Research On MEC Offloading Strategy Based On Deep Reinforcement Learning And Physical Layer Securit

With the increasing number of mobile devices and Internet of Things(Io T)devices,more and more computation-intensive and latency-sensitive applications have emerged,and higher requirements for computation and communication resources have been put forward.However,performing local computation is often limited by devices’ own computational capability,communication resources,and batteries.Traditional cloud computing chooses to offload tasks to the core network and processes them on a central cloud server,resulting in high latency and network congestion.In order to achieve wireless transmission with low latency,high energy efficiency,and high reliability,researchers have proposed mobile edge computing(MEC),which deploys the storage and computing services originally located in the cloud data center to the edge of the mobile network and provides computing,storage and communication resources at the edge of the mobile network.Due to the complexity of the MEC environment and the openness of wireless networks,how to design and quickly obtain an offload strategy to improve the performance of the MEC system is a key issue that needs to be addressed urgently.Therefore,this thesis focuses on task offloading and model offloading in MEC.Specifically,from the perspective of physical layer security,this thesis optimizes task offloading and system resource scheduling while ensuring system security.Firstly,this thesis studies the task offloading in MEC networks under interference attacks from malicious unmanned aerial vehicles(UAV),where multiple users process computational tasks with the assistance of a computational access point(CAP).Considering the dynamic changes in the MEC environment,deep reinforcement learning and transfer learning are used to optimize system performance to reduce latency and energy consumption.Specifically,deep reinforcement learning is used to provide offloading strategies that can improve system performance,while transfer learning is used to accelerate the training process and improve the performance of reinforcement learning.Simulation results show that the proposed offloading strategy can outperform traditional offloading strategies,and using transfer learning can significantly reduce training time while achieving better system performance.Secondly,based on the above research,this thesis further studies task offloading in MEC networks under multiple eavesdropping attacks,where K eavesdroppers exist in MEC networks,threatening the safety of task offloading.Eavesdroppers can work in the collusion or the non-collusion mode,where the eavesdroppers collaborate to decode messages in the collusion mode,or they decode messages individually in the non-collusion mode.For the two eavesdropping modes,this thesis designs a secure MEC system by optimizing the task offload ratio,transmit power,and computational capability allocation to improve the latency performance of the system.In particular,a deep reinforcement learning and convex optimization based algorithm is proposed to solve the optimization problem,in which deep reinforcement learning is used to decide the task offloading ratio,while convex optimization is used to solve the allocation of transmit power and computational capability.Simulation results show that the proposed deep reinforcement learning and convex optimization based algorithm is superior to other traditional methods,and can provide guaranteed system latency performance.Finally,this thesis studies the model offloading in a relay-assisted federated edge learning system,where N users collaborate to train the model with the assistance of M relays and an edge server.Specifically,in order to fully utilize relay nodes,it is proposed to perform partial aggregation and spectrum resource multiplexing at the relays.Moreover,in order to analyze the system performance,this thesis derives analytical and asymptotic expressions of the system outage probability and convergence rate.In further,two bandwidth allocation schemes are proposed to maximize the number of successfully offloaded models,thereby further optimizing system performance.Simulation shows the test accuracy and convergence speed of the relayassisted federated edge learning,and the proposed bandwidth allocation scheme is superior to other benchmarks.