This is the official implementation of Latent Feature Relation Consistency for Adversarial Robustness.
- May 2023 - Release the code and models.
- March 2023 - Release the pre-print on arXiv.
This repo is built upon python==3.8.5
and pytorch==1.6.0
.
The similarity matrix of natural examples is more compact than adversarial examples, i.e., the similarity matrix of natural examples has brighter blocks on the diagonal than the adversarial similarity matrix, which indicates that the latent representations of natural examples from the same class are more concentrated than those of adversarial examples.
The white-box robustness of LFRC with ResNet-18 on CIFAR-10 and CIFAR-100 datasets is shown below. Please refer to our paper for more details. The models' weight and train-logs can be found in here.
This project is under the Apache 2.0 license as found in LICENSE file.
Please consider citing LFRC and giving a star if LFRC helps your research:
@article{liu2023latent,
title={Latent Feature Relation Consistency for Adversarial Robustness},
author={Liu, Xingbin and Kuang, Huafeng and Liu, Hong and Lin, Xianming and Wu, Yongjian and Ji, Rongrong},
journal={arXiv preprint arXiv:2303.16697},
year={2023}
}