Radio Frequency Fingerprinting Method Based on Empirical Modal Decomposition and Frequency Domain Phase Characterization

Radio frequency (RF) fingerprinting is an RF signal identification method based on unintentional features of radio devices. In order to improve the accuracy of RF fingerprinting of autocorrelated surveillance broadcast (ADS-B) signals, a feature extraction method based on feature fusion of ADS-B signals is proposed. First, we employed empirical mode decomposition (EMD) and Fourier Transform to extract the intrinsic mode functions (IMFs) and frequency domain phases of pulse sequences. Subsequently, the extracted features were downscaled using principal component analysis (PCA) and the two downscaled features were fused into a single feature matrix using the operation of matrix fork product. Ultimately, we have designed an innovative residual neural network (ResNet) that incorporates an attention mechanism and integrates depthwise separable (DWS) convolution with an LSTM layer for training on feature matrices. The performance of the model was rigorously assessed on a validation set. Validation was carried out using pulse sequences extracted from the ADS-B signals of 43 civil aircraft. The results show a classification accuracy of 99.65%, thus affirming the validity of the method for real data, multi-category classification.