Toward Receiver-Agnostic RF Fingerprint Verification
Kevin Merchant, Bryan Nousain · 2019
A number of successful RF fingerprint classifiers have been demonstrated, but relatively few results evaluate the impact of changing the receiver used for training and inference. In this work, we record a set of 25 ZigBee transmitters with 10 independent, unsynchronized receivers and first show that similar performance may be achieved by a neural network-based RF fingerprint verification system on all receivers when training and inference are performed on the same receiver. Next, we show significant performance degradation when different receivers are used for training and inference. We propose two methods to address this shortcoming. The first method shows that the performance degradation is reduced when the training process uses several receivers. The second method proposes a calibration procedure whereby a single neural network-based transformation is learned per receiver and allows fingerprint verification models learned for a different receiver to be reused without retraining.