As the Internet of Things (IoT) continues to expand, there is a growing necessity for improved techniques to authenticate the identity of wireless transmitters to prevent unauthorized network access. In this dissertation, we develop a series of physical-layer authentication, or radio frequency (RF) fingerprinting, techniques which utilize methods from deep learning to train convolutional and recurrent neural network models to verify the identity of wireless transmitters which meet the IEEE 802.15.4 standard.

First, we develop a technique which utilizes a convolutional neural network (CNN) to identify or verify the identity of a transmitter from which a time-domain complex baseband signal was recorded. This technique relies on an extensive pre-processing sequence to remove sources of potential bias and trivial features from the received waveforms, and derives an estimated error signal from each recording from which the CNN learns discriminatory features. We demonstrate the effectiveness of the technique on a set of seven off-the-shelf ZigBee devices recorded outside in an urban environment, as well as in a laboratory environment with artificial noise over a wide-range of signal-to-noise ratios (SNRs).

Next, we train a series of models which utilize both convolutional and recurrent elements to improve the performance of the previous technique in the presence of high levels of noise and expand the evaluation to a larger set of twenty-five devices. We evaluate several realistic scenarios, including the performance in typical multipath environments and the ability to correctly reject previously unseen devices. In order to justify the proposed pre-processing sequence, we present experimental results that demonstrate weaknesses in fingerprint verification classifiers in which frequency synchronization is not performed. Finally, we present a simple technique to reduce the amount of memory required for a collection of fingerprint models by up to 95% without loss of performance.

To further enhance the security of the trained fingerprint models, we propose a generative adversarial network (GAN) architecture and training procedure to provide additional training examples for the classifiers. We show that fingerprint classifiers that are trained exclusively on real devices cannot reliably reject GAN-generated signals. Furthermore, we illustrate that augmenting the training process of the fingerprint models with GAN-generated signals reduces this vulnerability, even if the GAN used for training and inference are different.

Finally, we assess the practicality of transferring an RF fingerprint model from one receiver to another. Experimentally, we demonstrate significant degradation in classification performance when a fingerprint model is learned using signals recorded on one receiver and evaluated using signals recorded on another receiver. First, we show that generalization may be improved by including multiple receivers in the training process. Then, we develop a calibration procedure whereby models learned on a single receiver can be transferred without alteration to another receiver by learning a transformation function, implemented as a residual neural network, to model the variations between the two receivers. We perform several experiments with ten commercial receivers to confirm the effectiveness of the technique under realistic constraints.