Classifying Web Exploits with Topic Modeling

This short empirical paper investigates how well topic modeling and database meta-data characteristics can classify web and other proof-of-concept (PoC) exploits for publicly disclosed software vulnerabilities. By using a dataset comprised of over 36 thousand PoC exploits, near a 0.9 accuracy rate is obtained in the empirical experiment. Text mining and topic modeling are a significant boost factor behind this classification performance. In addition to these empirical results, the paper contributes to the research tradition of enhancing software vulnerability information with text mining, providing also a few scholarly observations about the potential for semi-automatic classification of exploits in the existing tracking infrastructures.