Cross-Modulation Specific Emitter Identification Based on Source-Free Domain Adaptation

Recently, there has been a large amount of research conducted on specific emitter identification (SEI), with deep learning serving as the mainstream approach. The coupling of unintentional features (fingerprint features) and intentional features (modulation features) reduces the recognition accuracy. Current research is mainly focused on scenarios with the same intentional modulation, which limits the flexibility of SEI models. To solve the cross-modulation SEI problem, this paper proposes combining dual-input complex-valued convolutional neural network (CVCNN) with demodulated and reconstructed signals to improve the generalization ability of the SEI model. Furthermore, we adopt SHOT, a source-free domain adaptation (SFDA) method, to realize cross-modulation SEI without accessing source-domain data. Experimental results demonstrate the effectiveness of our method across three modulation types: QPSK, BPSK, and 16QAM.