Security scanner detecting Python Pickle files performing suspicious actions.

For more generic model scanning, Protect AI's modelscan is now available to scan not only Pickle files but also PyTorch, TensorFlow, and Keras.

Getting started

Scan a malicious model on Hugging Face:

pip install picklescan
picklescan --huggingface ykilcher/totally-harmless-model

The scanner reports that the Pickle is calling eval() to execute arbitrary code:

https://huggingface.co/ykilcher/totally-harmless-model/resolve/main/pytorch_model.bin:archive/data.pkl: global import '__builtin__ eval' FOUND
----------- SCAN SUMMARY -----------
Scanned files: 1
Infected files: 1
Dangerous globals: 1

The scanner can also load Pickles from local files, directories, URLs, and zip archives (a-la PyTorch):

picklescan --path downloads/pytorch_model.bin
picklescan --path downloads
picklescan --url https://huggingface.co/sshleifer/tiny-distilbert-base-cased-distilled-squad/resolve/main/pytorch_model.bin

To scan Numpy's .npy files, pip install the numpy package first.

The scanner exit status codes are (a-la ClamAV):

• 0: scan did not find malware
• 1: scan found malware
• 2: scan failed

Develop

Create and activate the conda environment (miniconda is sufficient):

conda env create -f conda.yaml
conda activate picklescan

Install the package in editable mode to develop and test:

python3 -m pip install -e .

Edit with VS Code:

code .

Run unit tests:

pytest tests

Run manual tests:

• Local PyTorch (zip) file

mkdir downloads
wget -O downloads/pytorch_model.bin https://huggingface.co/ykilcher/totally-harmless-model/resolve/main/pytorch_model.bin
picklescan -l DEBUG -p downloads/pytorch_model.bin

• Remote PyTorch (zip) URL

picklescan -l DEBUG -u https://huggingface.co/prajjwal1/bert-tiny/resolve/main/pytorch_model.bin

Lint the code:

black src tests
flake8 src tests --count --show-source

Publish the package to PyPI: bump the package version in setup.cfg and create a GitHub release. This triggers the publish workflow.

Alternative manual steps to publish the package:

python3 -m pip install --upgrade pip
python3 -m pip install --upgrade build
python3 -m build
python3 -m twine upload dist/*

Test the package: bump the version of picklescan in conda.test.yaml and run

conda env remove -n picklescan-test
conda env create -f conda.test.yaml
conda activate picklescan-test
picklescan --huggingface ykilcher/totally-harmless-model

Tested on Linux 5.10.102.1-microsoft-standard-WSL2 x86_64 (WSL2).

References

• pickletools.py -- The pickletool code is the most detailed documentation of the Pickle format.
• Machine Learning Attack Series: Backdooring Pickle Files, Johann Rehberger, 2022
• Hugging Face Pickle Scanning, Luc Georges, 2022
• The hidden dangers of loading open-source AI models (ARBITRARY CODE EXPLOIT!, Yannic Kilcher, 2022
• Secure Machine Learning at Scale with MLSecOps, Alejandro Saucedo, 2022
• Backdooring Pickles: A decade only made things worse, ColdwaterQ, DEFCON 2022
• Never a dill moment: Exploiting machine learning pickle files, Evan Sultanik, 2021 (tool: Fickling)
• Exploiting Python pickles, David Hamann, 2020
• Dangerous Pickles - malicious python serialization, Evan Sangaline, 2017
• Python Pickle Security Problems and Solutions, Travis Cunningham, 2015
• Arbitrary code execution with Python pickles, Stephen Checkoway, 2013
• Sour Pickles, A serialised exploitation guide in one part, Marco Slaviero, BlackHat USA 2011 (see also: doc, slides)