Research On Computation Offloading Based On Federated Learning In UAV-Assisted Mobile Edge Computing

With the rapid development of Internet of Things(IoT),emerging IoT applications have put forward higher and higher computing power demand for resource-constrained mobile devices.Facing the huge demand of massive IoT devices for network access and dynamic communication requirements,fixed deployed edge base stations are vulnerable to the network traffic overload,which reduce the quality of user experience.UAV-assisted Mobile Edge Computing(MEC)is an effective way to solve this problem.In the case of traffic explosion in the temporary hot spot area and damage to the base stations in the disaster areas,carrying edge servers as air base stations,UAV can provide agile MEC services for mobile devices by using UAV’s better line of sight,wider communication,more flexible on-demand deployment and lower deployment cost.By offloading the computing intensive tasks of users to the nearby MEC servers,mobile edge computation offloading is considered as a key technology in MEC to realize low-latency services and prolong the battery life of users.However,the communication and computing resources of UAVs and edge base stations are limited.How to optimize computation offloading decisions,so as to reduce task response delay and user energy consumption,has become an urgent problem to be solved.Moreover,traditional offloading methods based on optimization theories are often non-scalable,time-consuming,and high computational cost,while learning-based centralized computational offloading methods require IoT devices to share their data to train the global model,which will infringe on the user’s privacy and generate unnecessary communication overhead on the scarce spectrum.In order to solve the above problems,Federated Learning(FL)is proposed as a distributed collaborative learning method,which can be deployed on the edge devices of the MEC system,providing a suitable solution for realizing network edge intelligence.However,the success of federated learning largely depends on the participation of nodes,and the amount of data for users to train high-quality learning model is limited.If there is no satisfactory reward,users and UAVs will not be willing to participate in federated learning with the cost of computing and communication resources.Based on it,this thesis has carried out researches on the incentive mechanism of federated learning and the computation offloading scheme based on federated learning in the UAV-assisted mobile edge computing environment.Specifically,the main work of this thesis include:(1)In the UAV-assisted mobile edge computing environment based on federated learning,in view of the lack of appropriate ways to encourage users to participate in federated learning,and the low accuracy of federated learning caused by the random selection of federated learning participants,a federated learning incentive mechanism method based on Stackelberg game is proposed in this thesis.Firstly,the system framework of UAV-assisted mobile edge computing based on federated learning is introduced and the problem of incentive mechanism of federated learning is analyzed.Then,the selection of federated learning participants and the training process of hierarchical federated learning are described,and the hierarchical federated learning utility model and Stackelberg game model are constructed.Finally,a federated learning incentive mechanism algorithm based on Stackelberg game is proposed.In order to carry out experimental verification and comparative analysis of the proposed algorithm,a UAV-assisted mobile edge computing environment based on federated learning is built in this thesis,and the proposed federated learning incentive mechanism algorithm based on Stackelberg game is compared with the RandFL algorithm,the PriceFFL algorithm,and the SPFL algorithm.Experimental results show that the proposed algorithm can motivate user devices with high-quality data to participate in federated learning training,improve the training accuracy of federated learning,as well as maximize social welfare.(2)In the UAV-assisted mobile edge computing environment,facing the problem that the computation offloading strategy lacks dynamic scalability,the centralized training method leads to excessive transmission overhead,and user privacy cannot be guaranteed.A computation offloading method based on federated deep reinforcement learning is proposed,aiming to achieve dynamic adaptive computation offloading while protecting users’ privacy.Firstly,the problem of computation offloading based on federated learning in the UAV-assisted mobile edge computing environment is analyzed.Then,the communication model,computation offloading delay and energy consumption model are constructed,and the objective function to minimize the normalized weighted cost of task response delay and energy consumption of user equipment is built.In this thesis,the execution cost on the mobile device side and the time-varying network conditions on the MEC side is comprehensively considered,and the dynamic computation offloading decision problem is modeled as a Markov decision process.A federated deep reinforcement learning algorithm is used to solve the approximate optimal computation offloading strategy.Finally,the proposed federated deep reinforcement learning algorithm is compared with the Greedy algorithm,the AC algorithm,and the DQN algorithm.Experimental results show that the proposed algorithm can obviously reduce the task response delay,energy consumption of user equipment and the weight cost of the system.