A fast security authentication scheme based on meta-learning under the changing channel environment

Abstract

This paper presents a meta-learning-based scheme for secure authentication at the physical layer. The solution focuses on efficiently adapting neural network-based physical layer security authentication to new environments. It aims to minimize the need for extensive re-learning and human intervention in the process. Specifically, the meta-learning-based physical layer authentication scheme uses the inner and outer cycle of the meta-learning algorithm to find a common initialization between tasks. When applying identical parameters across various tasks and fine-tuning them, consistent good results indicate the model’s ability to extract common features across tasks. On this basis, this paper utilizes a convolutional neural network to collect channel state information (CSI) data from multiple historical environments to train the network. This enables feeding a small amount of CSI data into a network model trained in the historical environment of the new setting, quickly generating an authenticator suited to the new environment. To evaluate the performance of the proposed scheme, this paper conducts experiments on real datasets. The results show that the scheme outperforms the transfer learning approach in both authentication accuracy and convergence. This highlights the economical and efficient nature of meta-learning-based physical layer authentication schemes.