Non-cooperative OFDM Spectrum Sensing Using Deep Learning

Although spectrum sensing, a key technique in dynamic spectrum access, has been widely investigated, conventional methods suffer from carrier frequency offset (CFO), timing delay and noise uncertainty, which can significantly degrade the sensing performance. In this paper, we aim to tackle those challenging issues by developing a stacked autoencoder based spectrum sensing approach (SAE-SS). The SAE architecture is employed to effectively learn useful and hidden information from the original received signals. Compared to the existing sensing methods, our approach is more robust to CFO, noise uncertainty and timing delay. Unlike the traditional feature-based detection approaches, the proposed framework does not require the prior knowledge or specific features of incumbent users (IUs). Moreover, in comparison with machine learning based sensing approaches, our solution does not need any external feature extraction algorithms to extract specific features (that is essential for ML-based ones). Through extensive experimental results, our proposed method is demonstrated to achieve notably higher sensing accuracy, e.g., 29% reduced probability of miss detection, than that of state-of-the-art approaches.