Security-Preserving Federated Learning via Byzantine-Sensitive Triplet Distance

This is an official implementation of the following paper:

Youngjoon Lee, Sangwoo Park, and Joonhyuk Kang. Security-Preserving Federated Learning via Byzantine-Sensitive Triplet Distance
IEEE International Symposium on Biomedical Imaging (ISBI) 2024.

Requirements

The implementation runs on

bash docker.sh

Additionally, please install the required packages as below

pip install tensorboard medmnist

Byzantine attacks

This paper considers the following poisoning attacks

• Targeted model poisoning (Bhagoji, Arjun Nitin, et al. ICML 2019): Targeted model poisoning attack for federated learning
• MPAF (Xiaoyu Cao, Neil Zhenqiang Gong. CVPR Workshop 2022): Untargeted model poisoning attack for federated learning

Byzantine-Robust Aggregation Techniques

This paper considers the following Byzantine-Robust aggregation techniques

• Vanilla (McMahan, Brendan, et al. AISTATS 2017)
• Krum (Blanchard, Peva, et al. NIPS 2017)
• Trimmed-mean (Yin, Dong, et al. ICML 2018)
• Fang (Fang, Minghong, et al. USENIX 2020)

Dataset

• Blood cell classification dataset (Andrea Acevedo, Anna Merino, et al. Data in Brief 2020)

Experiments

Without Byzantine attacks experiment runs on

bash execute/run0.sh

Impact of Byzantine percentage runs on

bash execute/run1.sh

Impact of non-iid degree runs on

bash execute/run2.sh

Acknowledgements

Referred http://doi.org/10.5281/zenodo.4321561