/DnD

A decompiler to automatically reverse-engineer the DNN semantics from its compiled binary using program analysis

Primary LanguagePythonApache License 2.0Apache-2.0

DnD: A Cross-Architecture Deep Neural Network Decompiler

DND is a deep neural network (DNN) decompiler capable of reverse engineering DNN models from compiled binaries.

Environment

  1. Create a Python virtual environment
  2. pip install -r ./angr_env/requirements.txt
  3. Move ./angr_env/base.py to overwrite the counterpart in the virtual environment claripy (usually in $USERNAME/.virtualenvs/$VIRTUALENVNAME/lib/python$VERSION/site-packages/claripy/ast/base.py)

Docker container

We provide a docker container. To use it, just run:

docker build . -t dnd
docker run -it dnd

Usage

  • Run python decompiler.py <model_binary> <model_onnx> to decompile a binary sample (<model_binary>) and export it into an ONNX file (<model_onnx>).For instance:
python ./decompiler.py binary_samples/evkbimxrt1050_glow_lenet_mnist_release.axf onnx_models/mnist.onnx
python ./decompiler.py binary_samples/evkbimxrt1050_glow_cifar10.axf onnx_models/resnet.onnx

  • Two samples are provided:
    ./binary_samples/evkbimxrt1050_glow_lenet_mnist_release.axf: a MNIST binary on NXP imrt1050-evk board
    ./binary_samples/evkbimxrt1050_glow_cifar10.axf: a Resnet binary on NXP imrt1050-evk board

  • The folder onnx_models containes ONNX models exported by DND for the binaries in binary_samples.

  • The folder patches showcases how to patch binaries implementing DNNs.

  • Please check decompiler.py for more details.

Citing this work

  title={$\{$DnD$\}$: A $\{$Cross-Architecture$\}$ deep neural network decompiler},
  author={Wu, Ruoyu and Kim, Taegyu and Tian, Dave Jing and Bianchi, Antonio and Xu, Dongyan},
  booktitle={31st USENIX Security Symposium (USENIX Security 22)},
  pages={2135--2152},
  year={2022}
}