Recurrent Neural Unit With Frequency Attention for Specific Emitter Identification

Abstract

With the development of wireless communication, specific emitter identification (SEI) used to enhance communication security is becoming increasingly important. Due to the inherent timing characteristics of signals, many new variants of recurrent neural networks (RNNs) have been proposed to extract temporal features and identify individual emitters. In the present article, frequency attention is applied to two classical RNN methods, namely long and short-term memory (LSTM) unit and gated recurrent unit (GRU), to extract both time and frequency features. Three implementations of frequency attention in LSTM and GRU are also discussed and described. Massive experiments have been conducted on a dataset collected from 9 devices of the same type. The results on the one hand show that LSTM and GRU with in-place frequency attention module have improved performance and outperform a large number of SEI methods, and on the other hand show that the performance of different implementations of frequency attention in LSTM and GRU varies a lot. The LSTM unit that combines the original hidden state and the hidden state with frequency attention through a controllable gate has the best performance with the minimum parameters.