Disassembly is the cornerstone of many cybersecurity solutions (vulnerability search, malware classification, etc.). Deepbits leverages the recent advance in deep learning to develop DeepDi , a fast and accurate disassembler. It is comparable or superior to the state-of-the-art disassemblers in terms of accuracy, and is robust against unseen compilers and platforms, obfuscated binaries, and adversarial attacks. Its CPU version is two times faster than IDA Pro, and its GPU version is over 300 times faster.
This is a community edition free for non-commercial use only. An online disassembler powered by DeepDi can be found at https://deepdi.deepbits.com/.
[USENIX Security'22] Sheng Yu, Yu Qu, Xunchao Hu, and Heng Yin, DeepDi: Learning a Relational Graph Convolutional Network Model on Instructions for Fast and Accurate Disassembly, to appear in the 31st USENIX Security Symposium, August 2022.
- Fast and accurate x86 disassembler
- Obfuscation resilient
- GPU acceleration
- Good scalability
- Python interface
- Produce disassembly instructions and function boundaries
- Docker
- DeepDi runs on most modern x86_64 machines, with optional GPU support provided via CUDA. The GPU version requires Maxwell cards and above, and the minimum required driver version for CUDA is
>=450.80.02for Linux and>=452.39for Windows.
- Install Docker (and NVIDIA Container Toolkit to pass NVIDIA GPUs to containers)
- Clone the master branch
- Run
docker build -t DeepDi .(ordocker build -t DeepDi -f Dockerfile-gpu .for GPU support). - Run
docker run --rm -it DeepDi. Add--gpus allto the docker command line if you intend to use GPU. - Usage:
python3 DeepDi.py --key KEY [--gpu] --path PATH. You can obtain a key for free at https://www.deepbits.com/deepdi.
The core functionalities are provided via DeepDiCore.so. See API.md for the low-level APIs.
We also provide DeepDi.py, an example wrapper of the low-level APIs to extract inputs for DeepDi and return instruction and function virtual addresses.
If you have any questions, please create an issue or reach us via https://www.deepbits.com/about#contact-us.