Abstract:

Security vulnerabilities in network protocol software occur frequently and pose serious threats to cyberspace security. Gray-box protocol fuzzing tools, such as AFLNet, have improved vulnerability detection by introducing coverage feedback and state modeling mechanisms. However, constrained by a persistent “semantic barrier”, these tools struggle to comprehend protocol syntax structures and contextual logic, resulting in limited testing efficiency. In recent years, large language models have demonstrated exceptional generalization and comprehension capabilities in tasks such as semantic modeling, contextual reasoning, and code generation, providing a promising pathway to overcome this barrier. This paper proposed LPF (LLMProFuzz), a protocol fuzzing framework guided by large language models, which addressed the limitations of traditional methods from three perspectives: firstly, automatically extracting protocol syntax templates through few-shot prompt engineering; secondly, designing a seed enrichment mechanism based on historical vulnerability characteristics to generate high-value initial cases that cover boundary and exceptional scenarios; thirdly, introducing a structure-aware mutation location selection strategy to increase the proportion of effective test cases. Experimental results on representative protocol stacks, including HTTP, FTP, and RTSP, demonstrate that LPF significantly outperforms baseline tools such as AFLNet and StateAFL in terms of code coverage, state coverage, and testing efficiency.

Key words: large language models, network protocol, fuzzing, structure-aware mutation, prompt engineering