Reliable KMDF Instruction Set Generation for User Queries Based on Endogenous Generative LLMs

This paper proposes a method for using endogenous generative large language models (LLMs) to generate reliable Microsoft Kernel Mode Driver Framework (KMDF) instruction sets for user queries in the natural language form. The reliability is defined to ensure the validity, correctness, and safety of the generated instruction sets through a multi-phase verification procedure. Three phases are designed in the verification pro-cedure, leveraging both traditional and advanced AI techniques to minimize the risk of errors and enhance the security and effi-ciency of the framework-based anti-virus programs. Meanwhile, m = 20 is figured out to be the maximum count of attempts, where $m$ is the maximum count of the attempts. With roles and layers shown clearly, security proofs like secure multi-party computation are discussed for local communication security. The experimental results demonstrate the effectiveness of our approach, showing significant improvements in the reliability of KMDF instruction set generation. Eventually, the paper presents some possible future research directions, showing its portability and extendability,