End-to-End Learning of Secure Wireless Communications: Confidential Transmission and Authentication

Abstract

Aiming to provide more efficient and robust physical layer security strategies for wireless communications, this article investigates the endogenous security of end-to-end learning of communication by addressing two main security issues of communication: confidential transmission and user authentication. For confidential transmission, we have redesigned the loss function of the autoencoder-based deep learning communication model to combat illegal eavesdropping over wireless broadcast channels. While assuming that the eavesdropper has three different ways of decoding prior information, the probability of successful eavesdropping attack is evaluated using the bit error rate criterion. In terms of user authentication, an authentication scheme using symbol-level fingerprints is designed for a single user, which takes advantage of the high complexity of parameters of the deep learning model and its natural sensitivity to training conditions. In addition, by leveraging a denoising autoencoder, we extend the authentication to adapt to the multi-user access situation. Experiments have shown that the proposed authentication scheme could guarantee reliability under dynamic channel and resistance to wireless attacks. The results inspire us to rebuild an efficient physical layer secure framework for wireless communication through a new deep learning method.