New Untrained Emitter Detection Based on SK-GAND Network

The increasing complexity of the spatial electromagnetic environment has triggered the need to discriminate individuals with unknown radiation sources, and the recognition effect of traditional deep networks and other methods based on supervised learning is poor. To address the problem of identifying new untrained devices in RF fingerprinting, this paper proposes an unsupervised SK-GAND neural network-based method for detecting new untrained radiation sources, adding the SK module to the traditional GAN network and setting two thresholds γand ω for identifying trained and new radiation sources. The thresholds
$y$
and ω are determined by training the difference between the samples obtained in the SK-GAND network for the existence of encoding and decoding re-encoding, and a multi-discriminator model is used to determine whether the input radiation source samples are new untrained sources. Based on the experimental comparison of different numbers of radiation sources, the detection effectiveness is obtained up to 80% for up to 12 sources and up to 90% for up to 5 sources.