High-Confidence Sample Augmentation Based on Label-Guided Denoising Diffusion Probabilistic Model for Active Deception Jamming Recognition

Abstract

Accurate recognition of the types of mainlobe active deception jamming is essential for radar systems to take anti-jamming countermeasures. A bunch of deep learning (DL)-based recognition methods that require largescale datasets for training have shown promising results. However, capturing a sufficient number of diverse deception jamming samples is particularly intricate in actual dynamic and complex battlefields, the yielding limited or unbalanced datasets presents a significant challenge in training and generalizing DL models. This letter proposes a deep generative model, called label-guided denoising diffusion probabilistic model (LG-DDPM), to address the issue of limited or class-imbalanced active deception jamming samples through data augmentation. By embedding label information into the diffusion process, the proposed model can generate and expand the active deception jamming samples specific to a pre-defined class even under low jamming-to-noise ratios (JNR) scenarios. The proposed method demonstrates superior performance in terms of both the fidelity and diversity of the generated jamming samples as well as the recognition accuracy of the DL recognizer when compared to state-of-the-art methods. Experimental results demonstrate the effectiveness and robustness of the proposed method.