Two-Stage Anomaly Detection in LEO Satellite Network

We introduce a novel two-stage method for detecting anomaly signal in Low Earth Orbit (LEO) satellite network in response to increasing clutter signals and interference. First, a convolutional neural network (CNN) classifier is trained on real sensor signals and synthesized wireless modulated signals. Then, an anomaly detector is used to detect the classified signal. To address the limited computing resources on the satellite, we utilize transfer learning to reduce the scale of the classifier and anomaly detector. Our proposed method consists of a multi-class CNN model that reliably detects the modulation methods used in a specific satellite environment by I/Q signals and a recurrent neural network model that identifies anomalies when events significantly deviate from expected or predicted values. Experimental results show the effectiveness of our proposed method.