Researchers have introduced LiteSGNet, an ultra-lightweight automatic modulation classification network designed for edge spectrum sensing in next-generation mobile networks. Evaluated on the RML2016.10a and RML2016.10b benchmark datasets, the proposed model achieved overall classification accuracies of 63.55% and 66.11%, respectively, while operating with 9.62K parameters and 0.401M FLOPs.
To address the substantial parameter counts and computational demands of existing deep learning approaches on resource-constrained edge terminals, LiteSGNet explicitly constructs four-channel features consisting of in-phase, quadrature, amplitude, and normalized-phase components. Its architecture combines Mobile Inverted Bottleneck Convolution blocks as a compact time-domain backbone with an inserted LiteSpectralGate frequency-domain gating module to adaptively reweight spectral components.
Under the evaluated benchmark setting, LiteSGNet demonstrated the smallest parameter count among compared methods, establishing a favorable accuracy-to-parameter trade-off for front-end edge perception modules in electromagnetic spectrum embodied intelligence. The researchers noted, however, that its FFT-based operator does not yield the lowest single-sample CPU latency.
Subject of Research: Automatic modulation classification for edge spectrum sensing
Article Title: LiteSGNet: An Ultra-Lightweight Automatic Modulation Classification Network for Edge Spectrum Sensing
Article References: Kong, X., Zuo, Y., & Ma, J. (2026). LiteSGNet: An Ultra-Lightweight Automatic Modulation Classification Network for Edge Spectrum Sensing. Mobile Networks and Applications. https://doi.org/10.1007/s11036-026-02543-3
Image Credits: AI Generated
DOI: 10.1007/s11036-026-02543-3
Keywords: Automatic Modulation Classification, Spectrum Sensing, LiteSGNet, Edge Computing, Deep Learning, Mobile Networks, MBConv, LiteSpectralGate, Ultra-Lightweight, Automatic, Modulation, Classification
News Source: Denise Maddox. (November 5, 2026). LiteSGNet Enables Ultra-Lightweight Signal Modulation Classification for Edge Spectrum Sensing. Scienmag.
