DC-FTNet: A Parameter-Free Fourier-Based Lightweight Network for Automatic Modulation Classification in OTFS

Abstract

Automatic modulation classification (AMC) for orthogonal time frequency space (OTFS) systems is crucial in non-cooperative scenarios. While vision-transformer (ViT)-based classifiers achieved high accuracy, they suffered from quadratic computational complexity due to self-attention. To address this limitation, this letter proposes a lightweight dual-channel Fourier-based token mixing neural network (DC-FTNet), based on token mixing with Fourier transforms, that replaces self-attention with parameter-free Fourier-based token mixing, enabling efficient global feature learning with quasi-linear complexity. By processing the in-phase and quadrature components as separate channels, DC-FTNet preserves the physical 2D OTFS structure without requiring complex-valued operations. Experimental results show that DC-FTNet achieves superior overall accuracy, outperforming ViT baselines while maintaining approximately $2.5\times $ fewer parameters and lower latency. Additional evaluations under representative hardware impairments, including carrier frequency offset, timing offset, phase noise, I/Q imbalance and DC offset, further demonstrate its robustness, making it a practical and efficient solution for OTFS-based intelligent receivers.