tuning was applied to improve the performance of the biLSTM-CRF extractor. Russian attribute extraction was performed with a micro-averaged F2-score of 0.886, using the biLSTM-CRF extractor developed with transfer learning. The ensemble extractor, that combines for each attribute the best extraction technique, resulted in a micro-averaged F2-score of 0.859 for English attribute extraction. Transfer learning applied to the biLSTM-CRF extractor resulted in a performance improvement of 0.014 and 0.002 micro-averaged F2-score for Russian and English respectively. With attributes that improved with up to 0.074 F2-score. These results show that attribute extraction and transfer learning have a great potential to gain insight in DNMs and effectively include DNMs in intelligence-led policing.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page