Machine-Learning PUF-based Detection of RF Anomalies in a Cluttered RF Environment

With the emergence of software defined radio (SDR) where a computer program defines transceivers’ physical layer functions, waveforms can change dynamically. SDR benefits new protocol deployment, enabling smart wireless communication applications. However, SDR makes it easier to mimic authorized transmission, leaving wireless networks vulnerable to spoofing attacks. This work explores ways to detect such radio frequency (RF) anomalies. Specifically, a machine-learning structure called convolutional neural network (CNN) possesses merits of local perception and shift invariance, matching the characteristics of our sampled SDR data. Therefore, we design a CNN for detection of RF anomalies. Furthermore, a physical unclonable function (PUF) provides physical-layer security by identifying a device analogous to human fingerprint. Our CNN extracts waveform features as well as PUFs of transmission devices, from which we train and validate a classification model. The trained model can detect and identify spoofed signals. As proof-of-concept experiments, we generate RF signals with Ettus Universal Software Radio Peripherals (USRPs) and GNU Radio software. We then use the dataset to train our CNN classification model that analyzes features of the RF signals and the USRPs’ PUFs. To expand the robustness of our CNN model in cluttered RF environments typical in the Internet of Things (IoT), we generate satellite signals of Automatic Dependent Surveillance- Broadcast (ADS-B) for aircraft tracking. The testing results confirm the promise of machine-learning PUF-based security enforcement in cluttered RF environments.