Lightweight Neural Network With Fixed Classifier for Enhanced Drone RF Signal Recognition

Although small drones provide significant convenience through various applications, their widespread use poses a major threat to public safety, particularly when operated without authorization. To address this challenge, a deep learning (DL) recognition system based on drone radio frequency (RF) signals has emerged as a potentially effective method for efficiently identifying drone classes. However, traditional DL models require large computing resources, making them unsuitable for deployment on mobile devices. In this paper, a novel network model consisting of a fixed classifier architecture inspired by neural collapse (NC) is proposed. Moreover, a soft thresholding and broadcasting mechanism is integrated, an enhanced Squeeze-and-Excitation (ESE) module is introduced, and information theory principles are employed to design the loss function for the Equiangular Tight Frame (ETF) classifier. Simulation results show that the average accuracy for recognizing drone classes is one percentage point higher than that of other methods at high signal-to-noise ratio (SNR). Furthermore, experiments show that the proposed drone recognition method achieves an average accuracy of 96.64% and demonstrates strong generalization capability. Additionally, the advantages of the fixed classifier are validated by analyzing the sample distribution in the 3D semantic space. It also exhibits lower model complexity and fewer parameters compared to traditional approaches. Finally, ablation experiments confirm that our proposed model makes a trade-off between the computational complexity and the average accuracy.