Design And Implementation Of Personalized Federated Learning Algorithm

Recently,with the emergence of data island and users’ concerns about the security of private data,the traditional centralized machine learning technology has encountered constraints.While as a special distributed machine learning paradigm with privacy protection,federated learning can decentralized clients collaboratively learn a model while protecting users’ data privacy.Since federated learning was proposed,it has been closely focused on in both industry and academia,but there are still many challenges in federated learning.For the challenge of data heterogeneity,personalized federated learning has been regarded as an effective method because it can customize to process information for each client and train a personalized model with strong performance for each participant.However,the existing personalized federated learning methods to overcome data heterogeneity still have a large room for improvement in classification tasks.In personalized federated learning,the local information of each client and the global information between different clients can be further utilized.In order to deal with the above problems,we propose corresponding improvement methods:1.Aiming at the problem of underutilized global knowledge,we propose a personalized federated learning model based on prototype contrastive learning.In the scenario of data heterogeneity,we utilize the output representation of the information enhancement layer as the data prototype,and calculate the contrastive loss with the global prototype representation of all categories.This design can make full use of the global knowledge as well as alleviate the impact of data heterogeneity,which improve the performance of the model.We evaluate the effectiveness of the model through sufficient experiments.2.Aiming at the lack of personalization in parameter aggregation,we propose a new personalized parameter aggregation strategy.The traditional federated learning paradigm applied parameter aggregation according to local data capacity,which lacks personalization.In this paper,we propose a personalized parameter aggregation strategy,which utilizes the client-level prototype information to calculate the similarity between clients,so as to achieve personalized parameter aggregation.The effectiveness of our parameter aggregation strategy is evaluated through sufficient experiments.3.Aiming at the balance between local personalization and global generalization,we propose a personalized federated learning model based on multi-task learning.The previous researches on personalized federated learning weakened global generalization ability while pursuing individualization ability.Therefore,we propose a personalized federated learning model based on multitask learning.The model learning process is divided into two tasks: local information learning and global information learning.By constructing expert networks and gated networks,the model can automatically balances the local information and global information used in local model training,and fully explores the implicit relationship between local data distribution and global distribution,which improves the performance of personalized federated learning model.We verify the effectiveness of our model through sufficient experiments.In generally,to improve the shortcomings of the existing personalized federated learning methods in data heterogeneous scenarios,in this paper,we propose a personalized federated learning model based on prototype contrastive learning,a new personalized parameter aggregation strategy and a personalized federated learning model based on multi-task learning.Extensive experiments verify the feasibility and effectiveness of the proposed methods.The research results have certain application value in the field of personalized federated learning.