mdzhangst/dbd

DBD

This a Pytorch implementation of our paper "Backdoor Defense via Decoupling the Training Process".

Table of Contents:

• DBD
  • Setup
    • Environments
    • Datasets
    • Log and checkpoint directories
  • Usage
    • No Defense
    • DBD
  • Pretrained Models
    • Test
    • Results
• License
• Citation

Setup

Environments

We recommend conda as the package manager to setup the environment used in our experiments. Create the environment dbd from the environment.yml file and activate it:

conda env create -f environment.yml && conda activate dbd

Datasets

Download CIFAR-10 dataset from its official website and extract it to dataset_dir specified in the YAML configuration files.

Note: Make sure dataset_dir contains the sub-string cifar.

Log and checkpoint directories

Create saved_dir and storage_dir specified in the YAML configuration files to save logs and checkpoints respectively:

mkdir saved_data && mkdir storage

Usage

We give examples to compare the standard supervised training (No Defense) and DBD on CIFAR-10 dataset under BadNets attack with ResNet-18. Other settings can also be found in the YAML configuration files. Please have an overview before running the codes.

No Defense

Run the following script to train a backdoored model:

python supervise.py --config config/supervise/badnets/cifar10_resnet18/example.yaml \
                    --resume False \
                    --gpu 0

DBD

1. Self-Supervised Learning

   Run the following script to train a purified feature extractor:

   python simclr.py --config config/defense/simclr/badnets/cifar10_resnet18/example.yaml \
                    --resume False \
                    --gpu 0
   

2. Semi-Supervised Fine-tuning

   Run the following script to finetune a clean model:

   python mixmatch_finetune.py --config config/defense/mixmatch_finetune/badnets/cifar10_resnet18/example.yaml \
                               --resume False \
                               --gpu 0
   

Pretrained Models

We provide pretrained models here.

Test

Run the following script to test No Defense under BadNets attack:

python test.py --config config/supervise/badnets/cifar10_resnet18/example.yaml \
               --ckpt-dir checkpoint/supervise/badnets/cifar10_resnet18/example \
               --resume latest_model.pt \
               --gpu 0

Run the following script to test DBD under BadNets attack:

python test.py --config config/supervise/badnets/cifar10_resnet18/example.yaml \
               --ckpt-dir checkpoint/defense/mixmatch_finetune/badnets/cifar10_resnet18/example \
               --resume latest_model.pt \
               --gpu 0

Run the following script to test No Defense under Blended attack:

python test.py --config config/supervise/blend/cifar10_resnet18/example.yaml \
               --ckpt-dir checkpoint/supervise/blend/cifar10_resnet18/example \
               --resume latest_model.pt \
               --gpu 0

Run the following script to test DBD under Blended attack:

python test.py --config config/supervise/blend/cifar10_resnet18/example.yaml \
               --ckpt-dir checkpoint/defense/mixmatch_finetune/blend/cifar10_resnet18/example \
               --resume latest_model.pt \
               --gpu 0

Results

Method	BadNets BA (%)	BadNets ASR (%)	Blended BA (%)	Blended ASR (%)
No Defense	95.13	100	94.26	98.15
DBD	92.50	0.88	92.60	0.31

License

Our codes is released under GNU General Public License v3.0.

Citation

If our work or this repo is useful for your research, please cite our paper as follows:

@inproceedings{huang2022backdoor,
  title={Backdoor Defense via Decoupling the Training Process},
  author={Kunzhe Huang, Yiming Li, Baoyuan Wu, Zhan Qin and Kui Ren},
  booktitle={ICLR},
  year={2022}
}