readme.md

Detecting Backdoor Attacks in Federated Learning via Direction Alignment Inspection

This is the official implementation for CVPR'25 ✨Highlight✨ paper "Detecting Backdoor Attacks in Federated Learning via Direction Alignment Inspection".

You can find the full version of our paper here.

Related Methods

If you found AlignIns is interesting, you may also refer to our other related works on robust federated learning systems:

• OBA & BNGuard (MobiHoc 2025, Oral presentation): https://github.com/JiiahaoXU/SoDa-BNGuard
• MASA (WACV 2025): https://github.com/JiiahaoXU/MASA
• LASA (WACV 2025, Oral presentation): https://github.com/JiiahaoXU/LASA

All these methods are based on simple and effective ideas for protecting federated learning systems. They are easy to implement and can be directly integrated into this codebase or into existing federated learning frameworks.

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 --beta 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

We provide the following BibTeX entry for citation for your convenience.

@InProceedings{Xu_2025_CVPR_AlignIns,
    author    = {Xu, Jiahao and Zhang, Zikai and Hu, Rui},
    title     = {Detecting Backdoor Attacks in Federated Learning via Direction Alignment Inspection},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {20654-20664}
}

Acknowledgment

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