HEDiff

Paper

Testing and Understanding Deviation Behaviors in FHE-hardened Machine Learning Models

Run

1. Train models with plain_models.py and plain_models_tf.py. If you want to use HElayers, you need to convert torch model into onnx with fhe_onnx_convert first.
2. Run diff_{name}.py for Margin-guided Differential Testing. name: 'tenseal' for TenSEAL, 'zama' for Concrete-ML, 'helayer' for HElayers.
3. If needed, run other files for the statistic information about deviation inputs in FHE-ML models.

Datasets List

MNIST: http://yann.lecun.com/exdb/mnist/

DIGITS: https://scikit-learn.org/stable/auto_examples/datasets/plot_digits_last_image.html

Credit: https://archive.ics.uci.edu/dataset/350/default+of+credit+card+clients

Bank: https://archive.ics.uci.edu/dataset/222/bank+marketing

Code for data processing is in ./dataset

Stealthiness Examples

---

---

Tips

1. Dependency information

concrete-ml               1.4.0
imbalanced-learn          0.12.3
scikit-learn              1.1.3
tenseal                   0.3.14
tensorflow-cpu            2.10.0
torchattacks              3.5.1
pandas                    2.1.1
pyhelayers                1.5.3.1
pytorch                   1.13.1
numpy                     1.23.5