WUT-ML / privacy

:mag: Research project: Utility-preserving data privatization

Research project: Utility-preserving data privatization

Overview | Requirements | Steps

Overview

This repository contains code for reproducing the experiments with a Generative Adversarial Privatizer , based on the Siamese Generative Adversarial Privatizer for Biometric Data. Note that the NIST Special Database 4 (FIGS), of the fingerprints, has been withdrawn from public use, thus it is not used in this repository.

Requirements

• Python 3.5

• PyTorch 0.3 & torchvision

Steps

0. Setup

pip install -r requirements.txt

1. Dataset preparation

Two datasets, FERG and CelebA can be downloaded and extracted by only setting --dataset argument to either "FERG" or "CelebA". For other datasets new dataloaders have to be written.

2. Siamese GAN training

Run in background/separate session:

tensorboard --logdir runs

Train model on FERG dataset with TensorBoard visualizations:

python src/generative_siamese/generator_plus_siamese_main.py --tensorboard