Deep Learning Model Portability for Domain-Agnostic Device Fingerprinting

RF (Radio Frequency) device fingerprinting approaches using deep learning have recently emerged as potential methods of identifying devices solely based on their RF transmissions. However, these recently proposed approaches suffer from the domain portability problem, in that when the deep learning models used for fingerprinting are trained on data collected on one (source) domain but tested on data collected on a different (target) domain, the models will not perform well. For example, a change in the receiver used for data collection, in the day on which data was captured, or in the configuration settings of transmitters amounts to a change in the domain. This work proposes a technique that uses metric learning and model calibration to enable a model trained with data from one source domain to perform well on data collected on another target domain. This is accomplished with only a small amount of training data from the target domain and without changing the weights of the model, which makes the technique computationally quick. The effectiveness of the proposed technique is assessed using RF data captured using a testbed of real devices and under various different setup scenarios. Our results show that the proposed technique is viable and useful for networks with limited computational resources and applications with time-critical missions.