Learning Maliciousness in Cybersecurity Graphs

Abstract

Statistical relational learning is concerned with inferring patterns from data explicitly modeled as graphs. In this work, we present an approach to learning latent topological and attribute features of multi-relational property graphs in settings where a fraction of node attributes are missing. This work draws upon prior work based on tensor factorization. We demonstrate how learned latent embeddings can be used to approximate the missing attributes. The methods explored are applied to the problem of detecting malicious entities in a novel cybersecurity ontology in which emails are explicitly modeled as graphs.