Prototypical Network for Few-Shot Signal Recognition

Abstract

With the rapid development of 5G and IoT technologies, Wireless Physical Layer Identification (WPLI) is essential for securing wireless communication technology, facing with more complex communication environment and access to heterogeneous networks. How to carry out device authentication in a small sample environment is an urgent and challenging problem. However, most data-driven deep learning (DL)-based WPLI approaches currently are not suitable for scenarios with insufficient samples. To address the issue, we learn from the prototype network in the image domain and use the meta-training method to solve the Few-Shot Learning (FSL) problem in signal recognition. To compare the prototype network performance, we conduct a 1-shot/5-shot test and compare with two baseline models. We further explore the impact of cross-domain on few-shot signal recognition. Simulation results show that prototype network achieves excellent performance on the same domain few-shot classification problem, but when there is distribution drift, the prototype network model adaptability is lower than the baseline model.