CVCA: A Complex-Valued Classifiable Autoencoder for MmWave Massive MIMO Physical Layer Authentication

For protecting millimeter wave (mmWave) communications from clone attacks, this paper employs the deep learning to propose a physical layer authentication (PLA) approach for detecting attackers and classifying multiple legitimate nodes simultaneously. Different from conventional upper-layer authentication mechanisms, the proposed PLA approach exploits the spatial and temporal characteristics of mmWave channels to extract the unique fingerprints for building a lightweight channel-based authentication method. However, the existing threshold-based PLA methods could not discriminate multiple nodes, and supervised learning based approaches have limited application due to the unavailability of attackers' channel state information (CSI) in practice. Besides, traditional real-valued deep neural networks cannot exploit the phase information of complex channels efficiently, which is unsuitable for designing the PLA scheme. Considering these, we propose a complex-valued classifiable autoencoder induced PLA scheme that includes a novel complex-valued long short-term memory (LSTM) module. Simulation results validate the superiority of our proposed PLA approach by comparing it with existing approaches and demonstrate that the detection probability of clone attacks positively correlates with antenna number. The classification performance is satisfactory even under the challenging experimental condition.