RF-Based UAV Classification Using CNN in Noisy Environments

Unmanned Aerial Vehicles (UAVs) have proliferated across numerous sectors, introducing sig nificant security concerns due to potential misuse. This study presents a convolutional neural network (CNN) architecture specifically customized for drone classification using radio frequency (RF) signals under adverse noise conditions. Our contribution lies in demonstrating the synergistic effectiveness of two key components: (i) an RF-specific data augmentation strategy (frequency masking, time masking, and random scaling) to improve robustness against real-world signal variations, and (ii) residual blocks to enhance deep spectral feature extraction while preserving gradient flow. Through a comprehensive ablation study across multiple signal-to-noise ratio (SNR) levels, we show that data augmentation alone improves classification accuracy by up to +24.7% at–14 dB, while the addition of residual blocks yields a further +14.1% improvement, achieving a total gain of +38.8% over the baseline CNN. Our proposed model achieves 96.65% accuracy at–14 dB SNR, significantly outperforming existing approaches and maintaining stable performance across higher SNR levels. These results validate that the integration of augmentation and residual learning provides a practically significant advancement toward robust UAV classification systems capable of real-world deployment in challenging RF environments.