GraphBinMatch: Graph-based Similarity Learning for Cross-Language Binary and Source Code Matching

TL;DR

GraphBinMatch introduces a graph neural network approach for cross-language binary and source code matching, significantly improving accuracy over existing methods in various code matching tasks.

Contribution

It presents a novel graph neural network-based method for cross-language binary-source code matching, addressing a key challenge in multi-language software analysis.

Findings

Outperforms state-of-the-art by up to 15% F1 score

Effective in cross-language binary-to-source matching

Also performs well in source-to-source matching

Abstract

Matching binary to source code and vice versa has various applications in different fields, such as computer security, software engineering, and reverse engineering. Even though there exist methods that try to match source code with binary code to accelerate the reverse engineering process, most of them are designed to focus on one programming language. However, in real life, programs are developed using different programming languages depending on their requirements. Thus, cross-language binary-to-source code matching has recently gained more attention. Nonetheless, the existing approaches still struggle to have precise predictions due to the inherent difficulties when the problem of matching binary code and source code needs to be addressed across programming languages. In this paper, we address the problem of cross-language binary source code matching. We propose GraphBinMatch, an approach based on a graph neural network that learns the similarity between binary and source codes. We evaluate GraphBinMatch on several tasks, such as cross-language binary-to-source code matching and cross-language source-to-source matching. We also evaluate our approach performance on single-language binary-to-source code matching. Experimental results show that GraphBinMatch outperforms state-of-the-art significantly, with improvements as high as 15% over the F1 score.