Research And Implementation Of Federated Multitasking Learning Method Based On An Asynchronous Update

Federated learning is a machine learning paradigm that enables individual clients to jointly learn a shared global model while storing their training data locally rather than uploading it to a central server to complete the training of a global model under the condition of protecting data privacy.However,two challenges affect federated learning performance.Firstly,the data distribution of each client is inconsistent,which affects the single global model that cannot meet the needs of clients with different data distribution,and then affects the performance of the model.Secondly,there are differences in device performance between other clients,and they have different computing speeds,which prolongs the training time and affects the training performance.The above two challenges often appear simultaneously,but many researchers only address one,almost no research has considered and addressed the above two challenges simultaneously.Thus,it is worth studying how to alleviate the data heterogeneity and device heterogeneity at the same time.To solve the above problems,this paper proposes a clusterasynchronous federated multi-task learning,divided into four steps.Firstly,a clustering method based on the data distribution and device performance of clients,which is used to separate the clients with similar data distribution and device performance into the same cluster.The second step,the data feature calculation and extraction method,calculated and extracted the global important features.The third step,a global federated learning method based on feature masking.All clients mask non-important features in the data and perform global federated learning to update and share the model weights related to important features.In the fourth step,the cluster asynchronous federated learning method,clients perform federated learning in their respective clusters,where the intra-cluster training is synchronous,while the clusters do not interfere with each other,and their updates are asynchronous.Experimental results show that our algorithm achieves better model performance and training performance on benchmark datasets compared with existing algorithms.Furthermore,this paper designs and implements a clusterasynchronous federated multi-task learning system.The system is divided into three modules.The first is the information display module,which is mainly responsible for the display of task information and model result information,allowing users to view relevant task and model result information.The second is the task transceiver module,which is mainly responsible for task publishing,receiving and starting functions,allowing users to publish and participate in federated learning tasks.The third is the federated learning module,which is mainly responsible for importing the task training data,performing the proposed method,model accuracy evaluation,and communication interaction between the client and server.