GNN for LoRa Device Fingerprint Identification

This paper presents a deep learning-based system for the classification and identification of LoRa signals. The system initially preprocesses the RSSI and phase of LoRa signals and then maps them into a high-dimensional space via a FNN(Feed-Forward neural network). It calculates the similarity between signals using the Wasserstein distance and selects the top K most similar signals for each as first-order neighbors through heap sorting, constructing phase and RSSI(Received Signal Strength Indicator) graphs and transforming the signal classification problem into a graph-based node classification problem. The system employs GGC(Gated Graph Convolution) networks to extract feature representations from the graphs, which are then encoded by a time encoder, specifically an LSTM(Long Short-Term Memory) network, to obtain a coarse-grained temporal representation. Finally, a FNN is used to achieve a fine-grained representation of the results. Experimental results demonstrate that the proposed method achieves high classification accuracy, recall, F1 score, and accuracy across various datasets, proving the model’s robustness to environmental variations.