Code for master thesis "Using Large Language Models for Binary Decompilation"

Python code project for my master thesis which can be found here: Using Large Language Models for Binary Decompilation

Setup

The following packages are needed:

radare2
python3
pip3
pkg-config # Required to install r2ghidra

TODO: Add more packages

In addition a .env file should be created with an OpenAI API key, looking like:

export OPENAI_API_KEY="key here"

Setting up and installing python requirements:

python -m venv .venv
source init_shell.sh
pip3 install -r requirements.txt

Sourcing init_shell.sh has to be done each time you start a new terminal.

For the tests:

pip3 install -r requirements_dev.txt

Running

To see how to use the main.py program can be used to run the pipeline use python main.py -h:

usage: main.py [-h] [-d DATA_DIR] [-r RESULTS] [-m MODEL] [-s] [-o] {clean,prepare,print,evaluate}

Run pipeline commands.

positional arguments:
  {clean,prepare,print,evaluate}
                        Command to execute

options:
  -h, --help            show this help message and exit
  -d DATA_DIR, --data-dir DATA_DIR
                        Data directory
  -r RESULTS, --results RESULTS
                        Results filename
  -m MODEL, --model MODEL
                        Model name
  -s, --strip           Strip the binary during compilation
  -o, --objdump         Use objdump instead of r2 for disassembly

For the most simple use case, using the 4 sources in data/ (default directory), run python main.py evaluate

For decompile-eval tests, first run python extract_decompile-eval.py and then python main.py -d data_decompile_eval evaluate. The first scripts downloads and creates the separate source directory.

Testing

Run ptw -- --cov=src --cov-report=term-missing to contiously test and produce coverage tests.