DE-CROP: Data-efficient Certified Robustness for Pretrained Classifiers (WACV 2023) - Official Implementation

Paper Link: https://arxiv.org/pdf/2210.08929.pdf

Project Webpage: https://sites.google.com/view/decrop

---

Method Overview

---

Training & Evaluation

1. Pre-train denoiser using $L_{lc}$, $L_{cs}$, $L_{mmd}$ losses: pretrain_denoiser_decrop.sh
2. Train denoiser with additional $L_{dd}$ loss along with Gradient Reversal Layer (initialized using pre-trained weights in 1.): train_denoiser_decrop.sh
3. Certify denoiser: certify_denoiser_decrop.sh
4. Visualize Results: python code/graph_decrop.py

---

Citation

If you use this code, please cite our work as:

@inproceedings{
    nayak2023_DECROP,
    title={DE-CROP: Data-efficient Certified Robustness for Pretrained Classifiers},
    author={Nayak, G. K., Rawal, R., and Chakraborty, A.},
    booktitle={IEEE Winter Conference on Applications of 
    Computer Vision},
    year={2023}
}

---

Acknowledgements

This repo is adapted from Salman et al. 2020 and utilizes utility functions from torchattacks library for generating boundary and interpolated samples.