Adversarial diffusion synthesis for specific emitter identification: multi-scale signal augmentation with cross-domain constraints

Generative data augmentation offers a promising solution to data scarcity in Internet of Things(IoT) emitter identification, yet standard Denoising Diffusion Probabilistic Model (DDPM) often lacks semantic and structural control, hindering their capture of hierarchical signal features and cross-domain physical consistency. For this, this paper proposes a DDPM enhanced with a multi-scale hybrid attention mechanism, which jointly optimizes consistency across time-domain waveforms, frequency-domain modulation patterns, and sequential signal features. By integrating contrastive learning with autoregressive modeling, the proposed method captures long-range temporal dependencies and effectively represents complex signal interactions across multiple time scales. Then, we propose a multi-domain adversarial joint constraint to align the generated and real data distributions across physical, statistical, and semantic domains, significantly enhancing cross-domain feature consistency. The experimental results obtained show that, compared with the existing benchmark tests, our solution has increased the recognition accuracy by 7.91% in the case of small sample sizes, and also demonstrates stronger robustness even under low signal-to-noise ratio conditions.