Transferable Physical Layer Authentication Based on Time-Varying Patterns Toward Zero Training Deployment for Mobile IIoT Devices

Physical layer authentication (PLA) is a promising approach to ensure wireless network security. However, existing PLA algorithms require channel state information calibration and model training at each location, which limits the device's mobile ranges. To enable zero training deployment in uncalibrated scenarios, we propose a transferable PLA (TPLA) algorithm. It exploits the channel time-varying patterns as the scenario-independent features to avoid PLA failure in uncalibrated scenarios. To extract transferable time-varying patterns from dynamic industrial mobile environments, a neural network feature extractor is designed by a multibranch parallel architecture with multiscale channel receptive fields. Furthermore, the global and component features are fused by the model-level and decision-level fusion methods to accommodate different transferability and computational cost requirements. In experiments, TPLA achieves below 2.5% authentication error in a new scenario, which proves that TPLA is an important step toward zero training deployment of the PLA algorithm.