This is the repo for IPSN 2022 paper: "BalanceFL: Addressing Class Imbalance in Long-Tail Federated Learning".

BalanceFL is a long-tailed federated learning framework that can robustly learn both common and rare classes from a real-world dataset, simultaneously addressing the global and local data imbalance problems.

BalanceFL mainly includes two components: knowledge inheritance and inter-class balancing. They address the class missing issue and the local class imbalance issue, respectively.

The program has been tested in the following environment:

• Python 3.7.11
• Pytorch 1.8.1
• torchvision 0.9.1
• torchaudio 0.8.1
• numpy 1.21.2
• librosa 0.6.0
• PyYAML 5.4.1
• Pillow 8.3.2
• h5py 3.4.0

In Total, the evaluation involves three datasets. Each folder contains the code of one dataset.

• CIFAR10 and Speech Command are two public datasets, which will be automatically downloaded by running the related code.
• Our collected IMU dataset is available in ./dataset/IMU/.

quick start: run python3 XXX.py

• train_ours.py: ours
• train_central_bal.py: centralized training with balanced softmax (Balanced Meta-Softmax for Long-Tailed Visual Recognition)
• train_central.py: naive centralized training
• train_fedavg.py: FedAvg
• train_fedprox.py: FedProx
• train_local.py: Local Training
• train_per.py: Per-FedAvg

The dataset included in this repository is for non-commericial use. Please cite our work if you feel our work is helpful or use our collected IMU dataset:

@INPROCEEDINGS{9825928,
  author={Shuai, Xian and Shen, Yulin and Jiang, Siyang and Zhao, Zhihe and Yan, Zhenyu and Xing, Guoliang},
  booktitle={2022 21st ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN)}, 
  title={BalanceFL: Addressing Class Imbalance in Long-Tail Federated Learning}, 
  year={2022},
  volume={},
  number={},
  pages={271-284},
  doi={10.1109/IPSN54338.2022.00029}}