LLVM-based vulnerability fine-grained positioning method

The invention relates to a vulnerability fine-grained positioning method based on LLVM. The method comprises the following steps: (1) constructing an abstract syntax tree of a source code: generating the abstract syntax tree of the source code by using a front-end tool Clang based on LLVM; four types of vulnerability code features are defined respectively; (2) feature extraction; (3) model training: i, obtaining a sensitive intermediate code snippet, and generating IRVCs; ii, performing data vectorization representation based on a word embedding model; iii, training a vulnerability fine-grained positioning model: taking the vectorized representation of the obtained fine-grained program slice IRVCs containing vulnerability position information as the input of a BGRU model, establishing a BGRU-based vulnerability fine-grained positioning model, adding a fusion layer into the model, sequentially adding a K-max pooling layer and an average pooling layer behind the fusion layer to jointly realize granularity refin