CNN-Driven QAM-Modulation Classification for Cognitive Radio-based IoT Networks
Sainiwetha Saikrishnan, Poornima Pandian · 2025
Cognitive Radio (CR) has revolutionized wireless communication due to efficient spectrum access and utilization of radio frequency resources. This paper investigates deep learning framework to enhance sensing, classification and decision making capabilities of Cognitive Radio systems for IoT applications. We explore five Conventional Neural Network (CNN) architectures such as ResNet, DenseNet GoogleNet, AlexNet and ResNeXt for automatic modulation classification. Each CNN model was evaluated based on classification accuracy over varying signal-to-noise ratio conditions, and computational efficiency such as inference time. This effectively addresses the challenges in IoT environments, such as spectrum scarcity and dynamic channel allocation. Analysis of five CNN models on DeepSig RadioML 2018.01A dataset revealed that DenseNet outperformed other models in terms of classification accuracy. Furthermore, inference time analysis revealed that AlexNet and GoogleNet are more suitable to be deployed on edge devices with limited resources due to their lower computational overhead but had their head down in terms of classification accuracy. These findings highlight important trade-offs between accuracy and efficiency, providing valuable insights for selecting suitable CNN architectures in real-time spectrum monitoring and dynamic wireless environments. With IoT devices relying on sensors, operating in shared and unlicensed frequency bands, automatic modulation classification using CNN in cognitive radio ensures interference-free communication.