Memory-Augmented Reverse Distillation for Continuous Spectrum Anomaly Detection
Abstract
Effective spectrum anomaly detection (SAD) is of paramount importance for maintaining communication security and spectrum utilization efficiency. Most existing research still focuses on training dedicated anomaly detection models for specific electromagnetic scenarios. However, developing such environment-specific SAD models may incur a high computational cost, restricting their practical applicability. In this paper, we investigate a continuous learning environment. Specifically, we introduce the Memory-Augmented Reverse Distillation for Continual Learning (MARD-CL), which is an innovative SAD method that integrates reverse distillation (RD) and memory enhancement module (MEM). Moreover, in order to enable the method to continuously learn in a series of electromagnetic environments, we perform the dual pooling distillation on the reconstructed data from student model and fusion feature from feature fusion module, respectively. The dual pooling distillation prevents the model from forgetting previously learned knowledge when learning new scenarios data, thereby solving the problem of catastrophic forgetting. Experiments have confirmed that proposed SAD method outperforms the existing methods.