Abstract
With the popularity of Internet of Things (IoT) smart devices, the amount of data generated by devices has grown rapidly. To relieve the pressure on device computing and energy consumption, mobile edge computing (MEC) offloads computing tasks to the edge. In some specific private MEC environments, feguarding user data privacy is also critical. In this paper, due to the dynamics and complexity of the system, a multi-type task offloading based on multi-capability federated deep Q-network (M2FD) algorithm is proposed to optimize the bi-objective performance. In federated deep reinforcement learning, federated learning protects user data privacy by transmitting model for training instead of data; deep reinforcement learning trains model accuracy and identifies suitable offloading nodes with heterogeneous capabilities for multi-type tasks. In addition, under the constraints of the service experience guarantee (SEG) model, the tasks are offloaded with the goal of improving system utility while reducing system cost. Moreover, data privacy is safeguarded during the offloading process in the M2FD algorithm. Experiments demonstrate that the M2FD increases system utility, decreases task response time and energy consumption.