DAD: Data-free Adversarial Defense at Test Time (WACV 2022) - Official Implementation

Paper link: https://openaccess.thecvf.com/content/WACV2022/papers/Nayak_DAD_Data-Free_Adversarial_Defense_at_Test_Time_WACV_2022_paper.pdf

Poster: https://drive.google.com/file/d/16VvThgTOMY28dMTlk4-NA9O9WQMyWLEc/view

Project webpage: https://sites.google.com/view/dad-wacv22

---

Method Overview

---

Dependencies

• tqdm
• torch
• numpy
• torchattacks

---

Evaluating Combined Performance (Correction + Detection):

./scripts/combined.sh

---

Citation:

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

    @inproceedings{
        nayak2021_DAD,
        title={DAD: Data-free Adversarial Defense at Test Time},
        author={Nayak, G. K., Rawal, R., and Chakraborty, A.},
        booktitle={IEEE Winter Conference on Applications of 
        Computer Vision},
        year={2022}
    }

---

Acknowledgements

This repo borrows code from Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation and High Frequency Component Helps Explain the Generalization of Convolutional Neural Networks