This is a complete reimplementation of our Analysis by Synthesis model. The experiments in the paper were done with the original implementation. To the best of my knowledge, both are equivalent and I tried to carefully reproduce all results but I cannot make any guarantees.
Our paper can be found on arXiv. It was accepted at ICLR 2019.
All data from the accuracy-distortion curves can be downloaded in raw form:
- figures/L2_accuracy_distortion_curves.pickle
- figures/Linf_accuracy_distortion_curves.pickle
- figures/L0_accuracy_distortion_curves.pickle
@inproceedings{schott2018towards,
title={Towards the first adversarially robust neural network model on {MNIST}},
author={Lukas Schott and Jonas Rauber and Matthias Bethge and Wieland Brendel},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=S1EHOsC9tX},
}