Automatic modulation classification (AMC) plays a central role in adaptive and cognitive wireless communication systems by enabling reliable signal interpretation under varying channel conditions. This study presents a comprehensive evaluation of deep learning based AMC using constellation diagram representations under additive white Gaussian noise (AWGN) across six signal to noise ratio levels ranging from \(-5\) , 0, 5, 10, 15, 20 dB. A balanced dataset consisting of 24,000 samples from eight modulation formats, namely BPSK, QPSK, 8PSK, 16PSK, 32PSK, 16QAM, 64QAM, and 256QAM, is employed to benchmark four convolutional architectures. These include the classical AlexNet, two lightweight modern backbones, MobileNetV2 and EfficientNet-B0, and the proposed Squeeze–Excite Convolutional Attention Lightweight Network (SCALNet). SCALNet integrates depthwise separable convolutions, squeeze–excite channel reweighting, and residual connections to enhance feature discrimination with minimal computational cost. Experimental results demonstrate that SCALNet consistently achieves the best robustness at low and mid SNRs, obtaining the highest F1-score at \(-5\) dB (0.6004) and stable accuracy up to 20 dB, while requiring only 0.37M parameters and 1.36 MB of memory. Compared with AlexNet, SCALNet reduces parameter count and FLOPs by more than %99, while MobileNetV2 and EfficientNet-B0 provide intermediate performance–efficiency trade-offs. Efficiency analysis shows that SCALNet delivers the lowest inference latency (4.49 ms at 20 dB) and the smallest computational footprint across all SNRs, confirming its suitability for embedded or resource-constrained AMC deployments. These results position SCALNet as a lightweight yet highly reliable architecture, with future work planned to extend the evaluation toward fading, phase noise, and nonlinear RF impairments for emerging 6G communication scenarios.
