Home

Awesome

Stochastic Variance Reduced Ensemble (SVRE)

This repository contains code to reproduce results from the paper: Stochastic Variance Reduced Ensemble Adversarial Attack for Boosting the Adversarial Transferability(CVPR2022). We provide an example of the SVRE method, and the complete experimental code and data will be released soon.

Datasets And models

To run the code, you should download pre-trained models and the data.

Please place pre-trained models under the models/ directory.

Please unzip the data and place the data under the dataset/ directory.

Requirements

File Description

Experiments

We provide an example of generating the adversarial examples on the ensemble of four normally trained models, ie. Inc-v3, Inc-v4, Res-15 and IncRes-v2, and test the transferability of the crafted adversaries on defense models.

To generate adversarial exmples of SVRE-I-FGSM and Ens-I-FGSM:

CUDA_VISIBLE_DEVICES=[gpu id] python SVRE-I-FGSM.py
CUDA_VISIBLE_DEVICES=[gpu id] python Ens-I-FGSM.py

To eval the efficacy of SVRE-I-FGSM and Ens-I-FGSM:

CUDA_VISIBLE_DEVICES=[gpu id] python eval.py  --eval_file ./results/SVRE-I-FGSM/
CUDA_VISIBLE_DEVICES=[gpu id] python eval.py  --eval_file ./results/Ens-I-FGSM/

Acknowledgements

In order to ensure that our personal information is not leaked, we obtain the download link of the model from open source repositories, eg. SI-NI-FGSM. We thank the authors for sharing.