[Research purpose]Vulnerability information in threat intelligence refers to information about the presence of vulnerabilities in networks,systems,applications,or supply chains.Current search engines have shortcomings in vulnerability information retrieval sys-tems.Using the pre-training model to build a vulnerability retrieval system can improve the retrieval efficiency.[Research method]U-sing public vulnerability information as the data source,we construct a question answering dataset and incrementally pre-train Tiny Bert.The model is used to vectorize each query,the vulnerability information is built into a faiss vector database,and the HNSW index is used for multi-channel and single-channel recall retrieval.Then,the two-tower and single-tower models are generated by comparative learning fine-tun-ing,and a simple knowledge retrieval system is constructed by using two-tower recall and single-tower fine ranking.[Research conclusion]The experimental results show that the retrieval performance can be significantly improved by using the pre-trained model,and the TOP 1 re-call rate of the two-tower model fine-tuned by comparative learning is 92.17％in the constructed vulnerability information test set.Through the retrieval practice in the field of vulnerability information,it provides some reference for building the retrieval system of threat intelligence.
维普
万方数据