Heterogeneous model architecture federated learning method and system in edge computing

Abstract

The invention provides a federated learning method and system for a heterogeneous model architecture in edge computing. The method comprises the following steps: training respective models locally by a federated learning client; the client uploads model parameters to the central server after the current training round is finished; the central server aggregates the model parameters uploaded by the client; training a shared generator according to the model parameters uploaded by the client; after training is completed, aggregated model parameters and the trained generator are sent to each client so as to carry out the next round of training; and repeatedly executing until the model converges or reaches a target training round. According to the method, factors such as the computing power, the storage resources and the network bandwidth of each node are considered in the federated learning process, meanwhile, an additional generator is introduced to guarantee feature alignment among the heterogeneous models, the utilization efficiency of the overall computing resources can be effectively optimized, and while the performance of the individual nodes is guaranteed, the overall performance of the heterogeneous models is improved. And the average performance of the global model is improved.

Images