CNN-Zero: A Zero-Shot Learning Framework for Jamming Identification

Abstract

Anti-jamming is a critical issue of wireless communication security, where jamming identification is an important pre-stage of anti-jamming. However, it is challenging to perform a jamming identification task in the absence of some jamming classes. To overcome this obstacle, we propose a zero-shot learning framework CNN-Zero, which aims to identify the known and unknown jamming signals. Specifically, we employ CNN to learn the potential representation of the semantic feature space of jamming signals. Then, we build a hybrid loss function consisting of attribute distance loss, cross entropy loss and reconstruction loss to ensure the semantic features have greater minimum inter-class distance than maximum intra-class distance. Finally, we build an appropriate distance measurement matrix to identify known and unknown jamming signals. Experimental results prove that compared with the supervised method using neural networks, CNN-Zero achieves a better average accuracy between eight jamming signals even in the absence of training samples.