Enhanced RF Fingerprinting for IoT Devices with Recurrent Neural Networks

As the Internet of Things (IoT) continues to expand, there is a growing necessity for improved techniques to authenticate the identity of wireless transmitters. In this paper, we develop a physical-layer authentication technique using a neural network structure with both convolutional and recurrent components to distinguish transmissions originating from a particular target device from all others. In addition, we demonstrate strong performance in a realistic multipath channel environment, as well as show that classifier performance remains strong when presented with transmissions from devices that were never seen by the classifier during training. We explore the latter benefit in more detail via an experiment which measures the performance on unknown devices as a function of the number of devices seen during training. Next, we highlight the importance of frequency synchronization prior to fingerprint extraction by demonstrating that a network trained on unsynchronized transmissions is easily fooled by a simple frequency shift in a transmitted waveform. Finally, we increase the applicability of our approach to IoT devices by presenting a simple technique for reducing the memory footprint of trained models by 95% while maintaining strong performance.