Radio Frequency Fingerprinting System Based on Model Extension With Outlier detection

Abstract

Radio frequency (RF) fingerprinting is a way to find a transmitter using specific features in Radio Frequency signal. In recently, RF fingerprinting is getting more attention in electronic warfare or the internet of things network for security. It is an effective method to prevent malicious nodes’ access or support allies' electronic devices by identifying if the signal is from the unauthorized transmitter. In the past, RF Fingerprinting research focused on classification problems based on machine learning or classification algorithms. Recently, as deep learning is developed, RF fingerprinting based on deep learning has been studied. However, many other studies still assume a classification of known signal data. In the real world, there are limitations to collecting all signal information for RF Fingerprinting. Therefore, there can be an outlier in performing RF fingerprinting, and we need to consider the outlier. In addition, there is no research on RF Fingerprinting system considering outlier detection with learning. In this paper, we suggest self-learning RF Fingerprinting system. The proposed system is to use incremental learning with Outlier detection in RF Fingerprinting system. It consists of three phases: initial training, outlier detection, and model extension. In the initial training, we train the Convolutional Neural Network (CNN) using existing RF signal. In the outlier detection phase, we use the mahalanobis distance method to detect the outliers and relabel the outliers to new class. In the model extension phase, we retrain the CNN using relabeled data based on incremental learning, with minimal loss of existing knowledge. Our proposed system acquire a 96% accuracy on outliers after model extension.