Participant Selection And Efficient Communication Mechanism Based On Federated Learning In Mobile Crowd Sensing

Mobile crowd sensing offers low cost,wide coverage and easy maintenance features,setting it apart from traditional static sensor networks.By utilizing users’ mobile intelligent terminals as sensing units,mobile crowd sensing achieves task distribution and data collection.The collected data can be used to create centralized databases for training advanced artificial intelligence models,which provide various services.The improvement of intelligent terminal computing capabilities and the development of federated learning technology provide a new distributed paradigm for mobile crowd sensing.For mobile crowd sensing tasks involving large-scale data collection and numerous participants,traditional centralized data collection methods face issues of low task completion quality and excessive concurrent communication pressure on servers.To address these problems,this thesis integrates federated learning with mobile crowd sensing,designing participant selection and efficient communication mechanisms that better evaluate,select,and utilize the computing capabilities of participants’ mobile terminals.(1)To address the aforementioned participant selection issue,a Participant Selection Algorithm Based on Reputation Evaluation and Resource Prediction(PSRERP)is proposed.This allows mobile crowd sensing platforms to evaluate the quality of participants’ data without requiring access to the original data,thereby optimizing participant selection.Initially,considering the resource level of the participants’ smart terminals and their current interaction states,an evaluation mechanism for these terminal resources is constructed.Subsequently,an intelligent terminal resource prediction model based on linear regression and Long Short-Term Memory(LSTM)networks is proposed,aiming to enhance the success rate of task completion.Subsequently,by pre-training and testing the model,it assesses the quality of data provided by the participants.Combined with historical task completion data,a participant reputation assessment model is established to dynamically evaluate and select participants.Simulation results demonstrate that the proposed participant selection algorithm exhibits superior performance in terms of task completion quality,energy consumption,communication rounds,and task completion time.(2)To alleviate the problem of excessive concurrent server communication pressure in mobile crowd sensing task scenarios,A Communication Optimization Algorithm based on Hierarchical Federated Learning in Mobile Crowd Sensing(COHFL)is proposed.This algorithm first considers factors such as the computing power of the intelligent terminal carried by the participant,battery life,network status,and participant intent to build a cluster head election strategy.Then,to address the issue of participant mobility in mobile crowd sensing,it builds a mobility prediction model based on Markov Chains,using the historical mobility data of the participants.A participant clustering algorithm based on mobility prediction is proposed to minimize losses due to re-clustering.Additionally,to reduce the transmission of inefficient models and minimize redundant communication,a method for evaluating the effectiveness of participants’ local models is proposed,further enhancing communication efficiency.Through simulation experiments,the algorithm proposed in this study is compared with similar algorithms.It achieves equivalent accuracy on the MNIST and CIFAR-10 datasets while reducing the global communication rounds by6.7% and 10% respectively,and decreasing the total number of intra-cluster communications by 7% and 16% respectively.