pytorch implementation of SemiAdv

This repository contains code of SemiAdv (SemiAdv: Query-Efficient Black-Box Adversarial Attack with Unlabeled Data) implemented in Pytorch.

Requirements

• Python 3.9.2
• Pytorch 1.9
• Torchvision 0.1.8

Instructions

Quick Start

We prepare an easy demo for quick start.

python train_black_model.py
python train_substitute_model.py
python attack.py

The train_black_model.py contains the code of training a black-box model.

The train_substitute_model.py contains the code of training a substitute model of the black-box model.

The attack.py contains the code of implementing black-box attack by using the sustitute model against the black-box model.

The default setting is as follows:

• black model: MobileNet,
• sustitute model: WideResNet-28,
• default attack method: PGD with our method,
• dataset: CIFAR-10,
• query number (labeled data): 1600.

More defulat setting or info refers to source code.

Training black-box model

Training substitute model

More information will quickly arrive.

Attack

More information will quickly arrive.

Others

customizing your model or dataset:

If you have any questions, please contact us or leave a message here.

Citing this paper