Note - This repo was created so that I could cleanup AlphaPav's repo for personal use. All original work belongs to the concerned parties mentioned in the original repo. Please check Acknowledgement section below for details.
This repository contains the code for ICLR 2020 paper DBA: Distributed Backdoor Attacks against Federated Learning
Create a new conda environment using the following command (remember to replace env_name with your preferred environment name)
conda create --name env_name
Install PyTorch and other missing packages
- download the raw dataset lending-club-loan-data.zip into dir
./utils - rename the zip to
lending-club-loan-data.zip - preprocess the dataset.
cd ./utils
python load_preprocess.py
Note: If the above command fails, comment out the following snippet from loan_preprocess.py
with zipfile.ZipFile(source_dir, 'r') as zip_ref:
if not os.path.exists("../data"):
os.mkdir("../data")
zip_ref.extractall("../data")and then execute the process_loan_data.sh script in utils directory
cd ./utils
./process_loan_data.sh
- download the dataset tiny-imagenet-200.zip into dir
./utils - preprocess the dataset
cd ./utils
python tinyimagenet_reformat.py
Note: If the above command fails, comment out the following snippet from tinyimagenet_reformat.py
with zipfile.ZipFile(source_dir, 'r') as zip_ref:
if not os.path.exists("../data"):
os.mkdir("../data")
zip_ref.extractall("../data")and then execute the process_tiny_data.sh script in utils directory
MNIST and CIFAR will be automatically download
-
prepare the pretrained model: Our pretrained clean models for attack can be downloaded from Google Drive. You can also train from the round 0 to obtain the pretrained clean model.
-
we can use Visdom to monitor the training progress
python -m visdom.server
Note - If the default port 8097 is already in use, you can start visdom a custom port like so:
python -m visdom.server -p 8098
Remember to change the port number (VIS_PORT) in config.py
- run experiments for the four datasets
python main.py -p utils/X.yaml
or
python main.py --params utils/X.yaml
X = mnist_params, cifar_params,tiny_params or loan_params. Parameters can be changed in those yaml files to reproduce our experiments.
If you find our work useful in your research, please consider citing:
@inproceedings{
xie2020dba,
title={DBA: Distributed Backdoor Attacks against Federated Learning},
author={Chulin Xie and Keli Huang and Pin-Yu Chen and Bo Li},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=rkgyS0VFvr}
}