Multi-Task Terminal Radio Frequency Fingerprint Feature Extraction Framework Based on Shuffle Network

In the process of extracting and identifying wireless radio frequency signal fingerprints, the hardware differences between devices of the same type are too small to extract sufficiently distinct and effective features. To address this issue, this paper proposes a multi-task terminal radio frequency finger-print feature extraction framework based on a shuffle network, called RFFRe (Radio Frequency Fingerprint Re-identification). This framework can effectively capture the set of features corresponding to the subtle hardware differences from radio frequency signals. The framework employs a ResNet network model, removing the traditional fully connected layer and using the deep network model solely as a feature extractor for radio frequency signals. Additionally, it incorporates a shuffle network to apply additional perturbations to the extracted features, enhancing feature variability. Further, it performs multi-task learning on the perturbed feature space, using metric learning and representation learning to extract the distribution characteristics of feature vectors in the feature space. This compactly clusters intra-class features while improving inter-class discriminability. This model can complete the device registration process with a single training session when new devices are added and achieve quick identification through simple feature similarity comparison during device access. Experimental results show that RFFRe achieves a recognition Rank-1 accuracy of 99.4% on devices of the same type with minimal hardware differences, outperforming similar solutions.