Fine-Grained Vulnerability Location Based on LLVM

Locating the vulnerabilities of software is the basis for developers to patch vulnerabilities. Currently, most static detection tools detect code vulnerabilities at the function and file level, and such detection granularity is not suitable for accurately locating vulnerabilities. The mainstream deep learning based vulnerability detection methods slice the source code at the function level, and this slicing method is implemented at coarse granularity, so it cannot support the task of accurately locating vulnerabilities. Constructing fine-grained program slicing algorithms and deep learning models that can learn fine-grained vulnerability features are the difficulties of vulnerability localization technology. Therefore, we propose a fine-grained source code vulnerability location method, dubbed VuIR. VuIR is designed for detecting C/C++ source code vulnerabilities, while accurately locating the vulnerabilities. Firstly, we extract program slices containing four types of vulnerability characteristics from the Abstract Syntax Tree (AST) of the source code. Secondly, we slice the program slices based on LLVM again, and mark the locations of vulnerabilities. Finally, we design a deep learning model to mine the vulnerability characteristics in the source code.