Code for the implementation of the paper Adversarial Attacks on Variational Autoencoders - George Gondim-Ribeiro, Pedro Tabacof, Eduardo Valle, 2018.

Please cite our work:

George Gondim-Ribeiro, Pedro Tabacof, and Eduardo Valle. Adversarial Attacks on Variational Autoencoders. arXiv:1806.04646.

After downloading and preparing the data, to reproduce our experiments, first, run the script trainer.py to train a model and the attack.py to start the attack choosing the same directory. Then, you can compute the AUDDC on the normalized results.

For instance:

# Download data
python get_data.py

# Prepare CelebA dataset
python celeba_prepare_data.py

# Train model
python trainer.py --epochs 50 --dataset mnist --architecture vae --dir /tmp/test

# Attack model
python attack.py --dir /tmp/test/ --num_attacks 5

# Compute metrics
python compute_metrics.py --dir /tmp/test

vae: Variational autoencoders with only fully-connected layers

cvae: Variational autoencoders with convolutional layers

draw: DRAW

mnist: MNIST dataset

svhn: SVHN dataset

celeba: CelebA dataset

This code is based upon the repositories below, please check their license before any use. This code is intended ONLY for academic purposes.

• https://github.com/sjchoi86/advanced-tensorflow
• https://github.com/ericjang/draw