IQ-CMAE: Multi-Modal Contrastive Masked Autoencoder for IQ-Based Signal Classification

This brief proposes a multi-modal self-supervised learning (SSL) framework for classifying wireless signals from raw in-phase and quadrature (IQ) traces, enabling joint classification of transmission bandwidths, power levels, and modulation types. Unlike conventional modulation classification, jointly inferring these attributes yields a more discriminative signature for reliable signal identification. The proposed framework employs a contrastive masked autoencoder (CMAE) to exploit the intrinsic structure of the unlabeled data, while reducing reliance on manual annotation. IQ-CMAE integrates generative reconstruction and contrastive alignment objectives, reconstructing masked inputs across multiple IQ modalities, namely spectrograms, Gramian angular fields (GAFs), and constellations while enforcing instance-level discrimination via contrastive learning. To balance the dual objectives, we employ contrastive gradient stopping, restricting contrastive gradients to the top K encoder layers, which enhances stability and robustness under channel variations and low-data regimes. We further conduct a systematic analysis of multi-modal fusion depth-evaluating early, mid, and late fusion strategies to characterize their interaction with the hybrid learning objective. Extensive experiments on both controlled-lab and real-world datasets demonstrate that IQ-CMAE with mid-level fusion significantly outperforms standard MAE and contrastive SSL baselines in classification accuracy and representation robustness.