Contrastive Self‐Supervised Learning Enables Robust and Data‐Efficient Automatic Modulation Classification

ABSTRACT Accurate automatic modulation classification (AMC) is vital for intelligent cognitive radio (CR) systems, yet traditional supervised methods face significant challenges. Their performance is hampered by a heavy reliance on large, expensively labelled datasets and poor generalization in real‐world dynamic environments characterized by low signal‐to‐noise ratios (SNRs), fading channels and hardware impairments. To overcome these limitations, this paper introduces a novel contrastive self‐supervised learning (CSSL) framework. The aim is to enable robust and data‐efficient modulation recognition by learning discriminative signal representations directly from unlabelled raw in‐phase and quadrature (I/Q) samples, eliminating the dependency on extensive annotations. The proposed approach employs a carefully crafted pretext task that uses domain‐specific augmentations to maximize the similarity between differently corrupted views of the same signal. Comprehensive evaluation on the RadioML2016.10A dataset demonstrates that this model significantly outperforms supervised benchmarks, achieving a standout accuracy of 98.10% at 15 dB SNR. Crucially, it exhibits superior robustness, maintaining high performance even at −10 dB SNR under a strictly separated training/testing protocol, where conventional models falter and shows strong resilience to carrier frequency offset (CFO) and phase noise. This work is significant as it offers a scalable and practical approach for deploying advanced AMC in real‐world CR applications, significantly reducing the cost of data labelling while enhancing performance and adaptability in the face of unpredictable channel conditions and noise. Furthermore, additional experiments under limited labelled data regimes (1, 10 and 50%) were implemented to demonstrate the data efficiency of the proposed framework.