Efficient NOMA user detection in 5G using machine learning

Dalil Beknadj, Mohamed Azni, Claire Goursaud, Leonardo S. Cardoso, Cyrille Morin,
Efficient NOMA user detection in 5G using machine learning,
Physical Communication,
2025,
102901,
ISSN 1874-4907,
https://doi.org/10.1016/j.phycom.2025.102901.
(https://www.sciencedirect.com/science/article/pii/S1874490725003040)
Abstract: Accurate transmitter identification at the physical layer is essential in modern wireless communication systems to strengthen security, minimize overhead, and optimize overall efficiency. Using shallow descriptors and support vector machines (SVM), this work proposes a machine learning-based approach for transmitter identification in 5G networks, where transmitters access the radio resources using the Non-Orthogonal Multiple Access (NOMA) technique. Our approach mainly exploits the characteristics of the preamble in the physical Random Access Channel (PRACH). Unlike deep learning-based methods, which often suffer from high computational complexity and sensitivity to variations in radio propagation environments, the proposed approach provides a lightweight and precise alternative. We used local phase quantization (LPQ) and binarized statistical image features (BSIF) to extract robust features from Mel-Spectrogram representations of I/Q signals. By converting the in-phase (I) and quadrature (Q) components into image-based features, our method achieves high classification accuracy while ensuring rapid processing. In addition, we conducted a thorough analysis to identify the most relevant features of both the I and Q components, aiming to determine which characteristics play a crucial role in accurate transmitter identification. Experimental results demonstrate a classification accuracy of 97.32% for the imaginary part (Q), 91.49% for the real part (I), and 97.36% for the complex signal as a whole, highlighting both the effectiveness and efficiency of our approach in dynamic wireless environments.
Keywords: NOMA user detection; 5G networks; Machine learning; Mel-Spectrogram; Shallow descriptors; SVM