Open-Set Specific Emitter Identification Based on Activation Reshaping and Important Neurons

Aiming at the challenge that existing research on open-set specific emitter identification (SEI) has low identification performance due to high similarity between known classes and unknown classes, this paper proposes an approach based on activation reshaping and important neurons. This method utilizes convolutional neural network (CNN) to extract latent features from the signals of emitter individuals. Specifically, activation reshaping and pruning based on Shapley value are applied in the penultimate layer of the CNN to enhance differentiation between known and unknown emitter individuals. The open-set score derived from model logits and the Energy score function is leveraged in proposed method to effectively discriminate unknown emitter individuals from known emitter individuals, thereby transforming open-set recognition into closed-set recognition. A balanced trade-off between accuracy for known and unknown classes is achieved by the proposed method. Experimental results demonstrate that under various degrees of openness, the proposed method achieves more accurate open-set SEI classification results.