Deep Learning for Specific Emitter Identification in Modern Electronic Warfare: From Static Models to Cognitive Adaptation

The complexity of the electromagnetic spectrum in the modern battlefield has made it critical for Cognitive Electronic Warfare (EW) systems to identify hostile elements not only by type but also at the individual identity level using Radio Frequency Fingerprints (RFF) originating from hardware imperfections, a process known as Specific Emitter Identification (SEI). This study comprehensively examines the evolution of SEI literature from manual feature extraction to data-driven Deep Learning (DL) approaches from the perspectives of data preprocessing, architectural design, and dynamic environmental adaptation. The analyses demonstrate that although static deep architectures such as Deep Residual Networks (ResNet) and Temporal Convolutional Networks (TCN) exhibit accuracies exceeding 96% in controlled laboratory environments, their reliability drops to levels around 50% under real-world field conditions such as data scarcity (few-shot), the presence of unknown threats (open-set), and adversarial attacks. In this context, the study emphasizes the importance of multimodal feature fusion techniques, such as Variational Mode Decomposition (VMD) and Bispectrum analysis, to bridge the reliability gap; furthermore, it proposes the integration of Extreme Value Theory (EVT)-based open-set recognition and meta-learning strategies for the system to identify unknown threats and update itself autonomously. Consequently, it is revealed that a transition from static models to self-healing and adaptive cognitive architecture is imperative for the sustainability of EW superiority.