| Federated learning realizes the jointly training of the deep neural network model among different participants without sharing local data,which has become an important research directions in the field of artificial intelligence.Cellular-connected Unmanned Aerial Vehicle(UAV)based on federated learning realizes the collaborative training of the globle model while ensuring the security of private data.However,the traditional algorithm of federated learning faces many challenges,such as data heterogeneity and equipment heterogeneity,and is difficult to be directly used in cellular-connected UAV.First of all,the data distribution of different participants is quite different,which would reduce the accuracy of the global model.Besides,single global model is difficult to meet the personalized needs of all participants.Personalized federated learning can generate high-quality personalized models for each participant,and is an important methods to solve data heterogeneity.In addition,the cellular-connected UAV as participants has the problem of limited battery power.Reasonable resource allocation can effectively reduce the computing and communication costs of federated learning and prolong the survival time of UAVs in the system.Therefore,thesis studies a personalized federated learning algorithm for cellular-connected UAV and the corresponding computing and communication resource allocation algorithm.The main work of thesis is as follows:1)Reviewed the related research of federal learning.First,the architecture and classical algorithm of federated learning is introduced.Then,personalized federated learning is summarized aiming at the problem of data heterogeneity.Finally,the resource allocation of UAV in mobile edge computing and federated edge learning are introduced.2)To solve the problem of global model accuracy degradation caused by data heterogeneity in federated learning,a personalized federated learning algorithm for cellular-connected UAV is proposed.First of all,we use the attention mechanism to get the similarity of the model parameters between each participant,and aggregate the personalized model parameters for each participant with strong generalization according to the similarity weight.Then,each participant will adaptively integrate the local model and personalized model to further improve the personalization and accuracy of the model.The performance of the proposed algorithm are verified by experiments.The experimental results show that the proposed algorithm can effectively improve the accuracy and convergence speed of the personalized model.3)To solve the problem of limited computing and communication resources in the federated learning scenario for cellular-connected UAV,a computing and communication resource allocation algorithm for federated learning of UAV is proposed.First,we combine federated learning with mobile edge computing,and replace the central server with edge servers to reduce the communication overhead of federated learning.The feasibility and effectiveness of the proposed algorithm are verified by experiments.Simulation results show that the proposed algorithm can effectively reduce the computing and communication costs of federated learning,and achieve a compromise between delay and energy consumption through weight changes. |
