Prototype Similarity-Weighted Subdomain Adaptation With Dynamic Boundary for Open-Set Jamming Recognition

Jamming recognition is one of the key technologies for achieving anti-jamming communication. With the rapid development and application of deep learning, intelligent jamming recognition has gained a lot of achievement. However, identifying jamming in low Jamming-to-Noise Ratio (JNR) environments remains a challenging issue. Additionally, most existing intelligent jamming recognition models focus on closed-set scenarios and ignore the challenging open-set recognition problems caused by the variable electromagnetic environment. To address these issues, this paper proposes a Prototype Similarity-weighted Sub domain Adaptation (PSSA) method with Dynamic Boundary for open-set jamming recognition, based on deep transfer learning and prototype learning. First, for open-set jamming recognition, dynamic boundary anchor prototype learning is introduced to construct an outlier classifier to separate known and unknown jamming. Second, by combining transfer learning and prototype learning, this paper proposes Multi-layer, Multi-kernel Proto type Similarity-weighted Local Maximum Mean Discrepancy (MLMK-PSLMMD) metric. The metric enables the model to extract the fine-grained transferable information between sub domains and improves the recognition accuracy under low JNR conditions. Finally, this paper validates the performance of the proposed method on multiple datasets. When the model is pretrained on data with JNRs in the range of [0, 10] dB, it achieves over 90% accuracy in recognizing known jamming at-6 dB, at least 15% higher than the latest competitive methods. The F1-score for unknown jamming recognition reaches 90%, at least 25% higher than other methods.