Detecting Backdoor Attacks in Federated Learning via Direction Alignment Inspection

This is the official implementation for CVPR'25 $${\color{red}Highlight}$$ paper "Detecting Backdoor Attacks in Federated Learning via Direction Alignment Inspection".

You can find the paper here.

Usage

If you have any issues using this repo, feel free to contact Jiahao @ jiahaox@unr.edu.

The proposed aggregation rule AlignIns is placed in src/aggregation.py, and you can easily take it and integrate AlignIns with your code.

Environment

Our code does not rely on special libraries or tools, so it can be easily integrated with most environment settings.

If you want to use the same settings as us, we provide the conda environment we used in env.yaml for your convenience.

Dataset

CIFAR-10 and CIFAR-100 datasets are available on torchvision and will be downloaded automatically.

Tiny-ImageNet can be easily downloaded from Kaggle.

Example

Generally, to run a case with default settings, you can easily use the following command:

python federated.py \
--poison_frac 0.3 --num_corrupt 4 \
--aggr alignins --data cifar10 --attack badnet

If you want to run a case with non-IID settings, you can easily use the following command:

python federated.py \
--poison_frac 0.3 --num_corrupt 4 \
--non_iid --alpha 0.5 \
--aggr alignins --data cifar10 --attack badnet

Here,

Argument	Type	Description	Choice
aggr	str	Defense method applied by the server	avg, alignins, rlr, mkrum, mmetric, lockdown, foolsgold, rfa
data	str	Main task data	cifar10, cifar100, tinyimagenet
num_agents	int	Number of clients in FL	N/A
attack	str	Attack method	badnet, DBA, neurotoxin, pgd
poison_frac	float	Data poisoning ratio	[0.0, 1.0]
num_corrupt	int	Number of malicious clients in FL	[0, num_agents//2-1]
non_iid	store_true	Enable non-IID settings or not	N/A
beta	float	Data heterogeneous degree	[0.1, 1.0]

For other arguments, you can check the federated.py file where the detailed explanation is presented.

Citation

@article{xu2025detecting,
  title={Detecting Backdoor Attacks in Federated Learning via Direction Alignment Inspection},
  author={Xu, Jiahao and Zhang, Zikai and Hu, Rui},
  journal={arXiv preprint arXiv:2503.07978},
  year={2025}
}

Acknowledgment

Our code is constructed on https://github.com/git-disl/Lockdown, big thanks to their contribution!