PageLLM: Incremental approach for updating a Security Knowledge Graph by using Page ranking and Large language model

Chinmaya Mishra, Himangshu Sarma, Saravanan M.

Abstract: Due to increase in cyber crime and evolution of sophisticated tools and techniques, Threat Intelligence plays a critical role. It helps defenders to stay ahead of attackers by developing the right defense mechanism to invade those attacks. In this regards security knowledge graph plays a critical role which can be used to signify complex entities and their relationship in a graphical structure. Further projecting those entities and relationships in to the lower dimension using several embedding techniques such as TransE help in many down streaming task. The learned embedding can be used to predict new cyber threat which is very helpful for defenders to stay alert and develop necessary weapons to stay ahead of an attack. One of the major challenge security knowledge graph has its dynamic nature of changing intelligence. Active learning can be used to only update the substantial portion of embedding rather than retraining the knowledge graph from scratch which has higher time and space complexity. Also given the rise in generative AI and large language models which are super rich in context, there is a scope of utilizing those for building a robust and good quality security knowledge graph. We will discuss a novel methodology called PageLLM which utilizes page ranking and LLMs to enable active learning in an incremental way and will improve the quality of knowledge graph through enriched context.