Research On Resource Allocation Algorithm For Federated Learning In Hybrid VLC/RF Systems

With the explosive growth of data,users and regulators are increasingly concerned about data privacy and security.Federated Learning(FL)is a distributed learning algorithm created to break data silos,which can train neural networks while protecting data privacy and has the advantages of strong security and low latency.However,the federal learning algorithm still faces the challenge of communication bottlenecks due to tight wireless resources and high communication overhead.Meanwhile,Visible Light Communications(VLC)has abundant spectrum resources,and there is no interference between VLC links and traditional Radio Frequency(RF)links.Therefore,to address the above challenges,this paper constructs a hybrid VLC/RF system by innovatively integrating VLC links with RF links,and allocates resources for horizontal federal learning and vertical federal learning,respectively.Thus,the communication efficiency of federal learning can be effectively improved and more efficient federal learning algorithms can be realized.The topic of this thesis comes from the project of "Research on the theory and technology of spatial dimming in visible light communication"supported by National Natural Science Foundation of China(Project Number:61871047).The main research contents of this thesis are following:(1)To solve the communication bottleneck problem caused by limited communication resources,large transmission data volume and unstable wireless communication in horizontal federation learning.We propose a joint optimization algorithm of user selection,bandwidth allocation and model compression for horizontal federated learning in hybrid VLC/RF system(VLC/RF-HFL).The algorithm effectively combines VLC links and RF links to build a hybrid VLC/RF system that is suitable for FL scenarios.Then the resources required to transmit FL model parameters in the wireless link are reduced by the model compression algorithm,Finally,the wireless network is optimized by user selection and bandwidth allocation.The problem is an optimization problem with the objective of minimizing the FL model training loss,which is separated into two subproblems.The first subproblem is a user selection problem with a given bandwidth allocation,which is solved by using an iterative algorithm.The second subproblem is the bandwidth allocation problem under a given user selection,and is solved by using numerical methods.The final user selection and bandwidth allocation are obtained by iteratively solving the two subproblems and performing model compression.Simulation results show that the proposed VLC/RF-HFL algorithm can improve the model accuracy of the target identification task by 16.7%and the number of selected users by 68.7%compared to the conventional FL algorithm that uses only RP links for transmission.(2)To solve the communication bottleneck problem and the high accuracy requirement of transmitted data due to the transmitted encryption intermediate value and encryption key,we propose a joint optimization algorithm of transmission power,user selection and channel estimation for vertical federated learning in hybrid VLC/RF system(VLC/RF-VFL).This algorithm first builds a hybrid VLC/RF system by combining traditional RF links and VLC links to expand the communication resources.Then a multi-layer perception(MLP)channel estimation algorithm is used to reduce the packet error rate(PER).Finally,the PER is introduced into the resource allocation problem with the objective of minimizing the FL loss function.This problem is solved by collaboratively optimizing transmission power allocation and user selection.The problem is divided into two subproblems,The optimal transmission power for each user is first derived given the set of selected users,and then the Kuhn-Munkres algorithm is used to optimize the selection of participating users.Simulation results show that the model accuracy of the proposed VLC/RFVFL algorithm is improved by 7.2%compared to the algorithm using randomized determination of user selection and resource allocation in the VLC/RF hybrid system.And the model accuracy of the proposed VLC/RFVFL algorithm is improved by 18.2%compared to the algorithm using the same setup as in this paper only without introducing the VLC link in the RF system.