PrivacyFlow

Requirements

• Python 3.10.4
• (Optional) Cuda 11.8
• pip or conda (recommended: pip)

To install the required Python packages run the following code in the terminal.

pip install -r requirements.txt

Code Structure

The module privacyflow contains several python classes, which are used inside the notebooks. The main components are configs, datasets, models and preprocessing functionality. Some empty folders (containing only a .gitkeep) are also located inside this module, which are used for saving data.

There are 4 Jupyter notebooks, which contain executable code, where different models are trained and evaluated. The Differential_Privacy_*.ipynb notebooks contain code for training models with and without DPSGD. These models are used inside the Membership_Inference_Attack.ipynb notebook and the Model_Inversion_Notebook.ipynb to evaluate attacks against these models.

├── privacyflow
│   ├── configs
│   ├── datasets
│   ├── models
│   ├── preprocessing
├── Differential_Privacy_Cifar10.ipynb
├── Differential_Privacy_Face_Models.ipynb
├── Membership_Inference_Attack.ipynb
├── Model_Inversion_Attack.ipynb