Spectrum Anomaly Detection Method Fusing Time-Frequency Attention

Detecting spectrum anomaly effectively is vital for spectrum management. Existing deep learning (DL)-based spectrum anomaly detection (SAD) methods struggle to detect spectrum anomaly by extracting the latent feature of spectrum data, which ignores the time-frequency feature in the spectrum data. Regarding this problem, we introduce TFAM-EGAN, an innovative SAD framework that integrates the Encoder Generative Adversarial Network (EGAN) with Time Frequency Attention Mechanism (TFAM). This framework first employs encoder and TFAM to capture the latent, time-frequency feature. Then, decoder reconstructs the spectrum data based on the extracted feature. Finally, the spectrum anomaly are detected by calculating the reconstruction error. The simulations confirm the superiority of our proposed method in SAD tasks.