A Novel Vulnerability Exploit Prediction System Using the Relational Vulnerability-Vendor Network

The number of disclosed vulnerabilities is increasing rapidly. Prioritizing vulnerabilities becomes essential as it is challenging to patch them all. The exploitability prediction is a crucial factor in determining the prioritization of vulnerabilities. The EPSS is a widely used data-driven vulnerability prioritization model. However, the literature offers limited insights into the model’s efficacy in accurately classifying exploited vulnerabilities. Many existing works on exploitability prediction focus on building models based on different features, neglecting the relationships between vulnerabilities and their impacted stakeholders. Only a few studies have explored graph-based techniques in exploitability prediction. In this study, we propose a relational vulnerability-vendor network and a heterogeneous relational graph convolution network model. It enables to incorporate the co-exploitation behavior of vulnerabilities in the exploitability prediction. We propose a set of complementary features that have found significance in exploitability prediction. Our proposed model achieves an overall performance gain over the EPSS. Using the connoted EPSS cut-off label, the proposed model obtains an improvement for highly exploitable vulnerabilities. Also, our proposed model has a lower false-negative prediction rate than the EPSS cut-off. We conducted a case study to evaluate the efficacy of our suggested model compared to EPSS, resulting in favorable new insights and conclusions.