ACM International Conference on Multimedia

This is the MindSpore implementation of FSMAFL in the following paper.

Wenke Huang, Mang Ye, Xiang Gao, Bo Du. Few-Shot Model Agnostic Federated Learning, in ACM MM, 2022.

FSMAFL Description

FSMAFL(Few-Shot Model Agnostic Federated Learning) is a latent embedding adaptation framework that can address the large domain gap between the public and private datasets in federated learning process. It is based on two parts:

1. Latent embedding adaptation confuses domain classifier to reduce the impact of domain gap.

2. Model agnostic federated learning is responsible for public-private communication. The public dataset acts as the bridge for model communication and private dataset is used for avoiding forgetting.

Framework Architecture

Dataset

Our experiments are conducted on two datasets, MNIST and EMNIST-Letter. The public dataset on the server is set to MNIST. and the private dataset on the client is set to EMNIST-Letter.

Note: Data will be processed in data_init.py

Environment Requirements

Hardware

• Support Ascend environment.
• Ascend: Ascend 910.

Framework

• MindSpore

For more information, please check the resources below：

• MindSpore Tutorials
• MindSpore Python API

Quick Start

After installing MindSpore via the official website, you can start training and evaluation as follows:

# Initialize the local models
python model_initialization.py
# FSMAFL
python Communication_GAN.py

Script Description

Script and Sample Code

├── FSMAFL
    ├── Dataset
        ├── MNIST
            ├── test
                ├── t10k-images-idx3-ubyte
                ├── t10k-labels-idx1-ubyte
            ├── train
                ├── train-images-idx3-ubyte
                ├── train-labels-idx1-ubyte
        ├── emnist-letters
    ├── Temp
        ├── priv_data_72.npy
        ├── total_priv_data_72.pickle
    ├── collabporate_train.py
    ├── Communication_GAN.py
    ├── data_utils.py
    ├── model_initialization.py
    ├── model_utils.py
    ├── models.py
    ├── option.py
    ├── README.md

Model Description

Performance

Evaluation Performance

The experimental setting is slightly different from the original paper due to different platforms. In the heterogeneous model scenario, we assign five different networks in models.py. The accuracy(%) is based on the EMNIST-Letter dataset. Initial represents only do initialization without federated learning process.

	θ1	θ2	θ3	θ4	θ5	Avg
Initial	20.16	26.37	45.02	41.04	42.33	34.98
FSMAFL	23.67	30.31	47.08	43.92	47.25	38.45

Citation

@inproceedings{huang2022fewshot,
  title={Few-Shot Model Agnostic Federated Learning},
  author={Huang, Wenke and Ye, Mang and Gao, Xiang and Du, Bo},
  booktitle={ACM MM Industry Track},
  year={2022}
}