Spectrum Sensing and Signal Classification: An Out-of-Distribution Detection Approach

Abstract

This thesis presents an innovative approach for spectrum sensing and signal classification based on out-of-distribution (OOD detection. Our proposed scheme showcases remarkable resilience in low signal-to-noise ratio (SNR) environments, attaining 100% accuracy in spectrum sensing and 97.7% accuracy in signal classification when a SNR as low as -14dB in our simulation settings. Our method incorporates a voting mechanism for both spectrum sensing and signal classification, removing the need for any classification layer, such as fully connected or softmax layers. A distinguishing feature that sets our research apart from existing studies is its capability to detect a new class of samples that have not been encountered during training. This added capability enhances the robustness, reliability, and security of our system, rendering it highly suitable for real-world applications. Furthermore, our scheme exhibits excellent scalability and flexibility, allowing for the assignment of weights to specific signals for sensing and classification, if necessary. This feature enhances the adaptability of our approach, enabling tailored optimization according to specific requirements.