Predicting Temporal Exploitation of Software Vulnerabilities Using Knowledge-Fused Neural Learning

Abstract

Despite over 25,000 Common Vulnerabilities and Exposures (CVEs) being disclosed annually, fewer than 5% are exploited in real attacks, making vulnerability prioritization a persistent challenge. Existing methods, including the Common Vulnerability Scoring System (CVSS) and Exploit Prediction Scoring System (EPSS), treat exploitation prediction as binary classification, discarding the temporal dimension that determines whether a patch needs deployment today, next month, or later. This study presents a knowledge-fused neural framework that reformulates this as a five-class temporal classification: 0–7 days, 8–30 days, 31–90 days, 91–365 days, and never exploited. The framework constructs a cybersecurity knowledge graph that integrates 1,003 Common Weakness Enumeration (CWE) definitions and 559 Common Attack Pattern Enumeration and Classification (CAPEC) attack patterns, with 5,070 relationships. Graph embeddings from Node2Vec are fused with Bidirectional Encoder Representations from Transformers (BERT) vulnerability embeddings via cross-attention. Evaluated on 102,000+ CVEs (2002–2024), the framework achieves 78.4% temporal classification accuracy, 74.7% exploitation F1-score, and 89.1% Area Under the Receiver Operating Characteristic curve (AUC-ROC)—outperforming EPSS by 6.8 F1 points. Ablation studies confirm that knowledge graph integration with cross-attention fusion is the critical architectural contribution.