Embedding-Based Entity Alignment For Vulnerability Knowledge Graphs

The quality and integrity of knowledge graphs affect their downstream applications,and merging knowledge graphs from different sources can help expand them.Entity alignment techniques identify and fuse entity pairs that represent the same facts from multiple heterogeneous knowledge graphs,which can enhance graph integrity and reduce data redundancy.Despite years of research on such techniques,there is still a gap in entity alignment for highly relevant graphs in domains like cyberspace security.Vulnerability knowledge graph(Vul KG)is a type of knowledge graph in cyber security.It allows researchers to query and analyze vulnerabilities based on textual and structural features,and to support automated attack and defense scenarios by detecting potential security risks.To build and improve Vul KGs,knowledge graphs based on different vulnerability repositories need to be merged to increase knowledge completeness.Entity alignment is a key step in this merging process.However,entity alignment in Vul KGs faces some unique challenges:(1)The knowledge graphs have many entities but few types of relationships.If generic models are used to learn entity features,there will be many false matches between entities with similar feature representations.Therefore,a knowledge graph embedding method that is tailored to the Vul KG needs to be designed to generate suitable entity vector representations.(2)The entity alignment task needs to handle multiple types of heterogeneity in the Vul KGs,such as multilingualism,text heterogeneity,structural heterogeneity,and missing data.This requires a unified alignment model that can incorporate multiple dimensions and views.This paper focuses on the aligning entities in Vul KGs and presents the following main contributions.(1)To solve the problem of embedding vulnerability knowledge graphs,this paper proposes a knowledge graph embedding method based on potential relationship mining.It can identify blurry entities in the graph that have potential missing relationships without using any additional information and use this information to enhance the quality of graph embedding.Experiments show that our method learns a more reasonable graph embedding space,improves the accuracy of link prediction and can be applied to general graphs as well.(2)This paper proposes a multi-view interactive entity alignment framework,TG-INT,for the entity alignment task.It uses multi-dimensional information to perform the entity alignment task of vulnerability graphs in a multi-view way.The experimental results show that our framework not only improves the accuracy of entity alignment for vulnerability knowledge graphs but also performs better on the general knowledge graph.(3)This paper has implemented an entity alignment system for Vul KGs based on these techniques.The system can import different Vul KGs,automate the entity alignment task and output a more complete vulnerability knowledge graph after entity alignment.