Specific Emitter Identification via Convolutional Neural Networks

Abstract

Specific emitter identification (SEI) is a technique that distinguishes between unique emitters using the external feature measurements from their transmit signals, primarily radio frequency fingerprints. The SEI has been widely adopted for military and civilian spectrum management applications. We propose a deep-learning-based SEI approach that uses the features of the received steady-state signals. In particular, the bispectrum of the received signal is calculated as a unique feature. Then, we use a supervised dimensionality reduction method to significantly reduce the dimensions of the bispectrum. Finally, a convolutional neural network is adopted to identify specific emitters using the compressed bispectrum. This approach essentially extracts overall feature information hidden in the original signals, which can then be used to improve identification performance. Results from both the simulations and the software radio experiments are provided. A signal acquisition system is designed to collect steady-state signals from multiple universal software radio peripherals. Both the simulations and the experiments validate our conclusion that the proposed approach outperforms other existing schemes in the literature.