FedAcc Algorithm Based on Client Accuracy in Federated Learning

In the FedAvg (Federated Averaging) algorithm of Federated Learning, the server primarily updates parameters by performing a weighted average of local model parameters based on the size of the client datasets, giving more weight to clients with larger datasets. In practical applications, clients with larger datasets do not necessarily contribute the most to the model. To address this issue, this paper designs a FedAcc (Federated Averaging with Accuracy-based Weighting) algorithm based on client accuracy and conducts experimental comparisons to analyze the performance of FedAcc and FedAvg. In the experiments, four different types of datasets were selected: image datasets MNIST and CIFAR-10, text dataset IMDB, and regression task dataset California Housing. Two data distribution scenarios were set up: Independent and Identically Distributed (IID) and Non-Independent and Identically Distributed (Non-IID). The experimental results show that in the IID scenario, the FedAcc algorithm achieves higher accuracy on all datasets compared to FedAvg. In the Non-IID scenario, FedAcc performs better in linear regression prediction and color image classification, particularly on the CIFAR-10 dataset, where the accuracy of FedAcc is 35% higher than that of FedAvg. Through experimental analysis, it was found that when there is a certain difference in client accuracy, the FedAcc algorithm performs better than FedAvg. In practical applications, there are usually differences in client accuracy, so FedAcc is more practical than FedAvg.