In recent years,taking the known vulnerable function as the criteria to retrieve the similar implementation has been proven to be an effective vulnerabilities detection method.However,a vulnerable function often contains some statements that are irrelevant to the vulnerability of interest,which may heavily interfere with the similarity computation and lead to false positives and false negatives.This paper presents an approach to improve the precision of the retrieval-based vulnerabilities detection by leveraging the patch of the vulnerable function.The program slicing technique is adopted to exclude irrelevant statements from the original vulnerable function according to the patch.A denoised feature vector is generated from the obtained slice and is used to search the potential unknown vulnerabilities in the code base.This approach has been applied to some real-world projects.Experimental results show that the approach can effectively reduce the interference of irrelevant statements and improve the detection precision.Three confirmed unknown vulnerabilities are successfully detected from the projects.