Automated Generation and Compilation of Fuzz Driver Based on Large Language Models

Fuzz drivers are essential components of library fuzz testing, yet automatically generating correct and robust fuzz drivers and executing them is challenging. The predominant approach to fuzz testing libraries and their function interfaces involves security experts manually writing test drivers for fuzz testers. In contrast, generation based on LLMs (Large Language Models) is a promising direction, as it can operate with lower requirements on consumer programs, utilizes API (Application Programming Interface) usage information across multiple dimensions, and generates user-friendly output code. However, there is currently no fully automated method for driver generation and compilation. To address this issue, this paper has designed an automated method for generating and compiling drivers, and has evaluated the quality of the drivers it produces. Evaluation results indicate that drivers generated by large language models perform nearly as well in coverage as those written manually, and the generated compilation commands achieve an accuracy of 75%.