Specific Emitter Identification based on Multiple Gaussian Prototype Learning and Deep Attention Residual Network in Open-Set Scenarios

Specific emitter identification (SEI) can enhance the recognition of external devices and improve network communications security performance. Although many methods that combine deep and prototype learning have been proposed to identify specific emitters in open-set practical application scenarios, they are limited by low recognition performance and inadequate robustness. To resolve these challenges, this study proposed an SEI method based on multiple Gaussian prototype learning methods and a deep attention residual network. First, a one-dimensional deep attention residual network was designed to extract latent features from the I/Q signal sequences. Then, a channel attention mechanism and gated recurrent unit were added to the network to improve the feature extraction capability. Subsequently, the deep attention residual network was combined with multiple Gaussian prototype learning methods. A preprocessing dataset was used for training and testing. During the training process, several Gaussian prototypes were set for different types of radiation source signals, and the network parameters and Gaussian prototypes were updated using the joint loss function of generating and discriminant constraints. The distance between the latent features of the radiation source signals and each Gaussian prototype was used as the basis for classification. During the testing process, known targets were classified, and unknown targets were detected using a reasonable distance threshold. SEI experimentation on five types of specific emitter signals, conducted in open- and closed-set scenarios, demonstrated that the proposed method exhibited better recognition performance than the other methods, particularly in the open-set scenario; the area under the curve value of the proposed method increased by 13.1% compared with that of the other methods. Thus, the approach provides an accurate and improved technical path for SEI in both closed- and open-set scenarios.