Spectrum Transformer: Wideband Spectrum Sensing using Multi-Head Self-Attention

Data-driven machine learning techniques have been advocated to detect the existence of target signals in complex wireless environments. However, wideband spectrum sensing has to deal with special challenges, including enlarged data dimensionality, insufficient training data, and implicit inter-band dependencies. Facing these issues, conventional deep models unfortunately suffer from high model complexity, inferior sensing accuracy, and notorious over-fitting issues due to ignorance of domain knowledge on spectrum occupancy patterns given limited training data. All these factors lead to ineffective learning model design in the most current literature. To fill this gap, this paper proposes a novel multi-task learning solution for wideband spectrum sensing. Empowered by the multi-head self-attention mechanism, we design an efficient Spectrum Transformer architecture to effectively learn both the inter-band spectrum occupancy correlations and the inner-band spectrum features of the wideband spectrum. Spectrum Transformer outperforms the existing methods based on convolutional neural networks especially in small-data regimes, by achieving higher sensing accuracy with 89% reduction in model complexity.