Lightweight Deep Learning-Based Spectrum Sensing Under Dynamic PU Activity in Cognitive Radio
Luepol Pipanmekaporn, Wilaiporn Lee, Warodom Phungjununt, Kitipoth Wasayangkool, Akara Prayote, Kanabadee Srisomboon · IEEE Open Journal of the Communications Society · 2026
In practical scenarios, dynamic primary users (PU) activity causes intermittent signal appearances within a sensing slot resulting in only partial signal capture by the secondary user (SU). Under the dynamic activity conditions, it leads to a critical limitation in conventional spectrum sensing approaches, which are often unable to reliably detect the presence or absence of the PU. Despite its ability to analyze time-sequential signals, deep learning-based sensing remains impractical for resource-constrained cognitive radios due to high computational demands. To address both computation and detection reliability, we propose a lightweight deep learning-based spectrum sensing (LDS) model by exploiting knowledge distillation framework with long short-term memory (LSTM) to achieve high spectrum sensing performance as expensive complexity attention-based LSTM with significantly fewer hidden units. By distilling knowledge from the attention-based LSTM, the LDS model demonstrates superior performance in spectrum sensing comparable to the attention-based LSTM and consistently outperforms standard LSTM, convolutional neural network with recurrent neural network (CNN-RNN) and generative adversarial networks augmented spectrum sensing (GAN-Aug) framework while reducing the complexity of the attention-based LSTM by approximately 99.39%, GAN-Aug by about 98.04% and CNN-RNN by around 96.72%. Furthermore, the LDS model outperforms the performance of conventional spectrum sensing techniques. Given its low complexity and strong performance, the LDS model is particularly well-suited for spectrum sensing in dynamic, time-sensitive cognitive radio environments.