Data-Oriented Federated Learning Research

With the wide use of smart devices,data containing user information continue to emerge.The artificial intelligence(AI)technology based on big data has made a qualitative leap.However,opportunities and challenges coexist.While AI uses data to bring benefits to various industries,user privacy data disclosure also appears.So Federated Learning(FL)arises at the historic moment.This is a special distributed machine learning framework,which can protect data privacy while fully combining multi-client training.Because any client data does not need to leave the local,such a learning framework provides a solution to the challenges faced by AI technology and is another hot research direction in the information age.This paper aims to improve the shortcomings of existing federated learning:(1)To solve the problem of Non-independent and identical distribution(Non-IID)regular data,this paper proposes a deep Generative Adversarial Network with an advanced auxiliary classifier as a pre-training module.Because the Non-IID increases the discreteness of the parameters of local models,it is difficult for FL to aggregate an excellent global model.The pre-training module proposed in this paper can deeply mine hidden features and increase the correlation between model parameters.it alleviates the problem of Non-IID;In order to solve the problem of low fault tolerance of classical FL,a federated learning framework based on Federated Kalman Filter(FKF)is proposed in this paper.Because the average aggregation algorithm in general federation learning can not identify the model parameters with noise,this paper uses the idea of FKF to propose a set of personalized adaptive confidence,to improve the fault tolerance of FL.(2)For non-Euclidean spatial data,this paper proposes a federated learning framework based on Graph Convolution Neural Networks(GCN).Because few FL frameworks can deal with non-Euclidean spatial data,this paper designs an efficient graph convolution neural network to give FL the ability to deal with non-Euclidean spatial data.Moreover,the conventional federated averaging algorithm only averages the local model parameters,which is too rough.Therefore,this paper uses the attention mechanism to assign appropriate coefficients to each local model and then aggregates the global model.This paper proposes two kinds of federated learning frameworks,which are used to deal with regular data and non-Euclidean spatial data.The federal learning framework proposed in this paper can reduce the damage caused by noise data and Non-IID data,improve the ability to deal with non-European spatial data.Finally,the effectiveness of the federal learning framework is proved by experiments.