Research On User Scheduling Policy For Reliable Federated Learning In Wireless Networks

As a distributed machine learning method,federated Learning(FL)allows several participants to jointly train a machine learning model,which can effectively avoid the leakage of user privacy caused by the centralized use of original data to train models in traditional machine learning methods,and has broad application prospects in finance,medical care,education and other fields.However,in wireless network environments,since the edge devices participating in FL are usually very geographically dispersed,making their identities difficult to authenticate.At the same time,due to the uncontrolled local training process of edge devices,the FL model is vulnerable to malicious attacks,which will greatly impair the convergence performance of the global model.Therefore,this dissertation aims to solve the unreliable problem of FL caused by data poisoning attacks from malicious users in wireless networks,and proposes a reputation-based user scheduling policy,builds reliable FL frameworks for wireless networks.The main work of this dissertation can be briefly described as follows:1.Construct the working mode of the wireless FL system under the client-server(C/S)architecture and clustered-net(Clus)architecture.Propose a reputation model based on beta distribution function to measure the credibility of edge devices,which classifies a user's behavior into a binary form of positive behavior and negative behavior,and the system calculates the reputation value of a user by evaluating the contribution of different behaviors to the global model.Propose a reputation-based user scheduling policy(RP),which prioritizes users with higher reputation and takes into account the fairness of scheduling among users.The policy not only reduces the probability of malicious users being scheduled,but also improves scheduling fairness among trusted users.2.Based on the proposed user scheduling policy,this dissertation proposes the reputationbased reliable FL algorithms for C/S architecture and Clus architecture,respectively.Specifically,for C/S architecture,the base station acts as the central server to maintain and update the reputation table of all users in the cell,and adopts the RP scheduling policy to determine the users scheduled for each round;For Clus architecture,the dissertation design a cluster head selection algorithm and network clustering criterion(ELEACH),which comprehensively considers factors such as the remaining energy of equipment,regional average energy,equipment performance and user reputation.Through ELEACH,all users maintain the reputation table of the whole network in a distributed manner,which constructs an efficient,energy-saving and reliable FL framework.The dissertation also deduces the theoretical expression of the proposed algorithms' convergence rate when malicious users exist,and analyze the influence of key parameters such as the percentage of malicious users and average attack intensity.3.Build a simulation experiment platform,analyze and verify the performance of the proposed reliable FL algorithm through numerical analysis and simulation under C/S architecture and Clus architecture.The results show that the proposed user scheduling policy RP can effectively identify malicious users in the network and resist different types of data poisoning attacks.Compared with the three classical scheduling policies(RS,RR and PF),the convergence performance of the FL algorithm based on the RS,RR and PF policies deteriorates sharply in the presence of malicious users,while the RP policy proposed in this dissertation can still make FL maintain a high convergence rate.Further,this dissertation verifies the superiority of the proposed clustering rule(ELEACH)and user scheduling policy in improving the network lifetime and resisting data poisoning attacks under the clustered FL framework.