Official code implementation of BackdoorIndicator: Leveraging OOD Data for Proactive Backdoor Detection in Federated Learning(https://www.usenix.org/conference/usenixsecurity24/presentation/li-songze)
You first need to install relevant packages using:
pip install -r requirement.txt
For the version of important packages:
Python==3.7.15
torch==1.13.0
torchvision==0.14.0
For the edge-case datasets, you can acquire them following the instructions of https://github.com/ksreenivasan/OOD_Federated_Learning, which is the official repo of the Yes-you-can-really-backdoor-FL paper.
The code trains an Federated Learning global model from scratch when you run the
code for the first time. It is recommended that do not apply any defense
mechanism, and only saves a few checkpoints of the global model for the first
run. To do this, you may set the follow parameters in
utils/yamls/params_vanilla_indicator.yaml:
poisoned_start_round: 10000 #larger than the biggest global round index you want to save
global_watermarking_start_round: 10000
A recording folder will then be created based on the launching time under
saved_models, where checkpoints will be saved. Then, you may choose any
checkpoint, and put the path in the resumed_model:
resumed_model: "Jun.05_06.09.03/saved_model_global_model_1200.pt.tar"
save_on_round: [xxx, yyy, zzz] #any round you like
To launch any defense mechanism, you may then put the corresponding yaml file in
the command line. For example, to implement BackdoorIndicator, you may first
want to check global_watermarking_start_round and poisoned_start_round, as
these two parameters determine the round where BackdoorIndicator begins and the
poisoning begins. Then you run the code
python main.py --GPU_id "x" --params utils/yamls/indicator/params_vanilla_Indicator.yaml
The results are recorded in the corresponding saved_models folder.
Please feel free to change the following parameters in corresponding yamls to see their influence to the proposed method, as it is discussed in the paper:
ood_data_source
ood_data_sample_lens
global_retrain_no_times
watermarking_mu
We appreciate it if you would please cite the following paper if you found the repository useful for your work:
@inproceedings {299824,
author = {Songze Li and Yanbo Dai},
title = {{BackdoorIndicator}: Leveraging {OOD} Data for Proactive Backdoor Detection in Federated Learning},
booktitle = {33rd USENIX Security Symposium (USENIX Security 24)},
year = {2024},
isbn = {978-1-939133-44-1},
address = {Philadelphia, PA},
pages = {4193--4210},
url = {https://www.usenix.org/conference/usenixsecurity24/presentation/li-songze},
publisher = {USENIX Association},
month = aug
}