Open-Set Long-Tailed Radio Frequency Fingerprint Identification

Radio frequency fingerprint identification (RFFI) is considered as an important physical-layer authentication scheme for wireless device security. In practice, both low-activity devices, i.e., few-shot devices, and unknown devices could coexist in the environment. This leads to the extremely challenging open-set long-tailed RFFI problem due to the difficulty in differentiating between the few-shot devices and the unknown devices. To address this challenge, we propose a RFFI architecture based on dynamic meta-embedding and distance-based temperature control (DME-DTC). On the one hand, the dynamic meta-embedding (DME) part augments the extracted feature with memory features derived from the sample-centroid distances in trained feature space. On the other hand, the distance-based temperature control (DTC) part reshapes the classifier's output distribution with an adjustable temperature parameter. In these ways, we can amplify the differences between the classifier's output distribution of the few-shot devices and that of the unknown devices for effective device identification. After that, a simple divide-and-combine data augmentation method is applied to further enhance the performance. The experimental results show that the average accuracy of the proposed DME-DTC architecture exceeds existing model-agnostic meta-learning (MAML) based methods by about 10%, reaching a category-wise average accuracy of 91% (with data augmentation).