Stochastic-Serial-Attack

Introduction

This repository include the code for "Boosting the Transferability of Adversarial Examples via Stochastic Serial Attack"

Method

We propose a novel attack strategy called stochastic serial attack (SSA). It adopts a serial strategy to attack local models, which reduces memory consumption compared to parallel attacks. Moreover, since local models are stochastically selected from a large model set, it can ensure that the adversarial examples do not overfit specific weaknesses of local source models.

Citations

If you use Stochastic Series Attack in your research, please consider citing

Getting started

1. git clone https://github.com/haolingguang/SSA.git

2. Install dependencies:
   Python >= 3.6
Pytorch >=1.3
pip install pretrainedmodels
   pretrainedmodels is a pretrained models repository of pytorch, which include all pytorch models used in our paper.

3. archive include the dataset in NeurIPS 2017 adversarial competition. If you want use ImageNet, you need download corresponding dataset and use torchvision.datasets.ImageFolder to load dataset.

4. run attack
   python SSA.py

5. If you need load adversarial training models, you need use https://github.com/haolingguang/tensorflow-to-pytorch.git to convert tensorflow's pre-training models to pytorch's

6. In order to get the mini-ImageNet, you can refer to https://github.com/WZMIAOMIAO/deep-learning-for-image-processing/tree/master/pytorch_classification/mini_imagenet

7. Create attack dataset for the models trained by mini-ImageNet, you need run Rename_image.py in mini_dataset.

Note

Package pretrainedmodels maybe have some problem, But you can use another pytorch model package timm
For example: parallel computing
import timm
t = timm.create_model('resnet50', num_classes=1000, pretrained=True)
t = torch.nn.DataParallel(t).cuda()

Contact

haolingguang@mail.dhu.edu.cn