DC4S: A Dual-Contrast Self-Supervised Learning Framework for Robust and Efficient Spectrum Sensing With Limited Labeled Data
Zhengyang Su, Kah Chan Teh, Yihang Xie, Sirajudeen Gulam Razul, Alex Chichung Kot · IEEE Transactions on Cognitive Communications and Networking · 2025
Non-cooperative spectrum sensing is a critical aspect of efficient spectrum utilization in wireless communications. However, the reliance on extensive labeled data poses a significant challenge to the application of deep learning methodologies in this domain. To address this issue, we propose a novel contrastive self-supervised learning framework called Dual-Contrast Self-Supervised Spectrum Sensing (DC4S). Inspired by the success of contrastive self-supervised learning in computer vision (CV) and natural language processing (NLP), DC4S extends its application to the field of communications, aiming to reduce the dependency on labeled data while learning robust and discriminative signal representations. DC4S innovatively integrates cross-view prediction and signal-context contrasting to exploit the intrinsic characteristics of communication signals, facilitating robust feature extraction without the need for labeled datasets. The cross-view prediction module leverages the temporal dynamics of communication signals, enabling the model to predict future signal states by understanding the past, thus fostering a deep temporal coherence in the learned representations. Subsequently, the signal-context contrasting module is designed to refine these representations by focusing on the contextual features of the signals, ensuring that the extracted features are discriminatively robust and contextually enriched. Extensive experiments are conducted to evaluate the performance of DC4S with various hyper-parameters on orthogonal frequency division multiplexing (OFDM) signals. The evaluation results demonstrate that fine-tuning a linear classifier in conjunction with the encoder pre-trained in the self-supervised stage achieves comparable performance to supervised learning methods when there exist only 5 labeled samples (1%) per signal-to-noise ratio (SNR) level. Furthermore, DC4S exhibits robust performance in scenarios with limited labeled data and showcases its effectiveness in transfer learning contexts.