CFLF

Federated Learning Experimentation and Visualization System

📜 项目简介

本项目致力于复现联邦学习领域内的重要算法，并进行对比实验。通过高度模块化和可扩展的代码结构，实现对不同联邦学习算法的快速迭代和测试。 进一步，本项目还旨在建立一个联邦学习可视化系统，以增强算法理解和分析效率。

✨ 项目特性

• 🔄 算法复现：复现了诸如FedAvg, FedFAIM等联邦学习核心算法。
• 📊 性能对比：提供了标准化的性能对比工具，可直观比较不同算法。
• 📈 可视化系统：开发了可视化工具，以图形方式显示训练进度和结果。
• 🛠 模块化设计：代码结构清晰，模块化设计便于扩展和维护。
• 📝 扩展升级：框架设计考虑了联邦学习的潜在算法与应用的扩展。

🏗 项目结构

CFLF/
├── algo/
│   ├── base/
│   │   ├── client.py
│   │   └── server.py
│   ├── integrity/
│   │   ├── your integrities/
│   │   └── ...
│   └── method/
│   │   ├── your methods/
│   │   └── ...
├── data/
│   ├── dataset/
│   │   ├── your datasets.py
│   │   └── ...
│   ├── utils/
│   │   ├── distribution.py
│   │   ├── parition.py
│   │   └── visualization.py
│   └── get_data.py
├── datasets/
│   ├── your dataset folders/
│   └── ...
├── model/
│   ├── base/
│   ├── your models/
│   ├── ...
│   └── initialization.py
├── experiment/
│   ├── your experiments/
│   ├── ...
│   └── option.py
├── utils/
│   ├── drawing.py
│   ├── generator.py
│   ├── logging.py
│   ├── running.py
│   └── task.py
├── result/
│   ├── your results/
│   └── ...
├── Visual Interactive SubSystem/
│   └── comming soon...
│── README.py
└── requirements.txt

主要路径

• 实验运行路径：main.py
• 实验配置路径：option.py
• 算法API路径：algo
• 数据集逻辑路径：data
• 模型逻辑路径：model
• 工具集路径：utils
  
  

🚀 快速开始

1. Clone the repository locally:

git clone https://github.com/your_github/FL-Experiment.git

2. Install dependencies:

pip install -r requirements.txt

3. Run the main program:

python constant.py --args

📓 复现记录

Paper: FedFAIM: a model performance-based fair incentive mechanism for federated learning

梯度聚合

• FedAvg:
  • Approach: Built client and server models based on the original paper description.
  • Challenges: Encountered issues with inconsistent weights during model aggregation.
  • Outcome: Successfully addressed the issue and replicated performance similar to the original paper.
• FedFAIM:
  • Approach: add gradient aggreation, contribution assessment, reaward allocation on fedavg.
  • Challenges: Encountered issues with inconsistent weights during model aggregation.
  • Outcome: Successfully addressed the issue and replicated performance similar to the original paper.
• FairAvg:
  • Approach:
  • Challenges:
  • Outcome:
• FedQD:
  • Approach:
  • Challenges:
  • Outcome:

运行结果 虽然算法部分已经和原文对齐，但实验结果与原文对比，有以下待解决问题：

1. FedFAIM算法到最后的表现远不如其他的算法，需要搞清楚是不是梯度定制的问题
2. niid场景下所有算法的精度过于高，需要检验

贡献评估

• CI:
  • Approach:
  • Challenges:
  • Outcome:
• TMC-Shapley:
  • Approach:
  • Challenges:
  • Outcome:
• GTB:
  • Approach:
  • Challenges:
  • Outcome:
• COS-SIM:
  • Approach:
  • Challenges:
  • Outcome:

梯度定制

• CFFL:
  • Approach:
  • Challenges:
  • Outcome:
• RFFL:
  • Approach:
  • Challenges:
  • Outcome:

📓 复现记录

Paper: Improving_Federated_Learning_With_Quality-Aware_User_Incentive_and_Auto-Weighted_Model_Aggregation

实验设置

• 模型:
  • Approach:
  • Challenges:
  • Outcome:
• 数据集:
  • Approach:
  • Challenges:
  • Outcome:

自动权重的模型聚合

• FedQD:
  • Approach:
  • Challenges:
  • Outcome:

在线客户本地学习质量的估计

• FedQD:
  • Approach:
  • Challenges:
  • Outcome:

多任务场景下质量敏感的客户激励选择

• FedQD:
  • Approach:
  • Challenges:
  • Outcome:

FAIR全配置在niid场景下：客户数30；样本量[100,1000]；报价[1,3]；预算20

License

该项目基于Apache-2.0 License许可证开源
This project is released under the Apache-2.0 License.