FedKA

Official repository for the ACM MM 2023 paper:

Chen J, Zhu J, Zheng Q. Towards fast and stable federated learning: Confronting heterogeneity via knowledge anchor. In Proceedings of the 31st ACM International Conference on Multimedia (ACM MM), 2023, pp. 8697-8706.

Overview

This repository contains the implementation of FedKA, a federated learning method designed to improve training efficiency and stability under heterogeneous client data distributions.

The codebase includes:

• dataset generation scripts under PFL/dataset/
• federated training pipeline under PFL/system/
• experiment templates under PFL/system/template/
• environment specification in freeze.yml

Getting Started

Step 1. Create the environment from freeze.yml

At the repository root, create and activate the conda environment:

conda env create -f freeze.yml
conda activate FL

The environment name is defined as FL in freeze.yml.

Step 2. Generate the dataset

For CIFAR-10 with Dirichlet non-IID partitioning, run:

cd PFL/dataset
python generate_cifar10.py noniid - dir 20 0.1

This command is consistent with the argument parser in PFL/dataset/generate_cifar10.py, where the arguments are interpreted as:

python generate_cifar10.py <iid_or_noniid> <balance_or_-> <partition> <num_clients> <dirichlet_alpha>

So in the above example:

• noniid means non-IID partitioning
• - means not using balanced splitting
• dir means Dirichlet partitioning
• 20 is the number of clients
• 0.1 is the Dirichlet concentration parameter

After generation, the dataset directory will be created automatically, e.g.:

Cifar10_NonIID_Dir0.1_Client20/

Step 3. Run federated training

Go to the training directory and launch an experiment with the provided config file:

cd ../system
python main.py --config template/FedKA_cifar10_dir0.1.yaml

The training entry is PFL/system/main.py, and the example configuration file is located at:

PFL/system/template/FedKA_cifar10_dir0.1.yaml

Project Structure

FedKA/
├── freeze.yml
├── README.md
└── PFL/
    ├── dataset/
    │   ├── generate_cifar10.py
    │   └── ...
    └── system/
        ├── main.py
        ├── template/
        │   └── FedKA_cifar10_dir0.1.yaml
        └── ...

Notes

• The commands above are written to match the actual repository structure.
• There is no top-level main.py at the repository root; the correct training entry is PFL/system/main.py.
• Likewise, the CIFAR-10 data generation script is PFL/dataset/generate_cifar10.py.

Citation

If you find this repository useful, please cite:

@inproceedings{chen2023towards,
  title={Towards fast and stable federated learning: Confronting heterogeneity via knowledge anchor},
  author={Chen, J and Zhu, J and Zheng, Q},
  booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
  pages={8697--8706},
  year={2023}
}