RF Fingerprinting of Software Defined Radios Using Ensemble Learning Models

Abstract

Machine Learning (ML) is becoming a transformative technology in wireless communication. The deployment of large scale RF devices particularly in IoT applications escalates security threats and also setting up of secure networks using wireless devices is becoming a big challenge. Along with ensuring security, identifying each RF device in an autonomous network is essential and the RFML (Radio Frequency Machine Learning) can play a crucial role here. This paper focuses on the RF characterization of a set of Software Defined Radios (SDR) using advanced machine learning models. This helps to identify each SDR module in the deployed network which runs only a specific protocol in a particular network. The SDRs will be configured for a particular specification and the test will be conducted. The transmitted data from multiple radio nodes were collected using a reconfigurable radio’s receive chain in IQ-format, in the laboratory environment. The RF features like IQ-imbalance, DC-offset and the image leakages in the multicarrier modes were used to set fingerprints for identifying the reconfigurable radios. Two ensemble learning models Random Forest and AdaBoost were used to train and develop predictive models to identify the radio. At a SNR of 30dB Random Forest achieved an accuracy of 85% and AdaBoost achieved an accuracy of 78% with 32K multicarrier data. A maximum recognition rate of 92% is achieved with RF and 83% with AdaBoost.