Design And Implementation Of Asynchronous Federated Learning System For Dynamic Aggregation

Federated learning is a distributed machine learning paradigm with privacy protection as the premise.Users first complete local model training,then share local model parameters and obtain a high-precision global model through aggregation.Federated learning usually uses synchronous aggregation for model aggregation,that is,after collecting the local models of all clients,the global model aggregation is completed together.However,in practice,there are differences in data volume and processing capacity between clients,which makes the training time-inconsistent,which leads to problems such as long aggregation time and low aggregation efficiency in the global model,which seriously affects the performance of federated learning.In order to deal with the above problems,related work proposes that an asynchronous aggregation method can be adopted,that is,it is not necessary to wait for all clients to complete the training,and global aggregation can be performed whenever the training of a client is completed,thereby improving the efficiency of global model aggregation.However,the existing asynchronous federated learning research work only selects a fixed number of clients per round during global aggregation,which makes some clients do not participate in the aggregation on time,resulting in the outdated local model of the client,thus affecting the global model accuracy of federated learning.On the other hand,although federated learning can keep the data out of the local area,there are certain security risks in the processing process.There may be a malicious server to identify the source of model parameters,and even the original data can be inferred from the model submitted by the client multiple times.Therefore,it is necessary to introduce a reasonable privacy protection mechanism for the federated learning process.Based on the above analysis,this thesis designs and implements an asynchronous federated learning system for dynamic aggregation,which effectively solves the model outdated problem caused by the unreasonable number of aggregation clients and the privacy leakage problem caused by malicious server attacks.The main contributions of this thesis are as follows:(1)Design and implement a dynamic aggregation scheme based on training time clustering,estimate the local training time according to the client’s local data and computing power,and then classify the client through the K-means clustering algorithm to dynamically determine the aggregation.The number of clients solves the outdated problem of the local model and improves the accuracy of the global model.(2)Design and implement a differential privacy-based federated learning privacy protection scheme.Privacy protection is achieved by adding Gaussian noise to the gradient of the local model,and the differential privacy technology features low communication overhead and easy deployment to solve the privacy leakage problem caused by malicious servers speculating on client models.(3)A dynamic aggregation-oriented asynchronous federated learning experimental scenario is built,and the system is tested.The test results show that compared with the traditional asynchronous federated learning aggregation algorithm FedAsync,the system global model accuracy is improved by 5.6%.