Sensing, Estimation, and Security of the Frequency Spectrum Using Shallow and Deep Learning Techniques

Abstract

The frequency spectrum is a limited resource that has experienced a growing demand in recent years, especially with the advent of 5G and 6G technologies. Spectrum sensing, estimation, and security are essential factors for increasing efficiency and flexibility in spectrum utilization, enabling its usage optimization and ensuring security for a larger number of users. Spectrum sensing is crucial for filling spectral holes, relieving congested frequency bands. Estimating spectrum conditions also plays an extremely important role in designing and proposing specific solutions and services for different conditions. Last but not least, with the development of generative artificial intelligence, spectrum security becomes essential for developing measures to mitigate malicious user activities. In this context, this thesis presents research related to these approaches. Experimental results suggest promising prospects for these approaches, implying improvements in efficiency, robustness, and low latency in current communication systems. For instance, in the spectrum sensing approach, the proposed simplified ResNet achieved 98% accuracy under a noise level of -134 dBm/Hz, with a response time below 0.05 seconds, ensuring low latency. For spectrum estimation, the XGBoost and Transformer models achieved the best correlation coefficients for identifying the noise level and the distance between users in a spectrum sensing environment, with values of 0.98 and 0.84, respectively. Lastly, in spectrum security, the proposed GAN was able to deceive deep cooperative spectrum sensing models in over 98% of cases.

Keywords: Spectrum Sensing, Machine Learning, Deep Learning, Generative Adversarial Network, Transformer.