Incentive Mechanism Design for Unbiased Federated Learning with Randomized Client Participation

This repository contains the codes for the publication:

[How to Recruit Clients for Federated Learning: Incentive Mechanism Design with Randomized Client Participation] B.Luo, Y. Feng, S. Wang, J. Huang, L. Tassiulas, “Incentive Mechanism Design for Unbiased Federated Learning with Randomized Client Participation”, accepted in Proc. IEEE International Conference on Distributed Computing Systems (ICDCS), 2023. https://doi.org/10.48550/arXiv.2304.07981 Our code is based on the code for fedavgpy and FedProx.

Setup network for experiment

1. Turn on Wifi router

2. Connect to Wifi

3. Config Wifi

   3.1. Open browser, and enter URL

   3.2. PORT Management -> DHCP Setting

   3.3.1. Scan devices under this wifi
   3.3.2. Check device's connect by identity

   3.3. VLAN: address binding

   3.4.1. Address of server to a fix ip 
   3.4.2. Tips: fix ip of devices and note them

do experiment

Overview

on server python $MAIN --model server

on client python $MAIN --model client

1. Preparation 1.1. add ssh-key, remove existing files, send latest code

   sh prepare.sh

   1.2. Generate data by following code and create log/ folder. More help information could be found in FedProx.

   cd Stackelberg
   python data/synthetic/generate_synthetic.py
   python data/mnist/generate_random_niid.py
   python data/emnist/generate_random_niid.py

   1.3. run Pre.py to fetch alpha for different dataset

2. To compare benchmark, activate server python main_bench.py --model server To compare properties, activate server python main_property.py --model server

3. execute code on clients

   3.1. Modify USERNAME, HOSTS and other parameters in run.sh and run sh run.sh

Notes:

1.In args of Pre.py main_bench.py main_property.py, you can modify following parameters to get different results

'dataset': dataset name
'test_num': test number
'C': cost
'budget': budget
'v': intrinsic value
'experiment_folder': folder to save result
'num_round': number of simulation
'alpha': alpha

Results in .json format contains global accuary, loss, time and other imformation.

2.Different dataset requires different parameters setting so that the solver can work sucessfully, e.g., too large cost on average may result in no solution(negative q) due to the constraint.