This is a reproduction study of the paper Privacy-preserving Collaborative Learning with Automatic Transformation Search
by Wei Gao et al.
The original paper can be found here: https://arxiv.org/pdf/2011.12505.pdf
The original code can be found here: https://github.com/gaow0007/ATSPrivacy
This repository contains both the original code as well as a refactored version. There are notebooks present to reproduce the experiments from our study.
This codebase is meant to be ran on a Linux system with a gpu (cuda). In the root folder an environment.yml
anaconda environment is provided with the exact packages that this project requires.
The experiment that verifies the claim of the correlation between the privacy score and reconstruction PSNR can be found in the folder ATSRefactored under the name Result-Reproduction Correlation and Diversity.ipynb
. This notebook also contains the experiment for the second insight, which investigates the diversity of well-performing policies.
The experiment that verifies the reconstruction attack PSNR and accuracy of a trained model can be found in the root folder under the name Result-Reproduction Cifar100 Trained Models.ipynb
. This file contains the functions to show pre-run results, training the network for a single epoch and running the entire experiment. Additionally Result-Reproduction FMNIST Trained Models.ipynb
contains the code to run the entire experiment for the FMNIST dataset.
Old code from the original paper with slight edits in order to reproduce the results on our systems.
ATSPrivacy-Framework but edited to work for F-Mnist
New, refactored code, used for all our original experiments in the report. Symlinks in some places to the old code, in particular inversefed and the checkpoints is the same for both.