Backdoor attack via augmentation

Contents

1. Introduction
2. Requirements
3. Experiments

Agattack - Backdoor Attack via Augmentation

Introdution

Requirements

• Install required python packages:

$ python -m pip install -r requirement.py

Experiments

Train command

Clean label: only operate on images

$ python train.py --dataset <datasetName> --p <ratio> 

where the parameters are the following:

• <datasetName>: mnist , cifar10 , gtsrb , celeba.

The trained checkpoints should be saved at the path checkpoints\<datasetName>\<datasetName>_<attackMode>_<attackChoice>.pth.tar.

Test for trained models, run command (clean label)

$ python eval.py --dataset <datasetName> --p <ratio> 

Other parameters are the following:

• --target_label: Labels to be attacked