Enhancing Reverse Engineering: Investigating and Benchmarking Large Language Models for Vulnerability Analysis in Decompiled Binaries

This repository contains the DeBinVul dataset, source code, and experimental results of the paper Enhancing Reverse Engineering: Investigating and Benchmarking Large Language Models for Vulnerability Analysis in Decompiled Binaries.

DeBinVul Dataset

The dataset is a compressed CSV file stored as DeBinVul.zip. DeBinVul has 12 columns:

1. source_code: The complete source code function definition.
2. comments: Extracted comments from the function's source code.
3. label: The assigned CWE label (or none if benign).
4. file: The file containing the function.
5. function: The name of the function.
6. decompiled_code: Ghidra decompiled code of the function.
7. compiler_options: Compiler command used for compiling the function.
8. stripped_decompiled_code: Ghidra decompiled code with symbols stripped.
9. description: Description of the vulnerability (if applicable) and the function’s behavior.
10. prompt: The complete prompt given to the model for the task.
11. instruction: Specific instructions given to the model for the task.
12. task: The nature of the task, such as identification, classification, description, or function name recovery.

Other Repository Contents

All code related to training, inference, and evaluation can be found in the src subdirectory. The experimental results can be found under the Experimental_results subdirectory. Miscellaneous scripts associated with the dataset are located in the scripts directory.