Enhancing Reverse Engineering: Investigating and Benchmarking Large Language Models for Vulnerability Analysis in Decompiled Binaries
The paper "Enhancing Reverse Engineering: Investigating and Benchmarking Large Language Models for Vulnerability Analysis in Decompiled Binaries" addresses the challenges of identifying security vulnerabilities in decompiled binary code, especially when source code is unavailable.
The authors introduce DeBinVul, a comprehensive dataset comprising 150,872 samples of vulnerable and non-vulnerable decompiled binary code, focusing on C/C++ languages due to their prevalence in critical infrastructure and associated vulnerabilities.
By fine-tuning state-of-the-art Large Language Models (LLMs) such as CodeLlama, Llama3, and CodeGen2 with DeBinVul, the study reports performance improvements of 19%, 24%, and 21% respectively in detecting binary code vulnerabilities.
Additionally, the models achieved high performance (80-90%) in vulnerability classification tasks and showed enhanced capabilities in function name recovery and vulnerability description.
This work underscores the importance of specialized datasets in enhancing the effectiveness of LLMs for security analysis in decompiled binaries
https://arxiv.org/abs/2411.04981
Comments
Post a Comment