Skip to content

Latest commit

 

History

History
82 lines (64 loc) · 6.17 KB

README.md

File metadata and controls

82 lines (64 loc) · 6.17 KB
Raw

Binary Function Similarity

This repository contains the code, the dataset and additional technical information for our USENIX Security '22 paper:

Andrea Marcelli, Mariano Graziano, Xabier Ugarte-Pedrero, Yanick Fratantonio, Mohamad Mansouri, Davide Balzarotti. How Machine Learning Is Solving the Binary Function Similarity Problem. USENIX Security '22.

The paper is available at this link.

Additional technical information

The technical report, with additional information on the dataset and the selected approaches, is available at this link.

Artifacts

The repository is structured in the following way:

  • Binaries: the compiled binaries and the scripts to compile them. Binaries are downloaded from GDrive via a Python script
  • IDBs: where the IDA Pro databases (IDBs) are stored after analysis. IDBs are generated via a Python script and IDA Pro
  • DBs: the datasets of selected functions, the corresponding features, and the scripts to generate them
  • IDA_scripts: the IDA Pro scripts used for the features extraction
  • Models: the code for the approaches we tested
  • Results: the results of our experiments on all the test cases and the code to extract the different metrics.

What to do next?

The following is a list of the main steps to follow based on the most common use cases:

  • Reproduce the experiments presented in the paper

    • Note: the binaries (Binaries) and the corresponding IDA Pro Databases (IDBs) are only needed to create a new dataset or to extract additional features. In order to reproduce the experiments or run new tests with the current set of features, DBs and Models already contain the required data.
    1. The DBs folder contains the input data needed to reproduce the results for each tested approach, including extracted features
    2. Refer to the README of each approach in the Models folder for detailed instructions on how to run it
    3. Follow the README and use the scripts in the Results folder to collect the different metrics.

  • Test a new approach on our datasets

    1. Check the README in the DBs folder to decide which data to use based on each test case
    2. Reuse the existing IDA Pro scripts codebase for the features extractions and pre/post-processing code to minimize evaluation differences
    3. Follow the README and use the scripts in the Results folder to collect the different metrics.

  • Use one of the existing approaches to infer new functions

    • Note: the current workflow and code has been written to optimize the evaluation of the similarity engines on a "fixed" dataset of functions and their features. This makes the inference on a new dataset slightly complex, as it requires to follow different steps for each approach. A simplification may be addressed in a future release.
    1. Refer to the README of each approach in Models for detailed instructions on how to run it in inference mode
    2. Use the corresponding IDA Pro script to extract the features that are needed by that specific approach
    3. Some approaches require to run a specific post-processing script to convert the extracted features into the requested format
    4. Be aware of the limitations of the ML models: new architectures, compilers and compiler options may require retraining them.

How to cite our work

Please use the following BibTeX:

@inproceedings {280046,
author = {Andrea Marcelli and Mariano Graziano and Xabier Ugarte-Pedrero and Yanick Fratantonio and Mohamad Mansouri and Davide Balzarotti},
title = {How Machine Learning Is Solving the Binary Function Similarity Problem},
booktitle = {31st USENIX Security Symposium (USENIX Security 22)},
year = {2022},
isbn = {978-1-939133-31-1},
address = {Boston, MA},
pages = {2099--2116},
url = {https://www.usenix.org/conference/usenixsecurity22/presentation/marcelli},
publisher = {USENIX Association},
month = aug,
}

Errata corrects

Our corrections to the published paper:

  • From Section 3.2 Selected Approaches: "First, the binary diffing tools grouped in the middle box [13,16,83] have all been designed for a direct comparison of two binaries (e.g., they use the call graph) and they are all mono-architecture." This sentence is inaccurate because Bindiff and Diaphora also support the cross-architecture comparisons.
  • The two plots in Figure 2 display the results from Dataset-2. The left one corresponds to the XO task, while the right one corresponds to the XA + XO task.

License

The code in this repository is licensed under the MIT License, however some models and scripts depend on or pull in code that have different licenses.

Bugs and feedback

For help or issues, please submit a GitHub issue.