Radiofrequency Fingerprint Feature Extraction and Recognition Using a Coordinate Attention-Guided Deep Residual Shrinkage Network

The current radiofrequency fingerprint (RFF) feature extraction and recognition methods using deep learning (DL) lack multiple dimensions for the extracted feature vectors. At the same time, due to the black box property of deep learning, problems such as overfitting often occur in experiments, leading to a decrease in classification effect. In this paper, we proposed a DL method for RFF feature extraction and identification using the coordinate attention-guided (CA) deep residual shrinkage network (DRSN-CA). DRSN-CA integrates the advantages of CA mechanism and DRSN, optimizes the internal attention mechanism for residual shrinkage building unit (RSBU) into a better CA mechanism, and generates a new RSBU-CA module. The module combines the two dimensions of space and channel, which not only reduces the loss of information because of dimensionality reduction, but also enhances the impact of feature extraction for the model under low signal-to-noise ratio. The quantitative results indicated that the effect of DRSN-CA is promising. Compared with the traditional RFF feature extraction and recognition methods, the classification recognition accuracy of 92.5% can be achieved at the maximum distance of 62ft at the lower signal-to-noise ratio of 0dB.