Research On Radio Frequency Fingerprint Identification Method Based On Transfer Learning

Radio frequency fingerprint(RFF)identification technology is an emerging technology that has received much attention in the field of wireless communication in recent years,and has been widely applied in the Internet of Things(Io T)domain.As a novel physical layer security technology,RFF identification can effectively address the authentication and identification tasks of Io T devices.Traditionally,feature-based machine learning(ML)methods have been widely used for RFF identification technology.However,these traditional methods heavily rely on complex feature extraction,which greatly limits their practical applications.In recent years,deep learning(DL)has been introduced to RFF feature extraction and identification,and has achieved good results.However,DL-based methods still have certain limitations.On one hand,the parameters and computational complexity of neural network models in deep learning are often large,which limits their application in practical Io T scenarios.On the other hand,DL-based methods are difficult to deal with variations in different recognition scenarios,such as changes in channel environment.In addition,DL-based methods require high-quality signal data for training,which is often difficult to obtain in practical applications.To address the uncertainties of DL-based RFF identification methods in practical applications,this paper mainly studies the use of transfer learning(TL)methods for RFF identification tasks under different channel environments.First,this paper designs a lightweight neural network model for RFF feature extraction and classification.Compared with traditional convolutional neural networks(CNN),the proposed Conv Mixer neural network(Conv Mixer Net)effectively reduces the parameters and computational complexity by using residual-connected depth-wise separable convolution modules,making the model more lightweight.Furthermore,the proposed model achieves better recognition performance on RFF dataset with different signal-to-noise ratios,especially under low signal-to-noise ratios.Then,this paper proposes three different deep transfer learning methods for RFF identification tasks under different channel environments.Based on the designed Conv Mixer Net as the base network,the pre-trained model under the original channel environment is fine-tuned using different strategies with new signal data under the different channel environment,which leads to better recognition performance and further reduction of model training parameters.Finally,for the practical applications where the signal data under a new channel environment are mostly unlabeled,this paper proposes an unsupervised TL method based on domain adaptation(DA)for RFF identification.The proposed method still uses the Conv Mixer Net as the feature extractor,and obtains a feature extraction network suitable for target signal data through adversarial training between the source domain and the target domain,which achieves better recognition performance even when the training samples in the target domain are unlabeled.