FUTUREs: Feature Utility Transfer for Uncooperative Recognition of Electromagnetic Signals Toward 6G Wireless Communications

Automatic modulation recognition (AMR) plays a crucial role in the field of wireless communications, with applications in spectrum sensing, link adaptation, interference mitigation, and electronic warfare. With the proliferation of wireless communication devices and increasingly complex channel environments, acquiring a large-scale labeled signal dataset has become exceedingly challenging for deep learning based AMR. To address this issue, this paper proposes a novel neural network architecture, referred to as Feature Utility Transfer for Uncooperative Recognition of Electromagnetic Signals (FUTUREs), which introduces the meta-transfer learning (MTL) for unseen signal recognition. The MTL-based framework combines the strengths of transfer learning and meta-learning, thereby enabling the uncooperative recognition of unknown signals with limited labeled signal samples by transferring the feature utility (i.e., learning how to extract feature for unknown data) from the source domain. The FUTUREs framework also employs parallel convolutional projections to extract latent information, such as the mean and variance of input features, thus enhancing the recognition accuracy. Experimental results on RML2018.01a datasets demonstrate that, compared to other advanced neural networks, the proposed architecture achieves higher recognition accuracy on unknown modulation signals while maintaining lower network complexity. In addition, we also generate 6G candidate modulation signals including orthogonal time frequency space (OTFS) and orthogonal frequency division multiplexing (OFDM) using GNU radio. The testing results show that the proposed architecture still has higher recognition accuracy on the dataset containing these two modulation types.