Research On Resource Allocation And Privacy Protection In Federated Edge Learning

In recent years,with the development of the 5th generation mobile communication technology,people’s life has made a qualitative leap,and the society has ushered in a new information age.At the same time,a large number of emerging industries have emerged,such as autonomous driving,virtual reality,smart city,energy consumption,etc.,these applications have high requirements for bandwidth,delay and computing energy consumption,which are far beyond the capacity of traditional communication modes.Although the traditional cloud computing model can meet some of the needs of users,it still can not meet the requirements of emerging industries due to the limitation of communication resources such as bandwidth.Thus,mobile edge computing(MEC),which can provide mobile devices with low delay and strong reliability at the edge of the network,has emerged.With the continuous development of edge computing,users have higher requirements for data privacy protection.However,the traditional central computing mode requires users to upload their original data to the central server,which is easy to lead to the disclosure of users’ privacy.In order to solve the problem of users’ privacy and security,a new learning mode,federated learning(FL),is proposed.In federated learning,users only need to calculate the local model through local data,upload the local model to the edge server without uploading the original data.The edge server aggregates and updates the received local models to obtain the global model,and then broadcasts the updated global model to each user.Resource allocation and user selection affect the training efficiency of federated learning.Users participating in training also need better privacy protection mechanism to ensure the privacy and security of their own data.This thesis considers the communication resource allocation and user selection in federated edge learning and the power optimization problem of introducing differential privacy into overthe-air computation.The main work is as follows:First,in the context of federated learning,combined with wireless communication resource allocation,select users with large size of dataset to participate in the training to speed up the learning process and reduce the consumption of communication resources.The optimization algorithm proposed in this thesis transforms the non-convex optimization problem into a nonlinear fractional programming problem,optimizes the communication time and calculation frequency by Lagrange multiplier method,and selects users combined with the linear function optimization method.Simulation results show that the proposed algorithm can not only speed up the learning process,but also greatly reduce the cost of the network.Second,in the context of federated learning,differential privacy mechanism is introduced into over-the-air computation to provide users with better privacy protection,and the power allocation scheme and denoising factor are optimized to minimize the mean square error of calculation.In this thesis,the proposed non-convex problem is transformed into two subproblems and the optimal results of the two subproblems are iterated repeatedly until the mean square error of the two subproblems tends to the same.The simulation results show that the proposed scheme can optimize the power control scheme and effectively reduce the mean square error of calculation.