Replication study of "Privacy-preserving Collaborative Learning with Automatic Transformation Search"

This is the code behind a replication study of Privacy-preserving Collaborative Learning with Automatic Transformation Search. The code for the original implementation can be found here.

Setup

1. Set up the conda environment.

conda env create -f environment.yml
conda activate reproducing-ats

2. (Optional) Download our logs and checkpoints and unzip in the root directory of this repository, so that you now have a logs folder.
3. Read the usage section below or dive straight into our report.ipynb Jupyter notebook.

Usage

The entrypoint for this project is main.py. It has three main capabilities:

• Search: perform a policy search.
• Train: train a model on a dataset, optionally using a particular policy and/or alternative defense.
• Attack: reconstruct an image given a gradient from a model.

See the prerendered report.ipynb Jupyter notebook for a tour through its functionality. If you download our logs and checkpoints, you can fully rerender our figures and tables.

usage: main.py [-h] [--data-dir DATA_DIR] {search,train,attack,test} ...

optional arguments:
  -h, --help            show this help message and exit
  --data-dir DATA_DIR

command:
  {search,train,attack,test}
                        Action to execute
    search              Automatic transformation search
    train               Model training
    attack              Perform reconstruction attack
    test                Test a model