CNN-Based Automatic Modulation Classification for Cognitive Radio: Benchmarking Against Fourth- Order-Cumulant Feature Classification

After a cognitive radio (CR) determines that a channel is occupied, identifying the primary user’s modulation scheme — automatic modulation classification (AMC) — supports interference-aware coexistence, standards identification, and physical-layer security. This paper presents a simulation-based benchmark of a one-dimensional convolutional neural network (CNN) operating directly on raw
in-phase/quadrature (I/Q) samples against the classical fourth order- cumulant feature classifier, for six modulation schemes
(BPSK, QPSK, 8PSK, 16QAM, 64QAM, CPFSK) under AWGN with random carrier-frequency offset and phase. On a held-out test set of 3,240 examples spanning −10 to +15 dB SNR, the proposed CNN reaches 81.1% accuracy at +15 dB and 57.5%overall, compared with 47.8% and 26.7% for the cumulant based classifier, a gap that is largest at moderate-to-high SNR and consistent with the modern automatic-modulation classification literature. The paper further reports per-class accuracy, confusion matrices at low and high SNR, model complexity comparison, and a genuine limitation: random carrier phase substantially degrades discrimination among
phase- and amplitude-based schemes (QPSK, 8PSK, 16QAM) relative to the constant-envelope, phase-agnostic classes (BPSK, CPFSK). All results are produced by Python-based Monte Carlo signal simulation and model training (NumPy, scikit-learn, TensorFlow/Keras); no captured RF data were used.