DPDS: A Systematic Framework for Few-Shot Specific Emitter Incremental Identification

Abstract

Specific emitter identification (SEI) technology is crucial for supporting emergency response and safety alerts, and it plays a significant role in maintaining the stable operation and information security of Internet of Things (IoT) systems. To mitigate the catastrophic forgetting and overfitting in few-shot specific emitter incremental identification, we propose a systematic identification framework called data processing and dynamic subnet (DPDS). This framework consists of a data processing module (DPM) and a dynamic subnet module (DSM), where DSM maintains recognition performance for both new and old tasks by adjusting the submodel of the original model dynamically. In DPM, we preprocess all received signals and perform additional data augmentation for few-shot to obtain more significant data representation and reduce overfitting. These processed signals serve as the inputs for DSM. In DSM, we identify an optimal subnet of the model, ensuring performance retention for old tasks while training new tasks on the remaining parts of the model, thus mitigating catastrophic forgetting. Additionally, when the model struggles to support learning new tasks, we expand the model nodes, single parameters in practical terms, appropriately and learn the importance of each expanded node. This approach enables us to compress the expanded nodes and achieve the optimal model architecture. Finally, we design a new recognition paradigm based on the proposed boundary disorder phenomenon, which indicates the differences in the feature space between data from distinct distribution domains. The experimental results indicate that DPDS significantly outperforms baseline methods and demonstrates superior performance compared to other state-of-the-art algorithms.