Autoencoder-based physical layer authentication in a real indoor environment

Authentication of wireless nodes, as in fifth-generation (5G) and Internet of Things (IoT) networks, is an increasingly pressing issue, in order to limit the required computational effort and the necessary overhead. A simplification of the authentication process may therefore be of interest to achieve the satisfaction of stringent performance requirements, such as those envisaged for sixth-generation (6G) networks. This paper provides a study on the feasibility of physical layer authentication (PLA) in a real indoor environment, as an alternative solution to the traditional authentication schemes. To ensure the reliability of the proposed approach a simulated scenario is firstly tested. Subsequently, real-world data are collected through a laboratory setup using a Vectorial Signal Transceiver (VST) and two Universal Software Radio Peripherals (USRPs) to emulate the behavior of the receiver, the legitimate transmitter, and the potential adversary. A machine learning (ML) algorithm is then exploited to act as authenticator. This means that channel fingerprint is extracted from signals to create a dataset used to train a sparse autoencoder. To emulate a real authentication scenario, the autoencoder is trained only on the class of the legitimate user. Once a new message arrives, the autoencoder task is to discern authentic signals from those forged by the adversary. It is shown that a geometric mean of accuracy of more than 90%, with corresponding low levels of false alarm and missed detection, is achievable irrespective of the nodes location, underlining the robustness and versatility of the proposed ML-based PLA approach.