Multi-index federated aggregation algorithm based on trusted verification

Motivated by the modern phenomenon of distributed data collected by edge devices at scale, federated learning can use large amounts of training data from diverse users for better representation and generalization. To improve flexibility and scalability, a new federated optimization algorithm, named multi-index federated aggregation algorithm based on trusted verification (TVFedmul) is proposed. It overcomes a series of problems caused by the original aggregation algorithm, which only takes the single index of data quantity as a reference factor to measure the aggregation weight of each client. The improved algorithm is based on multi-index measurement. It reflects the comprehensive ability of clients more thoroughly, to make overall judgments. Further, to achieve customized federated learning, a hyperparameter α is introduced. It can be changed to determine the importance of indexes. Finally, via extensive experimentation, it has been observed that the improved algorithm is faster, and the accuracy reaches 94.59%, which is 2.53% higher than that of FedAvg (92.06%).