Deep Learning for UAV Detection and Classification via Radio Frequency Signal Analysis

Unmanned Aerial Vehicles (UAVs) are advertised as great tool that benefits society and humanity. However, UAVs also pose significant security threats ranging from privacy invasions, to interfering with commercial aircraft landing and takeoff, to accidently crashing into vehicles or people, to military or terrorist attacks. Consequently, there is a pressing need to detect and identify UAVs to mitigate such potential risks. While image-based methods are crucial for UAV detection, radio frequency (RF) emissions offer additional valuable insights. Analyzing RF signals, such as those used in UAV-ground station communications, can provide information about UAV types based on distinct frequency usage or communication patterns. This work introduces a deep-learning-based approach for recognizing and identifying UAVs using their RF emissions. Captured RI' signals are transformed into spectrograms, which are subsequently analyzed using deep neural networks. Existing methods achieve low identification accuracy, for instance the ResNet-50V2 model achieves an accuracy of 85.39% even in controlled, laboratory, noise-free conditions. Moreover, in outdoor environments at distances of 50m and 1.00m, the accuracy drops to 68.90% and 56.88%, respectively. To improve classification accuracy in outdoors, a CNN model was developed, yielding an accuracy of 78.12%. Leveraging the ResNet 50 V2 architecture, remarkable accuracy of 95.08% was attained in binary classification tasks involving a dataset comprising 195 mixed UAV images and 290 non-mix UAV images.