Autoencoder based framework for drone RF signal classification and novelty detection

The increasing use of Unmanned Aerial Vehicles (UAVs) in modern civilian and military applications shows the urgency of having a robust drone detector that detects unseen drone RF signals. Ideally, the system can also classify known RF signals from known drones. This study aims to develop an incremental-learning framework which can classify the known RF signals, and further detect novel RF signals. We propose DE-FEND: A Deep residual network-based autoEncoder FramEwork for known drone signal classification, Novelty Detection, and clustering. The known signal classification and novelty detection are performed in a semi-supervised and unsupervised manner, respectively. We used commercial drone RF signals to evaluate the performance of our framework. With our framework, we obtained 100% novelty detection accuracy at 1.04% False Alarm Rate (FAR) and 97.4% classification accuracy with only 10% labelled samples. Furthermore, we show that our framework outperforms the state-of-The-Art (SoA) algorithms in terms of novelty detection performance.