Unauthorized unmanned aerial vehicles (UAVs) threaten Vehicle-to-Everything (V2X) spectrum security. Real-time edge detection faces strict hardware constraints, severe multipath fading, and Doppler distortions. This article proposes HA-EffNet, a physics-informed multi-task learning framework engineered for radio frequency (RF) sensing on roadside units (RSUs). The network restricts its temporal receptive field to align mathematically with the channel coherence time, thereby preventing deep noise overfitting. A hierarchical mechanism integrates Efficient Channel Attention (ECA) for shallow noise suppression and Receptive Field Attention (RFA) for deep signature extraction. Furthermore, the shared multi-task architecture simultaneously executes discrete classification and continuous spectral parameter regression, effectively halving computational overhead compared to redundant single-task deployments. Evaluations on the Microphase and DroneRFa datasets yield classification accuracies of 97.88% and 94.67%. Compound tests integrating Tapped Delay Line C (TDL-C) models and dynamic signal-to-noise ratio (SNR) variations validate algorithmic resilience against severe physical degradation. Utilizing a 0.12-million-parameter footprint, the network delivers a 0.84 ms inference latency and 1204.9 frames per second (FPS) throughput on the NVIDIA Jetson Orin Nano Super, providing a highly efficient edge-sensing solution.
Xu et al. (Wed,) studied this question.