Specific emitter identification (SEI) refers to the technique of identifying different individuals from the signals emitted by wireless devices. Recent studies have focused mainly on deep learning (DL) models that automatically learn valid inherent features from raw time-domain signals. However, current studies rarely consider real open-world scenarios, where new classes may emerge during the inference phase, and the utilized model must evolve as new classes incrementally appear. An incremental open-world learning (IOWL) framework is proposed in this paper, and we show how IOWL can continually recognize and learn new classes. The proposed method is based on a novel exemplar selection and generalization mechanism. First, by applying edge pattern detection (EPD) and shifting edge samples along the adversarial direction, a high-quality pseudo unknown dataset is generated to improve the open-set recognition (OSR) process. Second, a hybrid class-incremental learning method is proposed to maintain the previous identification capabilities through boundary exemplar generation, which not only benefits each individual paradigm but also highlights their synergies in a common framework. We provide a theoretical analysis of the obtained generalization error bounds to prove the benefits of the proposed method. Numerical results on real collected data indicate that IOWL consistently outperforms the other baseline algorithms.

Machine learning approaches are becoming increasingly popular to improve the efficiency of specific emitter identification (SEI). However, in most non-cooperative SEI scenarios, supervised and semi-supervised learning approaches are often incompatible due to the lack of labeled datasets. To solve this challenge, an unsupervised SEI framework is proposed based on information maximized generative adversarial networks (InfoGANs) and radio frequency fingerprint embedding (RFFE). To enhance individual discriminability, a gray histogram is first constructed according to the bispectrum extracted from the received signal before being embedded into the proposed framework. In addition to the latent class input and the RFFE, the proposed InfoGAN incorporates a priori statistical characteristics of the wireless propagation channels in the form of a structured multimodal latent vector to further improve the GAN quality. The probabilistic distribution of the bispectrum is derived in closed-form and the convergence of the InfoGAN is analyzed to demonstrate the influence of the RFFE. Numerical results indicate that the proposed framework consistently outperforms state-of-the-art algorithms for unsupervised SEI applications, both in terms of evaluation score and classification accuracy.