| Deep learning technology has attracted wide attention in various visual tasks.The success of deep learning largely depends on a large amount of valuable data.The data collected in a single domain is usually insufficient and has certain limitations,while distributed sharing involves privacy security issues,and the traditional centralized machine learning framework has been unable to meet the needs of data privacy protection.In recent years,the proposal of federated learning has solved the above problems.The core idea is that the client data is stored locally without uploading,and the global model is obtained through the coordination training of the central server,thus protecting the privacy and security of the local data.At present,federated learning has been widely applied in the fields of medicine,finance,artificial intelligence and so on.It is a research hotspot in recent years.However,the global model in federated learning has the problem of poor personalized performance.Therefore,how to train personalized models suitable for local data distribution for clients is one of the challenges faced by federated learning.In heterogeneous data scenarios,the current personalized federated learning methods mainly rely on clients downloading aggregate models from the server for local update or sorting similar clients for training to obtain personalized models,but these methods either ignore the use of similarities between users to assist training or fail to maximize the collaboration between similar users.To overcome the limitation that personalized models must remain structurally consistent,recent studies have introduced knowledge distillation into federated learning.However,in the server aggregation stage,the traditional federated knowledge distillation methods not only ignores the possibility of different quality among the model outputs uploaded by the clients,but also introduces the assumption condition that there is common data similar to the client data distribution.However,the public data is often difficult to obtain,which makes it difficult for this method to meet the conditions in practical application.In this paper,we focus on the personalized methods to federated learning,and propose a personalized federated learning method based on mutually beneficial collaboration of user similarity(FedMBC)and a personalized federated learning method based on semi-supervised distillation(pFedSD)to address the above issues.The main research contents are as follows:(1)For heterogeneous data scenario,this paper proposes a personalized federated learning method based on user similarity,which can maximize the collaboration among similar users and improve the convergence speed of model aggregation.Specifically,first,in each communication round,the similarity of users is measured using the task-layer output or the soft output of the client models on the server;Then,dynamic aggregation is performed for each user based on the similarity between users,and a personalized model is aggregated for each user to fit its local data distribution.In addition,during the client local update,the convergence rate of the model is accelerated by reviewing the knowledge of the aggregation model obtained from the previous iteration.Finally,the FedMBC method is compared with different personalized federated learning methods in different datasets.The results show that FedMBC method outperforms them in test performance and communication efficiency.In particular,FedMBC improved test accuracy by about 2.3% and reduced the number of communication rounds by up to 35% compared to FedAvg on the CIFAR-10 dataset.(2)In view of heterogeneous models and data,this paper proposes a personalized federated learning method based on semi-supervised distillation.On the one hand,it can adaptively generate data with similar distribution and remove the unreasonable premise of similar data;On the other hand,an adaptive aggregation method is designed,which can reduce the influence of low quality models on aggregation.Specifically,firstly,the method allows the clients to train the adversarial networks on the local unlabeled data to generate samples similar to the distribution of unlabeled data,and upload them to the server to obtain the shared unlabeled data before the federated learning;secondly,to reduce the aggregation weight of the soft output of the low-quality models uploaded by the clients during the federated learning,the server uses an adaptive aggregation method to adjust the aggregation weight to improve the quality of the aggregated soft output;Then,the server sends the aggregation results back to the clients for distillation training to obtain a personalized model.Finally,experiments are carried out on different image datasets to verify the effectiveness of the pFedSD.The results show that the performance of pFedSD is better than that of the federated knowledge distillation methods,especially on the MNIST dataset,and the test accuracy of pFedSD is improved by 3% on average. |
