MORA

This repository contains code for reproducing the results in our NeurIPS 2022 paper "MORA: Improving Ensemble Robustness Evaluation with Model-Reweighing Attack".

Please feel free to cite our paper with the following bibtex entry:

@inproceedings{mora,
 author = {Yu, Yunrui and Gao, Xitong and Xu, Cheng-Zhong},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {S. Koyejo and S. Mohamed and A. Agarwal and D. Belgrave and K. Cho and A. Oh},
 pages = {26955--26965},
 publisher = {Curran Associates, Inc.},
 title = {{MORA}: Improving Ensemble Robustness Evaluation with Model Reweighing Attack},
 url = {https://proceedings.neurips.cc/paper_files/paper/2022/file/ac895e51849bfc99ae25e054fd4c2eda-Paper-Conference.pdf},
 volume = {35},
 year = {2022}
}

Dependencies

Create the conda environment called mora containing all the dependencies by running:

conda env create -f environment.yml

We used PyTorch 1.4.0 for all the experiments, and the code were tested on an NVIDIA TITAN Xp GPU.

Pretrained models

The pretrained models for the ensemble defense strategies (ADP, DVERGE, GAL) can be accessed via this link. The pre-trained models are located in the folder named checkpoints. Download and place the checkpoints into a checkpoints/ folder under this repo before running evaluation scripts.

Usage

Examples of evaluation scripts can be found in scripts/evaluation.sh.