Introduction

This is the implementation of our paper Decoupling General and Personalized Knowledge in Federated Learning via Additive and Low-rank Decomposition (accepted by ACM MM 2024).

Citation

@misc{wu2024decouplinggeneralpersonalizedknowledge,
      title={Decoupling General and Personalized Knowledge in Federated Learning via Additive and Low-Rank Decomposition}, 
      author={Xinghao Wu and Xuefeng Liu and Jianwei Niu and Haolin Wang and Shaojie Tang and Guogang Zhu and Hao Su},
      year={2024},
      eprint={2406.19931},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2406.19931}, 
}

Dataset

Our experiments utilize three datasets: CIFAR-10, CIFAR-100, and Tiny ImageNet. CIFAR-10 and CIFAR-100 will automatically download when the code is run. For the Tiny ImageNet dataset, one can download the dataset from http://cs231n.stanford.edu/tiny-imagenet-200.zip and extract it to the ./data/ directory.

System

• main.py: Entry point of the program.
• ./utils/options.py: Configuration of experimental hyperparameters.
• ./src/client.py: Client-side code.
• ./src/server.py: Server-side code.
• ./models: Directory for storing backbone model code.

Simulation

Environment

The required experimental environments can be found in requirements.txt.

Training and Evaluation Demo

• Experiment with CIFAR-10 dataset.

  python main.py --num_users=40 --dataset=cifar --model=resnet8 --alpha=1.0 --Conv_r=0.6 --Linear_r=4 --local_p_ep=2

Experimental results can be found in the ./log directory.