binhuntertool

This is an implementation of the paper BinHunter: A Fine-Grained Graph Representation for Localizing Vulnerabilities in Binary Executables. BinHunter is a tool designed to find and localize various types of vulnerabilities within binary programs. It leverages a Graph Convolutional Neural Network (GCN) to train on vulnerabilities. For each type of vulnerability, we train a separate binary classifier. During the training phase, BinHunter trains on a subgraph of the Program Dependence Graph (PDG) of a function that locates vulnerability and patch patterns extracted using debug symbols (DWARF) in the binary. In the testing phase, however, it utilizes calls to external functions to slice the PDG.

Code Structure

BinHunter/
|
├── run_binhunter.py            # a module to run binhunter
├── binhunter_graph/            
│   ├── graphs/                 # This directory contains everything related to graph creation 
│   │   ├── cdg.py              # Control dependence Graph creation
│   │   ├── pdg.py              # Program Dependency Graph creation
│   │   └── ddg.py              # Data Dependency Graph creation
│   └── slicing/                
│       └── debug.py            # Provide all functions to extract dwarf information
│
└── binhunter_train/            # Training modules
    └── gcn.py                  # Graph Convolutional Network implementation

binhunter_graph directory contains all modules to create program dependence graph and graph representation and slicing the PDG. binhunter_train directory contains all modules to train and test using GCN. Note that the neural_models directory is related to kipf model.

requirements

As mentioned in our paper, we performed all experiments on a Linux server with 256 GB RAM and an Intel XeonE5-1650 CPU with 12 cores.

Installation

To install all dependencies, first download the Dockerfile from https://drive.google.com/drive/folders/1MrdlnvWPjTscoEviZRabamV8UVUmuTUp?usp=sharing.

Then create a directory called bindocker and put the dockerfile inside this directory.

mkdir bindocker

Then change your directory to bindocker.

cd bindocker

Now build the docker using this command.

docker build -t binhunter .

When the docker is built successfully, then run the docker using the command below.

docker run -it binhunter

you are now inside the docker. if you type whoami you should see binhuntuser. Now run the commands below to install pip.

sudo apt update
sudo apt install python3-pip

First give the permission to binhunt user to write and then clone the binhunter repo.

sudo chmod 777 .

git clone /~https://github.com/SimaArasteh/binhuntertool.git

Now change directory to binhuntertool.

cd binhuntertool

Now install all dependencies and requirements using following command.

pip3 install -r requirements.txt

When all dependencies installed successfully, download the data from https://drive.google.com/drive/folders/1MrdlnvWPjTscoEviZRabamV8UVUmuTUp?usp=sharing. Then copy the data.zip inside the docker. In your local machine find the container id of binhunter.

docker ps
docker cp data.zip container-id:/binhuntertool

After copying data into the docker, unzip the data.zip.

unzip data.zip -d .

usage

Now it is time to run binhunter. You can run binhunter for each type of vulnerability using following command.

python3 run_binhunter.py CWE121 

Note that for each round of running binhunter, we shuffle the training data and each time you run, you get different results but the results should be close to Table3.