CONIX Publication

With the rise of the Internet of Things (IoT), ensuring the security of smart cities that depend on IoT devices is of paramount importance. To satisfy the limited computational and power budget of such devices, machine-learning based passive physical layer transmitter authorization systems have been introduced recently. These systems generally rely on the hardware fingerprints of wireless transmitters to differentiate between a given set of authorized transmitters and to sieve out unauthorized transmitters. In this project, we demonstrate that neural networks are capable of mimicking such hardware fingerprints in a variety of settings. First, we show that given a dataset of wireless signals captured from authorized transmitters, variational autoencoders could be used to generate 1). more authorized samples and 2. samples from transmitters other than those in the authorized set (unauthorized samples). This is extremely useful for training more accurate classifiers for authorization. Next, by using reinforcement learning techniques, we show that unauthorized transmitters can add carefully learned perturbations to their transmitted signals to mimic authorized transmitters, thereby allowing them to fool standard deep-learning based transmitter authorization systems with high rates of success. We supplement this by showing that pre-training authorization classifiers with adversarial samples obtained using classical methods improves their robustness against such impersonators.