Research On Efficient Federated Learning Algorithm Based On Synchronous And Asynchronous

With the wide application of artificial intelligence technology in unmanned driving,medical and financial fields,a large number of intelligent mobile devices are connected to the network,and these devices will produce large-scale and valuable data in the process of operation.Traditional deep learning model needs to collect data on each device and train intensively,but such data storage mode will have data island and privacy security problems.Federated learning can solve this problem well.It is a distributed machine learning method,which can schedule mobile devices to use their own data to independently train the model locally,and then upload the trained model parameters to the server to form a new model.However,there are still some problems in the research of Federated learning algorithm:firstly,the mainstream federated learning algorithm uses synchronous communication scheduling in the process of communication between server and client.Although synchronous communication has higher prediction accuracy and more stable model convergence,it has the problem of communication congestion.To solve this problem,some researchers combine asynchronous communication with federated learning,but the upload of models on local devices will still be delayed,which leads to the fluctuation of model prediction accuracy.Secondly,training models on mobile devices are also faced with the problem of data noise,but most of the data sets used in the current federated learning algorithm research are well handled public data sets,and there is no problem of data noise in the actual scene.To solve these problems,this paper proposes a federated learning algorithm HySync based on synchronous and asynchronous.This method will improve the training efficiency of the model from two aspects of scheduling mode and data denoising.The main tasks of this paper are as follows:(1)HySync combines the advantages of synchronous and asynchronous methods.Experimental results show that HySync can update the global model flexibly without waiting for each selected client to complete the training work,and the training time is 33% shorter than the synchronization algorithm.At the same time,in the case of high concurrent updates,HySync can control the version difference caused by the delay of updates to 50% of the asynchronous algorithm,so the prediction accuracy is 5% higher than the asynchronous algorithm,and not lower than the synchronous algorithm.(2)In this paper,the method of data denoising in deep learning is applied to federated learning,and a neural network layer is used to solve the problem of data noise on equipment.The experimental results show that the prediction accuracy of the model is improved by 5% in the presence of noise.