Optimal-Feature-Driven Adversarial Adaptive Cascade Forest Method for Automatic Modulation Recognition

Automatic Modulation Recognition (AMR) plays a crucial role in tackling the challenges of spectrum scarcity within the Internet of Things (IoT), facilitating dynamic sensing and allocation of spectrum resources. However, traditional deep learning (DL) based AMR methods necessitate the addition of redundant layers and parameters to improve recognition accuracy, particularly under low signal-to-noise ratio (SNR) conditions. Furthermore, these methods heavily rely on features extracted from expert knowledge and high-quality data annotations, leading to significant reconstruction and full-process training costs when adapting to new samples. To tackle these issues, an optimal-feature driven adversarial adaptive cascade forest (OF-ACF) method for AMR is proposed. This method first constructs an optimal feature subset in the high-order domain feature space utilizing augmented recursive feature elimination (ARFE). Subsequently, it employs adversarial training (AT) of the adversarial adaptive cascade forest (AACF) algorithm driven by this subset, thereby enhancing recognition accuracy in low SNR environments and with unknown modulation schemes. The early stopping (ES) and adaptive growing (AG) mechanisms incorporated within the AACF method can effectively mitigate the resource consumption. Furthermore, the visualization of feature contributions through Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME) significantly enhances the interpretability of the model. The OF-ACF method achieves a recognition accuracy of 70.9% at an SNR of 0 dB on the RadioML2018.01A dataset, representing a 10.96% improvement over the MADWN method. At SNR values exceeding 10dB, the accuracy stabilizes above 98.85%. Ablation experiments confirm the effectiveness and efficiency of the proposed method, facilitating high-accuracy recognition speeds at the 1-ms level.