Research On Federated Learning For Different Structural Type

With the legislative protection of personal privacy data in various countries,the problem of data-isolated islands has become prominent.In this regard,federated learning is the best choice.There are some unsolved problems in traditional federated learning,such as data distribution,generalization of the global model and reliable central server.This thesis aims to overcome these shortcomings and solve these problems:(1)The problem of data distribution: In real life,data generally exist in the form of the Non-Independent Identically Distributed(Non-IID),which increases the degree of discretization between data,and the feature extraction of data is difficult.To solve the problem of the Non-IID,this thesis gives different solutions in Chapter 3 and Chapter 4,respectively.The third chapter proposes the DCMT model,which uses dilated convolution to obtain a large receptive field,integrates the Transformer mechanism to extract the node features,and strengthens the feature extraction of the Non-IID data.In the fourth chapter,this thesis proposes the M-Res Net model,which adds the attention mechanism to Res Net18,and the deeper network model strengthens the feature extraction of non-independent and identically distributed data.(2)The problem of the generalization of the global model: the existing federated learning algorithms are improved for specific problems,and the generalization performance of the global model could be better.In the third chapter of this thesis,the Fed MMD algorithm is proposed.Before the model aggregation,the client models are compared,and the domain differences between the models are given different weights.The global model obtained by the Fed MMD algorithm has higher generalization performance.(3)The problem of the reliable central server: The existing federated learning already has a decentralized algorithm.It does not need a reliable central server to calculate the global model,but its security and efficiency are low.In the fourth chapter of this thesis,the DFLBC algorithm is proposed,which does not need a reliable central server.The global model of federation learning is established based on blockchain technology,which can ensure the security of the framework.At the same time,in order to improve the global model performance of the DFLBC algorithm,different clients are given different weights according to the Markov property so that the DFLBC global model is improved to some extent.The experimental data show that two federated learning frameworks proposed in this thesis can alleviate the shortcomings of existing algorithms.The Fed MMD algorithm is proposed for centralized federated learning,and a global model with strong generalization ability is obtained.In decentralized,federated learning,the DFLBC algorithm is proposed,which can get a global model comparable to that of centralized servers without relying on centralized servers.