Learning Based Efficient Federated Learning for Object Detection in MEC Against Jamming

Abstract

Federated learning enables mobile edge computing (MEC) to train the object detection model with privacy protection and reduced communication overhead. However, the selection of the mobile devices and the training dataset that determines the energy consumption of the mobile devices and the detection accuracy and latency has to be optimized without relying on the known channel and jamming model against jamming attacks that aim to degrade the model training performance. In this paper, we propose a reinforcement learning (RL) based efficient federated learning training scheme against jamming. This scheme designs a fast RL algorithm with shared parameters to choose the training policy of the object detection model at the mobile devices based on the channel gain, the previous training, transmission and computation performance. The edge server uses a shared Q-table to determine the policy for each mobile device to accelerate the learning process. Simulation results show that this scheme can effectively improve the object detection accuracy, decrease the energy consumption and reduce the latency compared with the benchmark scheme.