Multi-label Positive and Unlabeled Learning and its Application to Common Vulnerabilities and Exposure Categorization

Abstract

The widely adopted Common Weakness Enumeration (CWE), which stores and manages software and hardware vulnerability reports known as Common Vulnerabilities and Exposures (CVE) in a hierarchical structure, provides common baseline standard for weakness identification, mitigation, and prevention efforts. In this paper, we propose a machine-learning based method to assign pertinent CWE identifiers to new CVE entries. The proposed method formulates the task as a multi-label classification problem and exploits positive and unlabeled learning to address the lack of multi-labelled samples in learning. In evaluations, the proposed method demonstrated preferable performance compared to traditional multi-label classifiers. In particular, case studies demonstrated that multiple CWE iden-tifiers assigned to CVE entries carry essential information that can benefit security practices.