Reinforcement Learning Based Anti-Jamming Mobile Edge Computing For Federated Target Recognition Technique

The data offloading process in mobile edge computing(MEC)is vulnerable to jamming attacks,which affects the received signal quality of the devices and increases the energy consumption and the latency of the target recognition significantly.Traditional MEC anti-jamming data offloading scheme usually relies on the channel model and the jamming model which is difficult to ensure long-term quality of service of the target recognition applications in the dynamically communication environment.Therefore,this paper applies the reinforcement learning to study the anti-jamming MEC technology for the federated target recognition,enhances the anti-jamming ability of data offloading,improves the target recognition performance and helps to improve the user experience.First,a reinforcement learning based anti-jamming MEC data offloading and model training(RLOT)scheme is proposed against jamming in the transmission process of the federated target recognition model.This scheme establishes the MEC data offloading network model for the target recognition and uses the federated learning to protect the data of devices.Then,this scheme utilizes the historical anti-jamming experiences of the edge devices for mining and transferring availably to resist the smart jamming attacks by optimizing the channel selection and data allocation strategies dynamically,without knowing the target recognition model,channel model and jamming model.Simulation results in the presence of jamming show that the proposed scheme increases the accuracy of the target recognition by 18.9%,saves the energy consumption by 27.3%and reduces the latency of the mobile devices by 35%compared with the Lagrange multipliers based MEC data offloading scheme.Secondly,a deep reinforcement learning based anti-jamming MEC data offloading and model training(DROT)scheme is proposed for the edge devices equipped with strong computation capacity.This scheme designs the deep neural network to compress the state space,which reduces the exploration time for channel selection and data allocation in the jamming environment,and further improves the target recognition performance.Experiment results show that the proposed scheme increases the accuracy of the target recognition to 92.5%,saves the energy consumption to 110 mJ and reduces the latency to 1.27 s of mobile devices,respectively.Meanwhile,experiment results show that the proposed scheme also increases the accuracy of the target recognition by about 11.4%,saves the energy consumption by 32.1%and reduces the latency by 23.4%of the mobile device compared with RLOT scheme.