Cyber Threat Analysis Through Densified Causal Knowledge Graph

This paper presents a novel framework to construct a dense, informative causal knowledge graph from a massive corpora of news articles leveraging both state-of-the-art transformers and traditional machine learning methods such as clustering. The proposed framework, specifically designed for modeling cause effect relationship through a graph-based data structure, consists of three steps, namely, i) cause-effect (CE) relation extraction, ii) argument clustering, and, iii) construction of causal knowledge graph (CKG). We thereafter perform experiments with 22,342 news articles crawled from the Internet with predetermined key words suggested from a centralized cyber security division within Panasonic Holdings. We demonstrate that the clustering-based strategy results in a dense CKG that conveys domain specific insightful information to the business user that would help in long and short term strategic and operational decision making. Finally, we design a metric to evaluate the performance of the CKG based on entropy motivated by information theory. We discuss the results and provide some future direction of research in this area.