A Self-Supervised Contrastive Learning Method Based on Transformer for Specific Emitter Identification

Aiming at the core role of Automatic Identification System (AIS) in Marine traffic management and the security threats it faces, this paper proposes a Transformer-based Self-supervised Siamese network framework (TSS). It aims to solve the problem that AIS signals are vulnerable to malicious interference and improve system security. Since traditional supervised learning methods are highly dependent on data annotation, it is difficult to be widely used in complex electromagnetic environments, TSS framework integrates the hierarchical self-attention mechanism of Swin-Transformer and the self-supervised learning strategy of SimSiam. Specifically, Swin-Transformer's local self-attention mechanism enables the network to capture local and global Radio Frequency Fingerprint (RFF) features of Specific Emitter Identification (SEI) at different scales. By constructing a two-branch Siamese network, TSS can automatically mine discriminative feature representation from unlabeled AIS signals. No need to rely on large amounts of manually annotated data. The experimental results based on the real world data show that the TSS algorithm has high accuracy and stronger generalization ability in the SEI task, and the accuracy is as high as 90.1%. Especially when dealing with SEI in a complex electromagnetic environment, TSS algorithm can still maintain stable recognition performance, and the identification accuracy is more than 10% higher than other advanced algorithms, which proves its potential and value in practical applications.