Abstract
Automatic modulation recognition (AMR) is vital for 5G/6G, yet existing methods often overlook the specific spectral structures of dominant OFDM systems. To address this, we propose the Fourier Adaptive Filter with Attention (FAFT), a parameter-efficient framework that explicitly models OFDM spectra. FAFT integrates a learnable FFT-based adaptive filter branch with a lightweight time-domain convolutional branch, fused via channel attention. Additionally, a novel frequency-domain regularizer is introduced to enhance spectral feature learning. Experiments on RML2016.10a, RML2016.10b, and the practical EVAS OFDM dataset demonstrate that FAFT achieves competitive accuracy with remarkable efficiency (0.13M parameters, 39.3M FLOPs). Its robustness under low SNR and multipath conditions highlights its strong potential for practical 5G/6G deployment.