T-2S: An Incremental Spectrum Prediction Framework Based on Bidirectional Distillation

In the dynamic electromagnetic environment, spectrum measurement results will exhibit drift. Thus, the datadriven model based on historical spectrum data struggles to maintain prediction accuracy, imposing challenges for dynamic spectrum management. To mitigate these limitations, this paper proposes an incremental spectrum prediction framework Teacher-2Students (T-2S), which includes a teacher model (TM) and two student models (SM). Initially, TM is trained using historical spectrum data to establish prior knowledge for the SM. Leveraging reverse distillation techniques, SM1 reconstructs the latent spectrum features captured by the TM. Spectrum drift detection subsequently operates by evaluating the feature discrepancy between SM1 and TM. When drift occurs, the lightweight SM2 is adaptively deployed for incremental prediction, employing knowledge distillation from the TM. Experimental results show that T-2S achieves both lower complexity of prediction model and higher prediction accuracy under spectrum drift conditions.