rruisong / H2-Fed

PyTorch implementation of H2-Fed Federated Learning Framework

Federated Learning Framework Coping with Hierarchical Heterogeneity in Cooperative ITS

H2-Fed is a federated learning framework addressing hierarchical heterogeneity in the different layers of cooperative Intelligent Transportation Systems (C-ITS).

Even when 90% (CSR=0.1) of the agents are timely disconnected, the pre-trained DNN model can still be forced to converge stably, and its accuracy can be enhanced from 68% to over 90% after convergence.

Framework Overview

Example Results on Non-IID MNIST:

Dependency

• Python (>=3.7)
• PyTorch (>=1.9)
• sklearn (<=0.20)
• numpy
• json
• OpenCV

Data Preparation

The training parameters can be edited in 'H2-Fed/config/*'

1. cd H2-Fed/data
2. python mnist_preparation.py
The data for train and test will be saved in 'data/mnist/'

Training Procedures

Example I: FL with individual parameter set

1. cd H2-Fed
2. python  main.py \
   -g 60 \
   -l 5 \
   -c 0.1 \
   -s 1 \
   -f 1 \
   -m 0.001 0.001 \
   -ei 10 10 \
   -rd 'rsu_config_NonIID_RSU' \
   -rc 'rsu_config_NonIID_RSU' \
   -la 'H2Fed' \
   -sm True \
   -bt True \
   -pt True \
   -rr 'results' 

Example II: Batch FL

1. cd H2-Fed/config
2. python main.py \
   -rd 'results' \
   -b True
Edit the training parameters in 'H2-Fed/config/batch_sim_config.py'

Explanations of arguments:

• --GAR -g: Global Aggregation Round
• --LAR -l: Local Aggregation Round
• --CSR -c: Connection Success Ratio
• --SCD -s: Stable Connection Duration
• --FSR -f: Full-task Success Ratio
• --mu -m: Core parameters M in framework
• --epoch_init -ei: Init epoch wrt. FSR
• --res_dir -rd: Directory of results under root directory
• --rsu_config -rc: RSU configuration dict data in ./config/*
• --label -la: Label for result data
• --batch -b: True, if the existing scenarios are batched run
• --save_model -sm: True, if the model should be saved
• --base_train -bt: True, if the centralized training should be implemented first
• --pre_train -pt: True, if the pre-training should be implemented first
• --res_root -rr: Root directory of the results

Evaluation procedures

1. cd H2-Fed/post_processing
2. python main_plot.py

Citation

If you find this work is useful, please cite our paper:

@INPROCEEDINGS{song2022h2fed,
  author={Song, Rui and Zhou, Liguo and Lakshminarasimhan, Venkatnarayanan and Festag, Andreas and Knoll, Alois},
  booktitle={2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC)}, 
  title={Federated Learning Framework Coping with Hierarchical Heterogeneity in Cooperative ITS}, 
  year={2022},
  pages={3502-3508},
  doi={10.1109/ITSC55140.2022.9922064}}

Authors:

• Rui Song (linkedin)
• Liguo Zhou
• Venkatnarayanan, Lakshminarasimhan
• Prof. Dr.-Ing. Andreas Festag (linkedin, web)
• Prof. Dr.-Ing. habil. Alois C. Knoll (linkedin, web)