A Spiking Few-Shot Learning Framework for Drone RF Signal Recognition

Radio-frequency (RF)-based drone recognition plays an increasingly important role in security surveillance and airspace management with the rapid development of drone communications and intelligent sensing technologies. However, most existing deep learning methods rely on large annotated datasets and energy-intensive artificial neural networks (ANN), which limits their applicability in sample-scarce and resource-constrained scenarios. To address these challenges, this paper proposes a spiking few-shot learning framework for RF-based drone identification that combines the data-efficient learning capability of Few-Shot Learning (FSL) with the inherent energy efficiency of Spiking Neural Networks (SNN). The proposed framework employs a pretrained SNN backbone and a Feature-Integrated Multi-Gated Attention (FIMA) module, where a multi-gated mechanism enables adaptive feature weighting and multi-head attention models cross-sample relationships, followed by task-adaptive fine-tuning to optimize FIMA and prototype representations for improved discriminative capability in few-shot tasks. Experiments conducted on the DroneRFb-Spectra dataset demonstrate that the proposed framework achieves accuracies of 99.27%, 98.47%, and 97.28% on the 3-way 1-shot, 5-way 1-shot, and 8-way 1-shot tasks, respectively, outperforming recent methods by approximately 2–5%. Meanwhile, the SNN-based feature extractor significantly reduces inference energy consumption to 41.9% of that of ANN-based counterparts, while also improving accuracy by 1–3%. Ablation studies further verify the effectiveness of the multi-gated multi-head attention mechanism and the fine-tuning stage, contributing approximately 1–4% and 4–9% performance improvements, respectively.