Weakly Supervised Attention Learning for Robust Cross-Temporal Specific Emitter Identification

Specific emitter identification (SEI) has emerged as a critical technology owing to its hardware independence and robust resistance to spoofing. However, significant challenges arise over long time spans, during which the signal transmission environment and radio frequency fingerprint (RFF) characteristics disrupt the consistency of the distribution between the training and testing data, severely degrading the model performance. To address these challenges, this letter introduces a new model based on weakly supervised attention learning and a swin transformer (WSAL-ST). This model initially preprocesses the collected signals and extracts their time-frequency feature maps using the Choi-Williams distribution (CWD), followed by the extraction of RFF via a swin transformer, which extracts discriminative RFF features with exceptional efficiency. Furthermore, it employs a WSAL to generate attention maps that delineate the spatial distribution of discernible regions to extract crucial local sequential features for precise identification. Concurrently, an attention supervision mechanism based on the class-center loss provides weak supervision for the attention generation process. The experimental results demonstrate that this model significantly enhances the performance of cross-temporal SEI, offering effective technical support for the practical application of SEI systems.