KAN-MAE: KAN-Based Masked Autoencoder With Correntropy-Aided Near-Homogeneity Strategy for Automatic Modulation Classification

Automatic modulation classification has attracted considerable research interest owing to its critical role in spectrum utilization for vehicular wireless communications. To address this task, a novel solution, the Kolmogorov–Arnold network–based masked autoencoder (KAN-MAE), is proposed. This solution employs a time-series representation and incorporates pre-training and fine-tuning. During pre-training, a masking operation helps extract core features, while the autoencoder captures contextual and mutual information within and between in-phase and quadrature sequences. This architecture uses modified Transformer blocks with KAN-based multi-head attention in the encoders and decoder. In the fine-tuning phase, the pre-trained encoders are retained, and KAN is used as the classifier. To reduce the influence of intensive noise across data subsets, a correntropy-aided near-homogeneity (CANH) strategy is designed. A comprehensive evaluation compares KAN-MAE with seven state-of-the-art (SOTA) methods across four datasets, considering fading channels, noise models and levels, dataset sizes, modulation types, and sample lengths. The impact of the masking ratio is also investigated. An ablation study validates the effectiveness of each design element, and the influence of sample length and sliding window length (SWL) on model size and computational complexity is further examined. Experimental and statistical results confirm the superiority and reliability of the proposed KAN-MAE method compared with existing SOTA methods under various conditions.