Research on federated model algorithm based on non-independent identically distributed data sets

Federated learning is a distributed machine learning framework based on privacy protection. Heterogeneity is unavoidable in distributed data sets. When the data sets of different participants are not independent and identically distributed, the performance of the existing federated learning algorithm is inevitably affected. At the same time, the uneven distribution of participants' own data sets will also aggravate the deterioration of federated learning performance. Therefore, based on the inconsistent distribution of the training set and test set and uneven distribution of every participant training set labels, this paper proposes an SPFL algorithm to solve this problem and compares it with the existing FedAvg, FedProx, and FedDC under the same experimental conditions. Experimental results show that the accuracy of the SPFL model is significantly better than other algorithms when the participant data set is non-independent and uniformly distributed. The optimal accuracy of SPFL is at least 5% higher than the other three algorithms.