Few Shot Specific Emitter Identification Based on Triplet Loss

Deep learning-based RF fingerprinting has emerged as a crucial approach for device authentication. However, this technology often requires a large number of labelled samples practically. To address this issue, this paper proposes a metric learning algorithm for few-shot RF fingerprinting. This method leverages the triplet loss and PCA to mitigate the impact of differences between the few-shot distribution and overall distribution on identification performance. Additionally, we introduce multi-size convolution kernels and channel attention mechanism to enhance the neural network’s feature extraction ability. In our research, we find and prove that low-frequency signals are more suitable for few-shot learning compared with high-frequency signals. In experiments conducted on Bluetooth data, our method accurately identified eight emitters with 96.13% accuracy with only 50 training samples per emitter.