Prototype-Driven Unsupervised Domain Adaptation for Specific Emitter Identification

Distribution shift is a prominent challenge for specific emitter identification (SEI) in noncooperative scenarios. Unsupervised domain adaptation (UDA) aims to address the aforementioned issue by transferring knowledge from labeled source domain to unlabeled target domain. However, existing UDA methods primarily focus on cross-domain alignment of global features, causing less discriminability between features of different classes. Building upon this observation, we leverage prototypes to extract domain-invariant class knowledge and propose a prototype-driven UDA (PDDA) framework to enhance the discriminability of different classes. First, we propose a pseudo-label mechanism to assist the target domain in self-supervised training, utilizing unlabeled samples to learn discriminative feature. Second, building upon the relationship between features and the corresponding prototypes, we propose a prototypical bidirectional alignment (PBA) method to achieve class-level transfer. Finally, we propose a prediction consistency constraint to balance transferability and discriminability of features. The proposed method is compared with state-of-the-art methods on the two publicly available radio frequency fingerprint data sets and the results demonstrate the effectiveness of PDDA.