A Novel Layer Pruning Method Based on Graph Structure of Feature Representations for Specific Emitter Identification

As an emerging method for physical layer security authentication, Specific Emitter Identification (SEI) plays a pivotal role in wireless communication network security. Leveraging its strong feature extraction capabilities, deep learning has exhibited excellent performance in SEI. However, most network architectures have many parameters, e.g., AlexNet, VGG, and ResNet. This brings a challenge for their deployment in devices especially those with constrained computation capacities. To address above concerns, this paper proposes a layer pruning method based on the graph structure of feature representations, which can simultaneously lightweight the neural network architecture, decrease the model complexity, and maintain the identification performance. This paper maps feature representations into a dynamic graph evolution process and determines redundant parameters based on the community detection of the graph. Moreover, the pruning process operates at the layer level, making it more convenient and not affecting the existing neural network training process. Experimental results demonstrate that the method can significantly reduce computational complexity, and the pruned models exhibit comparable or even better performance compared to the original models on multiple datasets. For instance, when compressing the computational cost of VGG16 by 95.4%, the identification accuracy is improved by 1.22%.