15 May 2025 in IEEE Communications Letters
We propose Denoising Masked Autoencoder (Deno-MAE), a novel multimodal autoencoder framework for denoising modulation signals during pretraining. DenoMAE extends the concept of masked autoencoders by incorporating multiple input modalities, including noise as an explicit modality, to enhance cross-modal learning and improve denoising performance. The network is pre-trained using unlabeled noisy modulation signals and constellation diagrams, effectively learning to reconstruct their equivalent noiseless signals and diagrams. DenoMAE achieves state-of-the-art accuracy in automatic modulation classification (AMC) tasks with significantly fewer training samples, demonstrating a 10× reduction in unlabeled pretraining data and a 3× reduction in labeled fine-tuning data compared to existing approaches. Moreover, our model exhibits robust performance across varying Signal-to-Noise Ratios (SNRs) and supports extrapolation on unseen lower SNRs. The results indicate that DenoMAE is an efficient, flexible, and data-efficient solution for denoising and classifying modulation signals in challenging, noise-intensive environments. Our codes are public at https: //github.com/atik666/denoMae/tree/master.
28 October 2025 in IEEE Transactions on Communications
We introduce DenoMAE2.0, an enhanced denoising masked autoencoder designed to significantly improve representation learning for Automatic Modulation Classification (AMC) in wireless communications. Unlike standard Masked Autoencoders (MAEs), which solely reconstruct masked inputs, DenoMAE2.0 jointly performs denoising and reconstruction by incorporating a position-aware local patch classification objective. This approach enables the model to simultaneously denoise corrupted signals and accurately reconstruct missing information, effectively capturing both global context and local structural patterns crucial for modulation classification under noisy and data-scarce scenarios. Extensive experiments demonstrate that DenoMAE2.0 achieves superior denoising and classification performance, outperforming baseline methods and its predecessor, DenoMAE, particularly under extremely low Signal-to-Noise Ratios (SNR) (down to -7dB), where other approaches significantly degrade. Our method achieves state-of-the-art accuracy of 82.4%, improving upon the original DenoMAE by 1.1%, while consistently enhancing reconstruction quality across all modulation classes. Additionally, DenoMAE2.0 exhibits robust transfer learning capabilities, achieving consistent improved accuracy over a Vision Transformer (ViT) at different SNRs with a 6.13% highest gain at noise-signal equilibrium point. These results highlight the effectiveness of our dual-objective, self-supervised framework for robust AMC in challenging wireless communication environments.