Fig 1.
Step 1: Server distributes the global model; Step 2: Clients train locally; Step 3: Clients upload the local model; Step 4: Server aggregates models.
Fig 2.
Crossover operator.
Fig 3.
Mutation operator.
Table 1.
Hyper-parameter settings.
Fig 4.
The accuracy of MNIST dataset under FedAvg, FedProx, and SCAFFOLD.
Fig 5.
The accuracy of FashionMNIST dataset under FedAvg, FedProx, and SCAFFOLD.
Fig 6.
The accuracy of FashionMNIST dataset under FedAvg, FedProx, and SCAFFOLD.
Fig 7.
The accuracy of MNIST dataset under FedAvg, FedProx, and SCAFFOLD.
Fig 8.
The accuracy of FashionMNIST dataset under FedAvg, FedProx, and SCAFFOLD.
Fig 9.
The accuracy of TodayNews dataset under FedAvg, FedProx, and SCAFFOLD.
Table 2.
Comparison between methods with EA and methods without EA.
Table 3.
Impact of the hyperparameter μ.
Fig 10.
Comparison of MNIST with different intervals under FedAvg, FedProx and SCAFFOLD.
Fig 11.
Comparison of TodayNew with different intervals under FedAvg, FedProx and SCAFFOLD.
Fig 12.
Comparison of FashionMNIST with different intervals under FedAvg, FedProx and SCAFFOLD.
Table 4.
Impact of the Dirichlet distribution.
Fig 13.
Visualization Comparison Between Client Model with NO_EA.
Fig 14.
Visualization Comparison Between Client Model with EA.
Table 5.
Comparison of model distances.