TFMix: A Robust Time-Frequency Mixing Approach for Domain Generalization in Specific Emitter Identification

Specific emitter identification (SEI) is a recognition technique based on the inherent features reflected in radio frequency signals due to individual differences among emitters. Deep learning (DL) methods for SEI have become a primary approach for wireless device identification and authentication. However, variations in electromagnetic signals across different environments lead to significant performance degradation in models when applied to unseen domains, making traditional DL methods insufficient. As a result, domain generalization (DG) has emerged as a crucial approach to address this challenge. To mitigate the limitations of SEI performance across diverse scenarios, we propose a mixing enhancement method, TFMix, which combines time-domain and frequency-domain data augmentation. Specifically, TFMix first performs time-domain interpolation (TDI) to generate augmented samples, followed by a Fourier transform. The method then applies amplitude interpolation mixing while introducing controlled phase perturbation. Afterward, an inverse Fourier transform is used to reconstruct more robust signal samples. This hybrid approach enhances model generalization by producing diverse and informative training samples. To validate the effectiveness of TFMix, we conducted evaluations on datasets from different scenarios. The experimental results demonstrate that TFMix significantly outperforms existing DG methods in terms of generalization across various scenarios, showcasing its robustness and superior generalization capability in SEI tasks. Our code can be downloaded from <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/frownean/TFMix</uri>.