Amplified Patch-Level Differential Privacy for Free via Random Cropping

This is the official implementation of our TMLR paper.

"Amplified Patch-Level Differential Privacy for Free via Random Cropping"

Kaan Durmaz, Jan Schuchardt, Sebastian Schmidt, Stephan Günnemann.

Requirements

To install the requirements, execute:

uv venv --python 3.11
source .venv/bin/activate
uv pip install -r requirements.txt

Installation

You can install this package via uv pip install -e .

Usage

In order to reproduce all experiments, you will need to execute the scripts in seml/scripts using the config files provided in seml/configs using the SLURM Experiment Management Library.

After computing all results, you can use the scripts in plotting to recreate the figures from the paper.

Cite

Please cite our paper if you use this code in your own work:

@article{durmaz2025amplified,
    title={Amplified Patch-Level Differential Privacy for Free via
    Random Cropping},
    author={Kaan Durmaz and Jan Schuchardt and Sebastian Schmidt and
    Stephan G{\"u}nnemann},
    journal={Transactions on Machine Learning Research},
    year={2025},
    url={https://openreview.net/forum?id=pSWuUF8AVP}
}