Research On RF Fingerprint Feature Selection And Dimensionality Reduction Method

With the development of mobile communication devices and Internet of Things technologies,wireless networks are vulnerable to malicious attacks due to their openness.The Radio Frequency Fingerprinting mechanism of wireless communication devices has become a new security mechanism that can effectively identify devices.In this paper,the identification of power amplifier RF fingerprinting is taken as an example.Aiming at the problem that the RF fingerprint features have high feature dimensions and the information between the features may be irrelevant or redundant,the RF fingerprint feature selection and dimensionality reduction methods are mainly studied.Firstly,the basic theory of RF fingerprinting feature extraction and recognition is studied.The basic model of RF fingerprint feature extraction and recognition of power amplifier is designed.The signal acquisition of power amplifier,test,covariance distribution feature extraction and K-nearest neighbor classification are completed respectively.It is verified that the eight amplifiers studied in this paper have subtle differences and the covariance distribution feature extraction method is an effective feature extraction method.Secondly,the RF fingerprinting feature selection method is studied.Based on the research of RELIEF-F method,F Score method and Laplacian Score method,an ensemble feature selection method based on these three methods is proposed according to the feature selection method ensemble principle,and the feature stability is evaluated by Spearman correlation coefficient.The simulation results show that the proposed feature selection method can achieve better classification results with fewer features,and the classification accuracy is higher under most SNR than the previous three methods.The stability of the ensemble feature selection method is also better.Finally,the RF fingerprinting dimensionality reduction method is studied.Based on the research of principal component analysis,linear discriminant analysis and auto-encoder,the principal component analysis-linear discriminant analysis method is studied,and the feature separability is evaluated by the distance ratio criterion.The simulation results show that the classification accuracy of principal component analysis-linear discriminant analysis method is better than that of principal component analysis method,linear discriminant analysis method and auto-encoding method under most SNR,and the features of the principal component analysis-linear discriminant analysis method are more separable.