A Novel Partial Domain Adaptation Method for Cross-Domain Specific Emitter Identification

In wireless communication networks (WCN), specific emitter identification (SEI) technology is significant for guaranteeing cybersecurity and can effectively defend against network attacks. Compared to traditional SEI methods that depend on specific training data or environmental conditions, cross-domain recognition study has attracted widespread attention for dynamic WCN. However, in a more general scenario, the target domain (TD) sometimes represents a subset of source domain (SD), i.e., in partial domain adaptation (PDA), which has not been well addressed. To solve this concern, we present a novel selective partial domain adaptation method based on complex-valued convolutional block attention module (CSPDA) for SEI, which tackles the domain shift challenge caused by the inconsistency in emitter categories between SD and TD. The proposed method includes a mixed feature extraction module that efficiently extracts representative features from both source and target raw signals at multiple levels, reducing feature information loss. In addition, by using the maximum of cosine (MoC) assessment mechanism and a data information filtering strategy, we can dynamically select the appropriate source signal samples to match each target signal sample. Experiment results on real-world data show that the CSPDA method achieves a recognition accuracy of 73.46% at -5 dB and 89.28% at 20 dB, outperforming other transfer learning and domain adaptation methods in PDA tasks.