DyCE-Net: Cross-Modal Radar Deception-Jamming Discrimination via Dynamic Convolution and Evidential Fusion

Existing intelligent radar anti-jamming methods relying on single-modal observations often suffer from unstable feature extraction and limited exploitation of cross-modal complementarity under complex deceptive interference. Motivated by these limitations, this letter proposes DyCE-Net, a cross-modal radar anti-jamming network based on dynamic convolution and evidential fusion. In the signal branch, a Phys-ODNet is designed to embed physics-guided filtering priors into dynamic expert convolution kernels, enabling input-adaptive signal feature extraction. In the track branch, a Transformer is employed to model the temporal motion characteristics of targets. At the fusion stage, subjective-logic-based Dempster-Shafer evidence theory is introduced to jointly infer dual-modal evidence and uncertainty, thereby improving discrimination accuracy under decision conflicts. Experimental results show that DyCE-Net achieves accuracy, Pd and Pfa of 0.9963, 0.9945 and 0.0019 respectively, demonstrating competitive discrimination performance. Further visualization results verify the effectiveness of the proposed method in feature representation, conflict suppression, and uncertainty modeling. The proposed method provides an effective solution for multimodal radar anti-jamming discrimination.