Answers are grounded in this paper and cite their sources.
Knowledge graph embedding (KGE) is widely used to complete cyber threat intelligence (CTI) graphs, yet most methods rely on graph structure alone and are validated on general-purpose benchmarks that differ sharply from security data. CTI graphs have a small relation vocabulary, and their entities carry much of their meaning in technical text. This study applies KGE-SymCL, a contrastive learning approach that considers the symmetry of relations, to a cyber threat intelligence (CTI) graph derived from MITRE ATT&CK Enterprise v19.1. The graph comprises 2,595 entities, 20,866 triples, and five distinct relation types. We propose an R-GCN encoder whose structural embeddings are merged with SecureBERT text embeddings through a learned gate. Fifteen KGE backbones are compared, with and without the SymCL term, under an identical protocol over three random seeds. SymCL improves the purely structural R-GCN (MRR 0.0904 to 0.0977) but adds little once SecureBERT fusion is present. The fused model reaches an MRR of 0.1102, a significant gain over plain R-GCN (p = 0.015), although CompGCN and SimKGC remain higher on Hit@1 and Hit@3. These results show that the benefit of contrastive alignment depends on what other signal the encoder already has. To support forensic use, we also add multi-hop path explanations, fusion-gate analysis and tamper-robustness tests that check how predictions behave when the graph is perturbed.
No takes yet. Share an insight, caveat, or question.
Pranitha et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: