Accurate Open Set Radio Frequency Fingerprinting Identification Enabled by a Novel Hierarchical Autoencoder Network

Radio frequency fingerprinting identification (RFFI) plays a vital role in achieving wireless communication security. The vast majority of the existing RFFI schemes are implemented under the closed set RFFI, which is not robust in the open electromagnetic spectrum environment. The open set RFFI is promising to solve the above drawback. However, the existing open set RFFI schemes are committed to reducing the distance between the same class without increasing the distance between different classes, which may cause the known samples and the unknown open samples to map to the same location in the feature space. In order to improve the robustness in the open electromagnetic spectrum environment while maintaining the identification accuracy of known samples, an accurate open set RFFI scheme is proposed by utilizing a novel hierarchical autoencoder network based on shuffle attention (SA-HANet). The SA-HANet is exploited to extract the high-level features for better distinguishing the known samples and the unknown open samples. Moreover, a loss function based on metric learning is used to minimize the distance to the ground-truth anchor point while maximizing the distance to other classes. The simulation results demonstrate that our designed SA-HANet scheme can better extract the characteristics of the known samples and retain the characteristics of the unknown open samples. Moreover, the identification accuracy of the open set RFFI is improved while the identification accuracy of the closed set RFFI is better than the baseline schemes.