Similarity-driven iterative aggregation algorithm for federated learning
To address the issue of model aggregation performance and Byzantine robustness in federa-ted learning,a similarity-oriented iterative aggregation algorithm is proposed.After collecting gradients uploaded by clients,the server initializes a similarity gradient,computes similarity distances between each client gradient and the similarity gradient,and between the client model and the global model of the previous round,and assigns different weights to each client according to these distances to aggre-gate a new similarity gradient.The above calculation and aggregation process is continuously repeated until the global gradient that is most similar to all clients is determined,which is used as the aggrega-tion result for the current round.Experimental results show that the aggregation algorithm can achieve excellent model performance on several datasets and a variety of neural network structures.Further-more,the robustness of the aggregation algorithm against Byzantine attacks can be effectively enhanced by adding a minimal spanning tree-based filter before aggregation.
万方数据
维普
